hangX/lerobot
1
0
Fork 0
forked from tangger/lerobot
Code Pull Requests Activity

[pre-commit.ci] auto fixes from pre-commit.com hooks

Browse Source 
for more information, see https://pre-commit.ci
This commit is contained in:
pre-commit-ci[bot]
2025-03-24 13:41:27 +00:00
committed by AdilZouitine
parent 2945bbb221
commit 7c05755823
123 changed files with 1161 additions and 3425 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lerobot/common/policies/act/configuration_act.py
View File

@@ -171,9 +171,7 @@ class ACTConfig(PreTrainedConfig):
def validate_features(self) -> None:
if not self.image_features and not self.env_state_feature:
raise ValueError(
"You must provide at least one image or the environment state among the inputs."
)
raise ValueError("You must provide at least one image or the environment state among the inputs.")
@property
def observation_delta_indices(self) -> None:

175
lerobot/common/policies/act/modeling_act.py
View File

@@ -63,9 +63,7 @@ class ACTPolicy(PreTrainedPolicy):
config.validate_features()
self.config = config
self.normalize_inputs = Normalize(
config.input_features, config.normalization_mapping, dataset_stats
)
self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
self.normalize_targets = Normalize(
config.output_features, config.normalization_mapping, dataset_stats
)
@@ -76,9 +74,7 @@ class ACTPolicy(PreTrainedPolicy):
self.model = ACT(config)
if config.temporal_ensemble_coeff is not None:
self.temporal_ensembler = ACTTemporalEnsembler(
config.temporal_ensemble_coeff, config.chunk_size
)
self.temporal_ensembler = ACTTemporalEnsembler(config.temporal_ensemble_coeff, config.chunk_size)
self.reset()
@@ -122,12 +118,8 @@ class ACTPolicy(PreTrainedPolicy):
batch = self.normalize_inputs(batch)
if self.config.image_features:
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = [
batch[key] for key in self.config.image_features
]
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = [batch[key] for key in self.config.image_features]
# If we are doing temporal ensembling, do online updates where we keep track of the number of actions
# we are ensembling over.
@@ -154,19 +146,14 @@ class ACTPolicy(PreTrainedPolicy):
"""Run the batch through the model and compute the loss for training or validation."""
batch = self.normalize_inputs(batch)
if self.config.image_features:
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = [
batch[key] for key in self.config.image_features
]
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = [batch[key] for key in self.config.image_features]
batch = self.normalize_targets(batch)
actions_hat, (mu_hat, log_sigma_x2_hat) = self.model(batch)
l1_loss = (
F.l1_loss(batch["action"], actions_hat, reduction="none")
* ~batch["action_is_pad"].unsqueeze(-1)
F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
).mean()
loss_dict = {"l1_loss": l1_loss.item()}
@@ -176,12 +163,7 @@ class ACTPolicy(PreTrainedPolicy):
# KL-divergence per batch element, then take the mean over the batch.
# (See App. B of https://arxiv.org/abs/1312.6114 for more details).
mean_kld = (
(
-0.5
* (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())
)
.sum(-1)
.mean()
(-0.5 * (1 + log_sigma_x2_hat - mu_hat.pow(2) - (log_sigma_x2_hat).exp())).sum(-1).mean()
)
loss_dict["kld_loss"] = mean_kld.item()
loss = l1_loss + mean_kld * self.config.kl_weight
@@ -235,9 +217,7 @@ class ACTTemporalEnsembler:
```
"""
self.chunk_size = chunk_size
self.ensemble_weights = torch.exp(
-temporal_ensemble_coeff * torch.arange(chunk_size)
)
self.ensemble_weights = torch.exp(-temporal_ensemble_coeff * torch.arange(chunk_size))
self.ensemble_weights_cumsum = torch.cumsum(self.ensemble_weights, dim=0)
self.reset()
@@ -253,9 +233,7 @@ class ACTTemporalEnsembler:
time steps, and pop/return the next batch of actions in the sequence.
"""
self.ensemble_weights = self.ensemble_weights.to(device=actions.device)
self.ensemble_weights_cumsum = self.ensemble_weights_cumsum.to(
device=actions.device
)
self.ensemble_weights_cumsum = self.ensemble_weights_cumsum.to(device=actions.device)
if self.ensembled_actions is None:
# Initializes `self._ensembled_action` to the sequence of actions predicted during the first
# time step of the episode.
@@ -270,22 +248,12 @@ class ACTTemporalEnsembler:
else:
# self.ensembled_actions will have shape (batch_size, chunk_size - 1, action_dim). Compute
# the online update for those entries.
self.ensembled_actions *= self.ensemble_weights_cumsum[
self.ensembled_actions_count - 1
]
self.ensembled_actions += (
actions[:, :-1] * self.ensemble_weights[self.ensembled_actions_count]
)
self.ensembled_actions /= self.ensemble_weights_cumsum[
self.ensembled_actions_count
]
self.ensembled_actions_count = torch.clamp(
self.ensembled_actions_count + 1, max=self.chunk_size
)
self.ensembled_actions *= self.ensemble_weights_cumsum[self.ensembled_actions_count - 1]
self.ensembled_actions += actions[:, :-1] * self.ensemble_weights[self.ensembled_actions_count]
self.ensembled_actions /= self.ensemble_weights_cumsum[self.ensembled_actions_count]
self.ensembled_actions_count = torch.clamp(self.ensembled_actions_count + 1, max=self.chunk_size)
# The last action, which has no prior online average, needs to get concatenated onto the end.
self.ensembled_actions = torch.cat(
[self.ensembled_actions, actions[:, -1:]], dim=1
)
self.ensembled_actions = torch.cat([self.ensembled_actions, actions[:, -1:]], dim=1)
self.ensembled_actions_count = torch.cat(
[
self.ensembled_actions_count,
@@ -356,9 +324,7 @@ class ACT(nn.Module):
config.dim_model,
)
# Projection layer from the VAE encoder's output to the latent distribution's parameter space.
self.vae_encoder_latent_output_proj = nn.Linear(
config.dim_model, config.latent_dim * 2
)
self.vae_encoder_latent_output_proj = nn.Linear(config.dim_model, config.latent_dim * 2)
# Fixed sinusoidal positional embedding for the input to the VAE encoder. Unsqueeze for batch
# dimension.
num_input_token_encoder = 1 + config.chunk_size
@@ -366,9 +332,7 @@ class ACT(nn.Module):
num_input_token_encoder += 1
self.register_buffer(
"vae_encoder_pos_enc",
create_sinusoidal_pos_embedding(
num_input_token_encoder, config.dim_model
).unsqueeze(0),
create_sinusoidal_pos_embedding(num_input_token_encoder, config.dim_model).unsqueeze(0),
)
# Backbone for image feature extraction.
@@ -385,9 +349,7 @@ class ACT(nn.Module):
# Note: The assumption here is that we are using a ResNet model (and hence layer4 is the final
# feature map).
# Note: The forward method of this returns a dict: {"feature_map": output}.
self.backbone = IntermediateLayerGetter(
backbone_model, return_layers={"layer4": "feature_map"}
)
self.backbone = IntermediateLayerGetter(backbone_model, return_layers={"layer4": "feature_map"})
# Transformer (acts as VAE decoder when training with the variational objective).
self.encoder = ACTEncoder(config)
@@ -416,18 +378,14 @@ class ACT(nn.Module):
n_1d_tokens += 1
self.encoder_1d_feature_pos_embed = nn.Embedding(n_1d_tokens, config.dim_model)
if self.config.image_features:
self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(
config.dim_model // 2
)
self.encoder_cam_feat_pos_embed = ACTSinusoidalPositionEmbedding2d(config.dim_model // 2)
# Transformer decoder.
# Learnable positional embedding for the transformer's decoder (in the style of DETR object queries).
self.decoder_pos_embed = nn.Embedding(config.chunk_size, config.dim_model)
# Final action regression head on the output of the transformer's decoder.
self.action_head = nn.Linear(
config.dim_model, self.config.action_feature.shape[0]
)
self.action_head = nn.Linear(config.dim_model, self.config.action_feature.shape[0])
self._reset_parameters()
@@ -437,9 +395,7 @@ class ACT(nn.Module):
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def forward(
self, batch: dict[str, Tensor]
) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, tuple[Tensor, Tensor] | tuple[None, None]]:
"""A forward pass through the Action Chunking Transformer (with optional VAE encoder).
`batch` should have the following structure:
@@ -475,13 +431,9 @@ class ACT(nn.Module):
self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
) # (B, 1, D)
if self.config.robot_state_feature:
robot_state_embed = self.vae_encoder_robot_state_input_proj(
batch["observation.state"]
)
robot_state_embed = self.vae_encoder_robot_state_input_proj(batch["observation.state"])
robot_state_embed = robot_state_embed.unsqueeze(1) # (B, 1, D)
action_embed = self.vae_encoder_action_input_proj(
batch["action"]
) # (B, S, D)
action_embed = self.vae_encoder_action_input_proj(batch["action"]) # (B, S, D)
if self.config.robot_state_feature:
vae_encoder_input = [
@@ -526,26 +478,20 @@ class ACT(nn.Module):
# When not using the VAE encoder, we set the latent to be all zeros.
mu = log_sigma_x2 = None
# TODO(rcadene, alexander-soare): remove call to `.to` to speedup forward ; precompute and use buffer
latent_sample = torch.zeros(
[batch_size, self.config.latent_dim], dtype=torch.float32
).to(batch["observation.state"].device)
latent_sample = torch.zeros([batch_size, self.config.latent_dim], dtype=torch.float32).to(
batch["observation.state"].device
)
# Prepare transformer encoder inputs.
encoder_in_tokens = [self.encoder_latent_input_proj(latent_sample)]
encoder_in_pos_embed = list(
self.encoder_1d_feature_pos_embed.weight.unsqueeze(1)
)
encoder_in_pos_embed = list(self.encoder_1d_feature_pos_embed.weight.unsqueeze(1))
# Robot state token.
if self.config.robot_state_feature:
encoder_in_tokens.append(
self.encoder_robot_state_input_proj(batch["observation.state"])
)
encoder_in_tokens.append(self.encoder_robot_state_input_proj(batch["observation.state"]))
# Environment state token.
if self.config.env_state_feature:
encoder_in_tokens.append(
self.encoder_env_state_input_proj(
batch["observation.environment_state"]
)
self.encoder_env_state_input_proj(batch["observation.environment_state"])
)
# Camera observation features and positional embeddings.
@@ -556,9 +502,7 @@ class ACT(nn.Module):
# For a list of images, the H and W may vary but H*W is constant.
for img in batch["observation.images"]:
cam_features = self.backbone(img)["feature_map"]
cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(
dtype=cam_features.dtype
)
cam_pos_embed = self.encoder_cam_feat_pos_embed(cam_features).to(dtype=cam_features.dtype)
cam_features = self.encoder_img_feat_input_proj(cam_features)
# Rearrange features to (sequence, batch, dim).
@@ -604,14 +548,8 @@ class ACTEncoder(nn.Module):
def __init__(self, config: ACTConfig, is_vae_encoder: bool = False):
super().__init__()
self.is_vae_encoder = is_vae_encoder
num_layers = (
config.n_vae_encoder_layers
if self.is_vae_encoder
else config.n_encoder_layers
)
self.layers = nn.ModuleList(
[ACTEncoderLayer(config) for _ in range(num_layers)]
)
num_layers = config.n_vae_encoder_layers if self.is_vae_encoder else config.n_encoder_layers
self.layers = nn.ModuleList([ACTEncoderLayer(config) for _ in range(num_layers)])
self.norm = nn.LayerNorm(config.dim_model) if config.pre_norm else nn.Identity()
def forward(
@@ -629,9 +567,7 @@ class ACTEncoder(nn.Module):
class ACTEncoderLayer(nn.Module):
def __init__(self, config: ACTConfig):
super().__init__()
self.self_attn = nn.MultiheadAttention(
config.dim_model, config.n_heads, dropout=config.dropout
)
self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
# Feed forward layers.
self.linear1 = nn.Linear(config.dim_model, config.dim_feedforward)
@@ -646,9 +582,7 @@ class ACTEncoderLayer(nn.Module):
self.activation = get_activation_fn(config.feedforward_activation)
self.pre_norm = config.pre_norm
def forward(
self, x, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None
) -> Tensor:
def forward(self, x, pos_embed: Tensor | None = None, key_padding_mask: Tensor | None = None) -> Tensor:
skip = x
if self.pre_norm:
x = self.norm1(x)
@@ -673,9 +607,7 @@ class ACTDecoder(nn.Module):
def __init__(self, config: ACTConfig):
"""Convenience module for running multiple decoder layers followed by normalization."""
super().__init__()
self.layers = nn.ModuleList(
[ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)]
)
self.layers = nn.ModuleList([ACTDecoderLayer(config) for _ in range(config.n_decoder_layers)])
self.norm = nn.LayerNorm(config.dim_model)
def forward(
@@ -700,12 +632,8 @@ class ACTDecoder(nn.Module):
class ACTDecoderLayer(nn.Module):
def __init__(self, config: ACTConfig):
super().__init__()
self.self_attn = nn.MultiheadAttention(
config.dim_model, config.n_heads, dropout=config.dropout
)
self.multihead_attn = nn.MultiheadAttention(
config.dim_model, config.n_heads, dropout=config.dropout
)
self.self_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
self.multihead_attn = nn.MultiheadAttention(config.dim_model, config.n_heads, dropout=config.dropout)
# Feed forward layers.
self.linear1 = nn.Linear(config.dim_model, config.dim_feedforward)
@@ -746,9 +674,7 @@ class ACTDecoderLayer(nn.Module):
if self.pre_norm:
x = self.norm1(x)
q = k = self.maybe_add_pos_embed(x, decoder_pos_embed)
x = self.self_attn(q, k, value=x)[
0
] # select just the output, not the attention weights
x = self.self_attn(q, k, value=x)[0] # select just the output, not the attention weights
x = skip + self.dropout1(x)
if self.pre_norm:
skip = x
@@ -785,14 +711,9 @@ def create_sinusoidal_pos_embedding(num_positions: int, dimension: int) -> Tenso
"""
def get_position_angle_vec(position):
return [
position / np.power(10000, 2 * (hid_j // 2) / dimension)
for hid_j in range(dimension)
]
return [position / np.power(10000, 2 * (hid_j // 2) / dimension) for hid_j in range(dimension)]
sinusoid_table = np.array(
[get_position_angle_vec(pos_i) for pos_i in range(num_positions)]
)
sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(num_positions)])
sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2]) # dim 2i
sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2]) # dim 2i+1
return torch.from_numpy(sinusoid_table).float()
@@ -837,9 +758,7 @@ class ACTSinusoidalPositionEmbedding2d(nn.Module):
x_range = x_range / (x_range[:, :, -1:] + self._eps) * self._two_pi
inverse_frequency = self._temperature ** (
2
* (torch.arange(self.dimension, dtype=torch.float32, device=x.device) // 2)
/ self.dimension
2 * (torch.arange(self.dimension, dtype=torch.float32, device=x.device) // 2) / self.dimension
)
x_range = x_range.unsqueeze(-1) / inverse_frequency # (1, H, W, 1)
@@ -847,15 +766,9 @@ class ACTSinusoidalPositionEmbedding2d(nn.Module):
# Note: this stack then flatten operation results in interleaved sine and cosine terms.
# pos_embed_x and pos_embed_y are (1, H, W, C // 2).
pos_embed_x = torch.stack(
(x_range[..., 0::2].sin(), x_range[..., 1::2].cos()), dim=-1
).flatten(3)
pos_embed_y = torch.stack(
(y_range[..., 0::2].sin(), y_range[..., 1::2].cos()), dim=-1
).flatten(3)
pos_embed = torch.cat((pos_embed_y, pos_embed_x), dim=3).permute(
0, 3, 1, 2
) # (1, C, H, W)
pos_embed_x = torch.stack((x_range[..., 0::2].sin(), x_range[..., 1::2].cos()), dim=-1).flatten(3)
pos_embed_y = torch.stack((y_range[..., 0::2].sin(), y_range[..., 1::2].cos()), dim=-1).flatten(3)
pos_embed = torch.cat((pos_embed_y, pos_embed_x), dim=3).permute(0, 3, 1, 2) # (1, C, H, W)
return pos_embed

9
lerobot/common/policies/diffusion/configuration_diffusion.py
View File

@@ -205,16 +205,11 @@ class DiffusionConfig(PreTrainedConfig):
def validate_features(self) -> None:
if len(self.image_features) == 0 and self.env_state_feature is None:
raise ValueError(
"You must provide at least one image or the environment state among the inputs."
)
raise ValueError("You must provide at least one image or the environment state among the inputs.")
if self.crop_shape is not None:
for key, image_ft in self.image_features.items():
if (
self.crop_shape[0] > image_ft.shape[1]
or self.crop_shape[1] > image_ft.shape[2]
):
if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
raise ValueError(
f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
f"for `crop_shape` and {image_ft.shape} for "

128
lerobot/common/policies/diffusion/modeling_diffusion.py
View File

@@ -70,9 +70,7 @@ class DiffusionPolicy(PreTrainedPolicy):
config.validate_features()
self.config = config
self.normalize_inputs = Normalize(
config.input_features, config.normalization_mapping, dataset_stats
)
self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
self.normalize_targets = Normalize(
config.output_features, config.normalization_mapping, dataset_stats
)
@@ -99,9 +97,7 @@ class DiffusionPolicy(PreTrainedPolicy):
if self.config.image_features:
self._queues["observation.images"] = deque(maxlen=self.config.n_obs_steps)
if self.config.env_state_feature:
self._queues["observation.environment_state"] = deque(
maxlen=self.config.n_obs_steps
)
self._queues["observation.environment_state"] = deque(maxlen=self.config.n_obs_steps)
@torch.no_grad
def select_action(self, batch: dict[str, Tensor]) -> Tensor:
@@ -127,9 +123,7 @@ class DiffusionPolicy(PreTrainedPolicy):
"""
batch = self.normalize_inputs(batch)
if self.config.image_features:
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack(
[batch[key] for key in self.config.image_features], dim=-4
)
@@ -138,11 +132,7 @@ class DiffusionPolicy(PreTrainedPolicy):
if len(self._queues["action"]) == 0:
# stack n latest observations from the queue
batch = {
k: torch.stack(list(self._queues[k]), dim=1)
for k in batch
if k in self._queues
}
batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
actions = self.diffusion.generate_actions(batch)
# TODO(rcadene): make above methods return output dictionary?
@@ -157,9 +147,7 @@ class DiffusionPolicy(PreTrainedPolicy):
"""Run the batch through the model and compute the loss for training or validation."""
batch = self.normalize_inputs(batch)
if self.config.image_features:
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack(
[batch[key] for key in self.config.image_features], dim=-4
)
@@ -201,9 +189,7 @@ class DiffusionModel(nn.Module):
if self.config.env_state_feature:
global_cond_dim += self.config.env_state_feature.shape[0]
self.unet = DiffusionConditionalUnet1d(
config, global_cond_dim=global_cond_dim * config.n_obs_steps
)
self.unet = DiffusionConditionalUnet1d(config, global_cond_dim=global_cond_dim * config.n_obs_steps)
self.noise_scheduler = _make_noise_scheduler(
config.noise_scheduler_type,
@@ -249,9 +235,7 @@ class DiffusionModel(nn.Module):
global_cond=global_cond,
)
# Compute previous image: x_t -> x_t-1
sample = self.noise_scheduler.step(
model_output, t, sample, generator=generator
).prev_sample
sample = self.noise_scheduler.step(model_output, t, sample, generator=generator).prev_sample
return sample
@@ -263,15 +247,11 @@ class DiffusionModel(nn.Module):
if self.config.image_features:
if self.config.use_separate_rgb_encoder_per_camera:
# Combine batch and sequence dims while rearranging to make the camera index dimension first.
images_per_camera = einops.rearrange(
batch["observation.images"], "b s n ... -> n (b s) ..."
)
images_per_camera = einops.rearrange(batch["observation.images"], "b s n ... -> n (b s) ...")
img_features_list = torch.cat(
[
encoder(images)
for encoder, images in zip(
self.rgb_encoder, images_per_camera, strict=True
)
for encoder, images in zip(self.rgb_encoder, images_per_camera, strict=True)
]
)
# Separate batch and sequence dims back out. The camera index dim gets absorbed into the
@@ -285,9 +265,7 @@ class DiffusionModel(nn.Module):
else:
# Combine batch, sequence, and "which camera" dims before passing to shared encoder.
img_features = self.rgb_encoder(
einops.rearrange(
batch["observation.images"], "b s n ... -> (b s n) ..."
)
einops.rearrange(batch["observation.images"], "b s n ... -> (b s n) ...")
)
# Separate batch dim and sequence dim back out. The camera index dim gets absorbed into the
# feature dim (effectively concatenating the camera features).
@@ -381,9 +359,7 @@ class DiffusionModel(nn.Module):
elif self.config.prediction_type == "sample":
target = batch["action"]
else:
raise ValueError(
f"Unsupported prediction type {self.config.prediction_type}"
)
raise ValueError(f"Unsupported prediction type {self.config.prediction_type}")
loss = F.mse_loss(pred, target, reduction="none")
@@ -443,9 +419,7 @@ class SpatialSoftmax(nn.Module):
# we could use torch.linspace directly but that seems to behave slightly differently than numpy
# and causes a small degradation in pc_success of pre-trained models.
pos_x, pos_y = np.meshgrid(
np.linspace(-1.0, 1.0, self._in_w), np.linspace(-1.0, 1.0, self._in_h)
)
pos_x, pos_y = np.meshgrid(np.linspace(-1.0, 1.0, self._in_w), np.linspace(-1.0, 1.0, self._in_h))
pos_x = torch.from_numpy(pos_x.reshape(self._in_h * self._in_w, 1)).float()
pos_y = torch.from_numpy(pos_y.reshape(self._in_h * self._in_w, 1)).float()
# register as buffer so it's moved to the correct device.
@@ -487,9 +461,7 @@ class DiffusionRgbEncoder(nn.Module):
# Always use center crop for eval
self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
if config.crop_is_random:
self.maybe_random_crop = torchvision.transforms.RandomCrop(
config.crop_shape
)
self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
else:
self.maybe_random_crop = self.center_crop
else:
@@ -510,9 +482,7 @@ class DiffusionRgbEncoder(nn.Module):
self.backbone = _replace_submodules(
root_module=self.backbone,
predicate=lambda x: isinstance(x, nn.BatchNorm2d),
func=lambda x: nn.GroupNorm(
num_groups=x.num_features // 16, num_channels=x.num_features
),
func=lambda x: nn.GroupNorm(num_groups=x.num_features // 16, num_channels=x.num_features),
)
# Set up pooling and final layers.
@@ -523,15 +493,11 @@ class DiffusionRgbEncoder(nn.Module):
# Note: we have a check in the config class to make sure all images have the same shape.
images_shape = next(iter(config.image_features.values())).shape
dummy_shape_h_w = (
config.crop_shape if config.crop_shape is not None else images_shape[1:]
)
dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
self.pool = SpatialSoftmax(
feature_map_shape, num_kp=config.spatial_softmax_num_keypoints
)
self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
self.feature_dim = config.spatial_softmax_num_keypoints * 2
self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
self.relu = nn.ReLU()
@@ -573,11 +539,7 @@ def _replace_submodules(
if predicate(root_module):
return func(root_module)
replace_list = [
k.split(".")
for k, m in root_module.named_modules(remove_duplicate=True)
if predicate(m)
]
replace_list = [k.split(".") for k, m in root_module.named_modules(remove_duplicate=True) if predicate(m)]
for *parents, k in replace_list:
parent_module = root_module
if len(parents) > 0:
@@ -592,9 +554,7 @@ def _replace_submodules(
else:
setattr(parent_module, k, tgt_module)
# verify that all BN are replaced
assert not any(
predicate(m) for _, m in root_module.named_modules(remove_duplicate=True)
)
assert not any(predicate(m) for _, m in root_module.named_modules(remove_duplicate=True))
return root_module
@@ -622,9 +582,7 @@ class DiffusionConv1dBlock(nn.Module):
super().__init__()
self.block = nn.Sequential(
nn.Conv1d(
inp_channels, out_channels, kernel_size, padding=kernel_size // 2
),
nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2),
nn.GroupNorm(n_groups, out_channels),
nn.Mish(),
)
@@ -647,13 +605,9 @@ class DiffusionConditionalUnet1d(nn.Module):
# Encoder for the diffusion timestep.
self.diffusion_step_encoder = nn.Sequential(
DiffusionSinusoidalPosEmb(config.diffusion_step_embed_dim),
nn.Linear(
config.diffusion_step_embed_dim, config.diffusion_step_embed_dim * 4
),
nn.Linear(config.diffusion_step_embed_dim, config.diffusion_step_embed_dim * 4),
nn.Mish(),
nn.Linear(
config.diffusion_step_embed_dim * 4, config.diffusion_step_embed_dim
),
nn.Linear(config.diffusion_step_embed_dim * 4, config.diffusion_step_embed_dim),
)
# The FiLM conditioning dimension.
@@ -678,16 +632,10 @@ class DiffusionConditionalUnet1d(nn.Module):
self.down_modules.append(
nn.ModuleList(
[
DiffusionConditionalResidualBlock1d(
dim_in, dim_out, **common_res_block_kwargs
),
DiffusionConditionalResidualBlock1d(
dim_out, dim_out, **common_res_block_kwargs
),
DiffusionConditionalResidualBlock1d(dim_in, dim_out, **common_res_block_kwargs),
DiffusionConditionalResidualBlock1d(dim_out, dim_out, **common_res_block_kwargs),
# Downsample as long as it is not the last block.
nn.Conv1d(dim_out, dim_out, 3, 2, 1)
if not is_last
else nn.Identity(),
nn.Conv1d(dim_out, dim_out, 3, 2, 1) if not is_last else nn.Identity(),
]
)
)
@@ -716,24 +664,16 @@ class DiffusionConditionalUnet1d(nn.Module):
nn.ModuleList(
[
# dim_in * 2, because it takes the encoder's skip connection as well
DiffusionConditionalResidualBlock1d(
dim_in * 2, dim_out, **common_res_block_kwargs
),
DiffusionConditionalResidualBlock1d(
dim_out, dim_out, **common_res_block_kwargs
),
DiffusionConditionalResidualBlock1d(dim_in * 2, dim_out, **common_res_block_kwargs),
DiffusionConditionalResidualBlock1d(dim_out, dim_out, **common_res_block_kwargs),
# Upsample as long as it is not the last block.
nn.ConvTranspose1d(dim_out, dim_out, 4, 2, 1)
if not is_last
else nn.Identity(),
nn.ConvTranspose1d(dim_out, dim_out, 4, 2, 1) if not is_last else nn.Identity(),
]
)
)
self.final_conv = nn.Sequential(
DiffusionConv1dBlock(
config.down_dims[0], config.down_dims[0], kernel_size=config.kernel_size
),
DiffusionConv1dBlock(config.down_dims[0], config.down_dims[0], kernel_size=config.kernel_size),
nn.Conv1d(config.down_dims[0], config.action_feature.shape[0], 1),
)
@@ -801,23 +741,17 @@ class DiffusionConditionalResidualBlock1d(nn.Module):
self.use_film_scale_modulation = use_film_scale_modulation
self.out_channels = out_channels
self.conv1 = DiffusionConv1dBlock(
in_channels, out_channels, kernel_size, n_groups=n_groups
)
self.conv1 = DiffusionConv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups)
# FiLM modulation (https://arxiv.org/abs/1709.07871) outputs per-channel bias and (maybe) scale.
cond_channels = out_channels * 2 if use_film_scale_modulation else out_channels
self.cond_encoder = nn.Sequential(nn.Mish(), nn.Linear(cond_dim, cond_channels))
self.conv2 = DiffusionConv1dBlock(
out_channels, out_channels, kernel_size, n_groups=n_groups
)
self.conv2 = DiffusionConv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups)
# A final convolution for dimension matching the residual (if needed).
self.residual_conv = (
nn.Conv1d(in_channels, out_channels, 1)
if in_channels != out_channels
else nn.Identity()
nn.Conv1d(in_channels, out_channels, 1) if in_channels != out_channels else nn.Identity()
)
def forward(self, x: Tensor, cond: Tensor) -> Tensor:

12
lerobot/common/policies/factory.py
View File

@@ -104,9 +104,7 @@ def make_policy(
PreTrainedPolicy: _description_
"""
if bool(ds_meta) == bool(env_cfg):
raise ValueError(
"Either one of a dataset metadata or a sim env must be provided."
)
raise ValueError("Either one of a dataset metadata or a sim env must be provided.")
# NOTE: Currently, if you try to run vqbet with mps backend, you'll get this error.
# TODO(aliberts, rcadene): Implement a check_backend_compatibility in policies?
@@ -136,12 +134,8 @@ def make_policy(
)
features = env_to_policy_features(env_cfg)
cfg.output_features = {
key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION
}
cfg.input_features = {
key: ft for key, ft in features.items() if key not in cfg.output_features
}
cfg.output_features = {key: ft for key, ft in features.items() if ft.type is FeatureType.ACTION}
cfg.input_features = {key: ft for key, ft in features.items() if key not in cfg.output_features}
kwargs["config"] = cfg
if cfg.pretrained_path:

25
lerobot/common/policies/hilserl/classifier/modeling_classifier.py
View File

@@ -7,9 +7,7 @@ from torch import Tensor, nn
from .configuration_classifier import ClassifierConfig
logging.basicConfig(
level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s"
)
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)
@@ -53,9 +51,7 @@ class Classifier(
super().__init__()
self.config = config
# self.processor = AutoImageProcessor.from_pretrained(self.config.model_name, trust_remote_code=True)
encoder = AutoModel.from_pretrained(
self.config.model_name, trust_remote_code=True
)
encoder = AutoModel.from_pretrained(self.config.model_name, trust_remote_code=True)
# Extract vision model if we're given a multimodal model
if hasattr(encoder, "vision_model"):
logging.info("Multimodal model detected - using vision encoder only")
@@ -81,9 +77,7 @@ class Classifier(
self.feature_dim = self.encoder.fc.in_features
self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
elif hasattr(self.encoder.config, "hidden_sizes"):
self.feature_dim = self.encoder.config.hidden_sizes[
-1
] # Last channel dimension
self.feature_dim = self.encoder.config.hidden_sizes[-1] # Last channel dimension
else:
raise ValueError("Unsupported CNN architecture")
@@ -103,9 +97,7 @@ class Classifier(
if hasattr(self.encoder.config, "hidden_size"):
input_dim = self.encoder.config.hidden_size
else:
raise ValueError(
"Unsupported transformer architecture since hidden_size is not found"
)
raise ValueError("Unsupported transformer architecture since hidden_size is not found")
self.classifier_head = nn.Sequential(
nn.Linear(input_dim * self.config.num_cameras, self.config.hidden_dim),
@@ -141,10 +133,7 @@ class Classifier(
return features
else: # Transformer models
outputs = self.encoder(processed)
if (
hasattr(outputs, "pooler_output")
and outputs.pooler_output is not None
):
if hasattr(outputs, "pooler_output") and outputs.pooler_output is not None:
return outputs.pooler_output
return outputs.last_hidden_state[:, 0, :]
@@ -160,9 +149,7 @@ class Classifier(
else:
probabilities = torch.softmax(logits, dim=-1)
return ClassifierOutput(
logits=logits, probabilities=probabilities, hidden_states=encoder_outputs
)
return ClassifierOutput(logits=logits, probabilities=probabilities, hidden_states=encoder_outputs)
def predict_reward(self, x, threshold=0.6):
if self.config.num_classes == 2:

36
lerobot/common/policies/normalize.py
View File

@@ -82,43 +82,25 @@ def create_stats_buffers(
if stats:
if isinstance(stats[key]["mean"], np.ndarray):
if norm_mode is NormalizationMode.MEAN_STD:
buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(
dtype=torch.float32
)
buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(
dtype=torch.float32
)
buffer["mean"].data = torch.from_numpy(stats[key]["mean"]).to(dtype=torch.float32)
buffer["std"].data = torch.from_numpy(stats[key]["std"]).to(dtype=torch.float32)
elif norm_mode is NormalizationMode.MIN_MAX:
buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(
dtype=torch.float32
)
buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(
dtype=torch.float32
)
buffer["min"].data = torch.from_numpy(stats[key]["min"]).to(dtype=torch.float32)
buffer["max"].data = torch.from_numpy(stats[key]["max"]).to(dtype=torch.float32)
elif isinstance(stats[key]["mean"], torch.Tensor):
# Note: The clone is needed to make sure that the logic in save_pretrained doesn't see duplicated
# tensors anywhere (for example, when we use the same stats for normalization and
# unnormalization). See the logic here
# https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/py_src/safetensors/torch.py#L97.
if norm_mode is NormalizationMode.MEAN_STD:
buffer["mean"].data = (
stats[key]["mean"].clone().to(dtype=torch.float32)
)
buffer["std"].data = (
stats[key]["std"].clone().to(dtype=torch.float32)
)
buffer["mean"].data = stats[key]["mean"].clone().to(dtype=torch.float32)
buffer["std"].data = stats[key]["std"].clone().to(dtype=torch.float32)
elif norm_mode is NormalizationMode.MIN_MAX:
buffer["min"].data = (
stats[key]["min"].clone().to(dtype=torch.float32)
)
buffer["max"].data = (
stats[key]["max"].clone().to(dtype=torch.float32)
)
buffer["min"].data = stats[key]["min"].clone().to(dtype=torch.float32)
buffer["max"].data = stats[key]["max"].clone().to(dtype=torch.float32)
else:
type_ = type(stats[key]["mean"])
raise ValueError(
f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead."
)
raise ValueError(f"np.ndarray or torch.Tensor expected, but type is '{type_}' instead.")
stats_buffers[key] = buffer
return stats_buffers

8
lerobot/common/policies/pi0/conversion_scripts/compare_with_jax.py
View File

@@ -44,9 +44,7 @@ def main():
else:
dataset_repo_id = "lerobot/aloha_sim_transfer_cube_human"
ckpt_torch_dir = (
Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
)
ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
ckpt_jax_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}"
save_dir = Path(f"../openpi/data/{model_name}/save")
@@ -72,9 +70,7 @@ def main():
# Create LeRobot batch from Jax
batch = {}
for cam_key, uint_chw_array in example["images"].items():
batch[f"observation.images.{cam_key}"] = (
torch.from_numpy(uint_chw_array) / 255.0
)
batch[f"observation.images.{cam_key}"] = torch.from_numpy(uint_chw_array) / 255.0
batch["observation.state"] = torch.from_numpy(example["state"])
batch["action"] = torch.from_numpy(outputs["actions"])
batch["task"] = example["prompt"]

4
lerobot/common/policies/pi0/conversion_scripts/conversion_utils.py
View File

@@ -54,9 +54,7 @@ def get_paligemma_config(precision: str):
"projector_hidden_act": "gelu_fast",
"vision_use_head": False,
}
final_config = PaliGemmaConfig(
text_config=text_config, vision_config=vision_config, **config
)
final_config = PaliGemmaConfig(text_config=text_config, vision_config=vision_config, **config)
return final_config

8
lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py
View File

@@ -322,9 +322,7 @@ def update_keys_with_prefix(d: dict, prefix: str) -> dict:
return {f"{prefix}{key}": value for key, value in d.items()}
def convert_pi0_checkpoint(
checkpoint_dir: str, precision: str, tokenizer_id: str, output_path: str
):
def convert_pi0_checkpoint(checkpoint_dir: str, precision: str, tokenizer_id: str, output_path: str):
# Break down orbax ckpts - they are in OCDBT
initial_params = slice_initial_orbax_checkpoint(checkpoint_dir=checkpoint_dir)
# process projection params
@@ -384,9 +382,7 @@ def convert_pi0_checkpoint(
# gemma_config=gemma_config, paligemma_config=paligemma_config)
pi0_model = PI0Policy(pi0_config)
paligemma_params = update_keys_with_prefix(
paligemma_params, "model.paligemma_with_expert."
)
paligemma_params = update_keys_with_prefix(paligemma_params, "model.paligemma_with_expert.")
gemma_params = update_keys_with_prefix(gemma_params, "model.paligemma_with_expert.")
projection_params = update_keys_with_prefix(projection_params, "model.")

88
lerobot/common/policies/pi0/modeling_pi0.py
View File

@@ -193,9 +193,7 @@ def aloha_gripper_to_angular(value):
# This is the inverse of the angular to linear transformation inside the Interbotix code.
def linear_to_radian(linear_position, arm_length, horn_radius):
value = (horn_radius**2 + linear_position**2 - arm_length**2) / (
2 * horn_radius * linear_position
)
value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
return safe_arcsin(value)
# The constants are taken from the Interbotix code.
@@ -246,9 +244,7 @@ class PI0Policy(PreTrainedPolicy):
super().__init__(config)
config.validate_features()
self.config = config
self.normalize_inputs = Normalize(
config.input_features, config.normalization_mapping, dataset_stats
)
self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
self.normalize_targets = Normalize(
config.output_features, config.normalization_mapping, dataset_stats
)
@@ -256,9 +252,7 @@ class PI0Policy(PreTrainedPolicy):
config.output_features, config.normalization_mapping, dataset_stats
)
self.language_tokenizer = AutoTokenizer.from_pretrained(
"google/paligemma-3b-pt-224"
)
self.language_tokenizer = AutoTokenizer.from_pretrained("google/paligemma-3b-pt-224")
self.model = PI0FlowMatching(config)
self.reset()
@@ -271,9 +265,7 @@ class PI0Policy(PreTrainedPolicy):
return self.parameters()
@torch.no_grad
def select_action(
self, batch: dict[str, Tensor], noise: Tensor | None = None
) -> Tensor:
def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
"""Select a single action given environment observations.
This method wraps `select_actions` in order to return one action at a time for execution in the
@@ -312,9 +304,7 @@ class PI0Policy(PreTrainedPolicy):
self._action_queue.extend(actions.transpose(0, 1))
return self._action_queue.popleft()
def forward(
self, batch: dict[str, Tensor], noise=None, time=None
) -> tuple[Tensor, dict[str, Tensor]]:
def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> tuple[Tensor, dict[str, Tensor]]:
"""Do a full training forward pass to compute the loss"""
if self.config.adapt_to_pi_aloha:
batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
@@ -330,9 +320,7 @@ class PI0Policy(PreTrainedPolicy):
actions_is_pad = batch.get("action_is_pad")
loss_dict = {}
losses = self.model.forward(
images, img_masks, lang_tokens, lang_masks, state, actions, noise, time
)
losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
loss_dict["losses_after_forward"] = losses.clone()
if actions_is_pad is not None:
@@ -359,9 +347,7 @@ class PI0Policy(PreTrainedPolicy):
img_masks = []
present_img_keys = [key for key in self.config.image_features if key in batch]
missing_img_keys = [
key for key in self.config.image_features if key not in batch
]
missing_img_keys = [key for key in self.config.image_features if key not in batch]
if len(present_img_keys) == 0:
raise ValueError(
@@ -373,9 +359,7 @@ class PI0Policy(PreTrainedPolicy):
img = batch[key]
if self.config.resize_imgs_with_padding is not None:
img = resize_with_pad(
img, *self.config.resize_imgs_with_padding, pad_value=0
)
img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
# Normalize from range [0,1] to [-1,1] as expacted by siglip
img = img * 2.0 - 1.0
@@ -414,9 +398,7 @@ class PI0Policy(PreTrainedPolicy):
return_tensors="pt",
)
lang_tokens = tokenized_prompt["input_ids"].to(device=device)
lang_masks = tokenized_prompt["attention_mask"].to(
device=device, dtype=torch.bool
)
lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
return lang_tokens, lang_masks
@@ -435,9 +417,7 @@ class PI0Policy(PreTrainedPolicy):
actions[:, :, motor_idx] *= -1
# Reverse the gripper transformation that is being applied by the Aloha runtime.
for motor_idx in [6, 13]:
actions[:, :, motor_idx] = aloha_gripper_from_angular(
actions[:, :, motor_idx]
)
actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
return actions
def _pi_aloha_encode_actions_inv(self, actions):
@@ -446,9 +426,7 @@ class PI0Policy(PreTrainedPolicy):
actions[:, :, motor_idx] *= -1
# Reverse the gripper transformation that is being applied by the Aloha runtime.
for motor_idx in [6, 13]:
actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(
actions[:, :, motor_idx]
)
actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
return actions
def prepare_state(self, batch):
@@ -498,25 +476,15 @@ class PI0FlowMatching(nn.Module):
train_expert_only=self.config.train_expert_only,
attention_implementation=self.config.attention_implementation,
)
self.paligemma_with_expert = PaliGemmaWithExpertModel(
paligemma_with_export_config
)
self.paligemma_with_expert = PaliGemmaWithExpertModel(paligemma_with_export_config)
# Projections are float32
self.state_proj = nn.Linear(self.config.max_state_dim, self.config.proj_width)
self.action_in_proj = nn.Linear(
self.config.max_action_dim, self.config.proj_width
)
self.action_out_proj = nn.Linear(
self.config.proj_width, self.config.max_action_dim
)
self.action_in_proj = nn.Linear(self.config.max_action_dim, self.config.proj_width)
self.action_out_proj = nn.Linear(self.config.proj_width, self.config.max_action_dim)
self.action_time_mlp_in = nn.Linear(
self.config.proj_width * 2, self.config.proj_width
)
self.action_time_mlp_out = nn.Linear(
self.config.proj_width, self.config.proj_width
)
self.action_time_mlp_in = nn.Linear(self.config.proj_width * 2, self.config.proj_width)
self.action_time_mlp_out = nn.Linear(self.config.proj_width, self.config.proj_width)
self.set_requires_grad()
@@ -560,9 +528,7 @@ class PI0FlowMatching(nn.Module):
# Normalize image embeddings
img_emb_dim = img_emb.shape[-1]
img_emb = img_emb * torch.tensor(
img_emb_dim**0.5, dtype=img_emb.dtype, device=img_emb.device
)
img_emb = img_emb * torch.tensor(img_emb_dim**0.5, dtype=img_emb.dtype, device=img_emb.device)
bsize, num_img_embs = img_emb.shape[:2]
img_mask = img_mask[:, None].expand(bsize, num_img_embs)
@@ -637,9 +603,7 @@ class PI0FlowMatching(nn.Module):
embs.append(action_time_emb)
bsize, action_time_dim = action_time_emb.shape[:2]
action_time_mask = torch.ones(
bsize, action_time_dim, dtype=torch.bool, device=device
)
action_time_mask = torch.ones(bsize, action_time_dim, dtype=torch.bool, device=device)
pad_masks.append(action_time_mask)
# Set attention masks so that image, language and state inputs do not attend to action tokens
@@ -677,9 +641,7 @@ class PI0FlowMatching(nn.Module):
prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
images, img_masks, lang_tokens, lang_masks
)
suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(
state, x_t, time
)
suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(state, x_t, time)
pad_masks = torch.cat([prefix_pad_masks, suffix_pad_masks], dim=1)
att_masks = torch.cat([prefix_att_masks, suffix_att_masks], dim=1)
@@ -703,9 +665,7 @@ class PI0FlowMatching(nn.Module):
losses = F.mse_loss(u_t, v_t, reduction="none")
return losses
def sample_actions(
self, images, img_masks, lang_tokens, lang_masks, state, noise=None
) -> Tensor:
def sample_actions(self, images, img_masks, lang_tokens, lang_masks, state, noise=None) -> Tensor:
"""Do a full inference forward and compute the action (batch_size x num_steps x num_motors)"""
bsize = state.shape[0]
device = state.device
@@ -763,16 +723,12 @@ class PI0FlowMatching(nn.Module):
timestep,
):
"""Apply one denoising step of the noise `x_t` at a given timestep."""
suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(
state, x_t, timestep
)
suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(state, x_t, timestep)
suffix_len = suffix_pad_masks.shape[1]
batch_size = prefix_pad_masks.shape[0]
prefix_len = prefix_pad_masks.shape[1]
prefix_pad_2d_masks = prefix_pad_masks[:, None, :].expand(
batch_size, suffix_len, prefix_len
)
prefix_pad_2d_masks = prefix_pad_masks[:, None, :].expand(batch_size, suffix_len, prefix_len)
suffix_att_2d_masks = make_att_2d_masks(suffix_pad_masks, suffix_att_masks)

28
lerobot/common/policies/pi0/paligemma_with_expert.py
View File

@@ -39,13 +39,9 @@ def apply_rope(x, positions, max_wavelength=10_000):
dtype = x.dtype
x = x.to(torch.float32)
freq_exponents = (2.0 / x.shape[-1]) * torch.arange(
d_half, dtype=torch.float32, device=device
)
freq_exponents = (2.0 / x.shape[-1]) * torch.arange(d_half, dtype=torch.float32, device=device)
timescale = max_wavelength**freq_exponents
radians = positions[..., None].to(torch.float32) / timescale[None, None, :].to(
torch.float32
)
radians = positions[..., None].to(torch.float32) / timescale[None, None, :].to(torch.float32)
radians = radians[..., None, :]
@@ -178,9 +174,7 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
def __init__(self, config: PaliGemmaWithExpertConfig):
super().__init__(config=config)
self.config = config
self.paligemma = PaliGemmaForConditionalGeneration(
config=config.paligemma_config
)
self.paligemma = PaliGemmaForConditionalGeneration(config=config.paligemma_config)
self.gemma_expert = GemmaForCausalLM(config=config.gemma_expert_config)
# Remove unused embed_tokens
self.gemma_expert.model.embed_tokens = None
@@ -297,9 +291,7 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
# so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
# the max len, then we (for instance) double the cache size. This implementation already exists
# in `transformers`. (molbap)
key_states = torch.cat(
[past_key_values[layer_idx]["key_states"], key_states], dim=1
)
key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
value_states = torch.cat(
[past_key_values[layer_idx]["value_states"], value_states],
dim=1,
@@ -392,9 +384,7 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
value_states,
):
num_att_heads = self.config.paligemma_config.text_config.num_attention_heads
num_key_value_heads = (
self.config.paligemma_config.text_config.num_key_value_heads
)
num_key_value_heads = self.config.paligemma_config.text_config.num_key_value_heads
num_key_value_groups = num_att_heads // num_key_value_heads
# query_states: batch_size, sequence_length, num_att_head, head_dim
@@ -442,9 +432,7 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
att_weights *= head_dim**-0.5
big_neg = -2.3819763e38 # See gemma/modules.py
masked_att_weights = torch.where(
attention_mask[:, None, :, :], att_weights, big_neg
)
masked_att_weights = torch.where(attention_mask[:, None, :, :], att_weights, big_neg)
probs = nn.functional.softmax(masked_att_weights, dim=-1)
probs = probs.to(dtype=value_states.dtype)
@@ -456,8 +444,6 @@ class PaliGemmaWithExpertModel(PreTrainedModel):
att_output = att_output.permute(0, 2, 1, 3)
# we use -1 because sequence length can change
att_output = att_output.reshape(
batch_size, -1, num_key_value_heads * num_key_value_groups * head_dim
)
att_output = att_output.reshape(batch_size, -1, num_key_value_heads * num_key_value_groups * head_dim)
return att_output

24
lerobot/common/policies/pretrained.py
View File

@@ -71,9 +71,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
def _save_pretrained(self, save_directory: Path) -> None:
self.config._save_pretrained(save_directory)
model_to_save = self.module if hasattr(self, "module") else self
save_model_as_safetensor(
model_to_save, str(save_directory / SAFETENSORS_SINGLE_FILE)
)
save_model_as_safetensor(model_to_save, str(save_directory / SAFETENSORS_SINGLE_FILE))
@classmethod
def from_pretrained(
@@ -112,9 +110,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
if os.path.isdir(model_id):
print("Loading weights from local directory")
model_file = os.path.join(model_id, SAFETENSORS_SINGLE_FILE)
policy = cls._load_as_safetensor(
instance, model_file, config.device, strict
)
policy = cls._load_as_safetensor(instance, model_file, config.device, strict)
else:
try:
model_file = hf_hub_download(
@@ -128,9 +124,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
token=token,
local_files_only=local_files_only,
)
policy = cls._load_as_safetensor(
instance, model_file, config.device, strict
)
policy = cls._load_as_safetensor(instance, model_file, config.device, strict)
except HfHubHTTPError as e:
raise FileNotFoundError(
f"{SAFETENSORS_SINGLE_FILE} not found on the HuggingFace Hub in {model_id}"
@@ -141,12 +135,8 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
return policy
@classmethod
def _load_as_safetensor(
cls, model: T, model_file: str, map_location: str, strict: bool
) -> T:
if packaging.version.parse(safetensors.__version__) < packaging.version.parse(
"0.4.3"
):
def _load_as_safetensor(cls, model: T, model_file: str, map_location: str, strict: bool) -> T:
if packaging.version.parse(safetensors.__version__) < packaging.version.parse("0.4.3"):
load_model_as_safetensor(model, model_file, strict=strict)
if map_location != "cpu":
logging.warning(
@@ -157,9 +147,7 @@ class PreTrainedPolicy(nn.Module, HubMixin, abc.ABC):
)
model.to(map_location)
else:
safetensors.torch.load_model(
model, model_file, strict=strict, device=map_location
)
safetensors.torch.load_model(model, model_file, strict=strict, device=map_location)
return model
# def generate_model_card(self, *args, **kwargs) -> ModelCard:

4
lerobot/common/policies/sac/configuration_sac.py
View File

@@ -48,9 +48,7 @@ class SACConfig:
"observation.state": {"min": [-1, -1, -1, -1], "max": [1, 1, 1, 1]},
}
)
output_normalization_modes: dict[str, str] = field(
default_factory=lambda: {"action": "min_max"}
)
output_normalization_modes: dict[str, str] = field(default_factory=lambda: {"action": "min_max"})
output_normalization_params: dict[str, dict[str, list[float]]] = field(
default_factory=lambda: {
"action": {"min": [-1, -1], "max": [1, 1]},

109
lerobot/common/policies/sac/modeling_sac.py
View File

@@ -18,8 +18,8 @@
# TODO: (1) better device management
import math
from typing import Callable, Optional, Tuple, Union, Dict, List
from pathlib import Path
from typing import Callable, Dict, List, Optional, Tuple, Union
import einops
import numpy as np
@@ -124,17 +124,13 @@ class SACPolicy(
self.actor = Policy(
encoder=encoder_actor,
network=MLP(
input_dim=encoder_actor.output_dim, **config.actor_network_kwargs
),
network=MLP(input_dim=encoder_actor.output_dim, **config.actor_network_kwargs),
action_dim=config.output_shapes["action"][0],
encoder_is_shared=config.shared_encoder,
**config.policy_kwargs,
)
if config.target_entropy is None:
config.target_entropy = (
-np.prod(config.output_shapes["action"][0]) / 2
) # (-dim(A)/2)
config.target_entropy = -np.prod(config.output_shapes["action"][0]) / 2 # (-dim(A)/2)
# TODO (azouitine): Handle the case where the temparameter is a fixed
# TODO (michel-aractingi): Put the log_alpha in cuda by default because otherwise
@@ -146,10 +142,11 @@ class SACPolicy(
def _save_pretrained(self, save_directory):
"""Custom save method to handle TensorDict properly"""
import os
import json
import os
from dataclasses import asdict
from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE, CONFIG_NAME
from huggingface_hub.constants import CONFIG_NAME, SAFETENSORS_SINGLE_FILE
from safetensors.torch import save_model
save_model(self, os.path.join(save_directory, SAFETENSORS_SINGLE_FILE))
@@ -177,12 +174,14 @@ class SACPolicy(
**model_kwargs,
) -> "SACPolicy":
"""Custom load method to handle loading SAC policy from saved files"""
import os
import json
import os
from pathlib import Path
from huggingface_hub import hf_hub_download
from huggingface_hub.constants import SAFETENSORS_SINGLE_FILE, CONFIG_NAME
from huggingface_hub.constants import CONFIG_NAME, SAFETENSORS_SINGLE_FILE
from safetensors.torch import load_model
from lerobot.common.policies.sac.configuration_sac import SACConfig
# Check if model_id is a local path or a hub model ID
@@ -302,14 +301,10 @@ class SACPolicy(
) -> Tensor:
self.temperature = self.log_alpha.exp().item()
with torch.no_grad():
next_action_preds, next_log_probs, _ = self.actor(
next_observations, next_observation_features
)
next_action_preds, next_log_probs, _ = self.actor(next_observations, next_observation_features)
# TODO: (maractingi, azouitine) This is to slow, we should find a way to do this in a more efficient way
next_action_preds = self.unnormalize_outputs({"action": next_action_preds})[
"action"
]
next_action_preds = self.unnormalize_outputs({"action": next_action_preds})["action"]
# 2- compute q targets
q_targets = self.critic_forward(
@@ -353,21 +348,15 @@ class SACPolicy(
).sum()
return critics_loss
def compute_loss_temperature(
self, observations, observation_features: Tensor | None = None
) -> Tensor:
def compute_loss_temperature(self, observations, observation_features: Tensor | None = None) -> Tensor:
"""Compute the temperature loss"""
# calculate temperature loss
with torch.no_grad():
_, log_probs, _ = self.actor(observations, observation_features)
temperature_loss = (
-self.log_alpha.exp() * (log_probs + self.config.target_entropy)
).mean()
temperature_loss = (-self.log_alpha.exp() * (log_probs + self.config.target_entropy)).mean()
return temperature_loss
def compute_loss_actor(
self, observations, observation_features: Tensor | None = None
) -> Tensor:
def compute_loss_actor(self, observations, observation_features: Tensor | None = None) -> Tensor:
self.temperature = self.log_alpha.exp().item()
actions_pi, log_probs, _ = self.actor(observations, observation_features)
@@ -408,11 +397,7 @@ class MLP(nn.Module):
if dropout_rate is not None and dropout_rate > 0:
layers.append(nn.Dropout(p=dropout_rate))
layers.append(nn.LayerNorm(hidden_dims[0]))
layers.append(
activations
if isinstance(activations, nn.Module)
else getattr(nn, activations)()
)
layers.append(activations if isinstance(activations, nn.Module) else getattr(nn, activations)())
# Rest of the layers
for i in range(1, len(hidden_dims)):
@@ -424,11 +409,7 @@ class MLP(nn.Module):
layers.append(nn.LayerNorm(hidden_dims[i]))
# If we're at the final layer and a final activation is specified, use it
if (
i + 1 == len(hidden_dims)
and activate_final
and final_activation is not None
):
if i + 1 == len(hidden_dims) and activate_final and final_activation is not None:
layers.append(
final_activation
if isinstance(final_activation, nn.Module)
@@ -436,9 +417,7 @@ class MLP(nn.Module):
)
else:
layers.append(
activations
if isinstance(activations, nn.Module)
else getattr(nn, activations)()
activations if isinstance(activations, nn.Module) else getattr(nn, activations)()
)
self.net = nn.Sequential(*layers)
@@ -639,15 +618,11 @@ class Policy(nn.Module):
# Compute standard deviations
if self.fixed_std is None:
log_std = self.std_layer(outputs)
assert not torch.isnan(log_std).any(), (
"[ERROR] log_std became NaN after std_layer!"
)
assert not torch.isnan(log_std).any(), "[ERROR] log_std became NaN after std_layer!"
if self.use_tanh_squash:
log_std = torch.tanh(log_std)
log_std = self.log_std_min + 0.5 * (
self.log_std_max - self.log_std_min
) * (log_std + 1.0)
log_std = self.log_std_min + 0.5 * (self.log_std_max - self.log_std_min) * (log_std + 1.0)
else:
log_std = torch.clamp(log_std, self.log_std_min, self.log_std_max)
else:
@@ -660,9 +635,7 @@ class Policy(nn.Module):
if self.use_tanh_squash:
actions = torch.tanh(x_t)
log_probs -= torch.log(
(1 - actions.pow(2)) + 1e-6
) # Adjust log-probs for Tanh
log_probs -= torch.log((1 - actions.pow(2)) + 1e-6) # Adjust log-probs for Tanh
else:
actions = x_t # No Tanh; raw Gaussian sample
@@ -709,9 +682,7 @@ class SACObservationEncoder(nn.Module):
freeze_image_encoder(self.image_enc_layers)
else:
self.parameters_to_optimize += list(self.image_enc_layers.parameters())
self.all_image_keys = [
k for k in config.input_shapes if k.startswith("observation.image")
]
self.all_image_keys = [k for k in config.input_shapes if k.startswith("observation.image")]
if "observation.state" in config.input_shapes:
self.state_enc_layers = nn.Sequential(
@@ -738,9 +709,7 @@ class SACObservationEncoder(nn.Module):
self.aggregation_size += config.latent_dim
self.parameters_to_optimize += list(self.env_state_enc_layers.parameters())
self.aggregation_layer = nn.Linear(
in_features=self.aggregation_size, out_features=config.latent_dim
)
self.aggregation_layer = nn.Linear(in_features=self.aggregation_size, out_features=config.latent_dim)
self.parameters_to_optimize += list(self.aggregation_layer.parameters())
def forward(self, obs_dict: dict[str, Tensor]) -> Tensor:
@@ -753,19 +722,13 @@ class SACObservationEncoder(nn.Module):
obs_dict = self.input_normalization(obs_dict)
# Batch all images along the batch dimension, then encode them.
if len(self.all_image_keys) > 0:
images_batched = torch.cat(
[obs_dict[key] for key in self.all_image_keys], dim=0
)
images_batched = torch.cat([obs_dict[key] for key in self.all_image_keys], dim=0)
images_batched = self.image_enc_layers(images_batched)
embeddings_chunks = torch.chunk(
images_batched, dim=0, chunks=len(self.all_image_keys)
)
embeddings_chunks = torch.chunk(images_batched, dim=0, chunks=len(self.all_image_keys))
feat.extend(embeddings_chunks)
if "observation.environment_state" in self.config.input_shapes:
feat.append(
self.env_state_enc_layers(obs_dict["observation.environment_state"])
)
feat.append(self.env_state_enc_layers(obs_dict["observation.environment_state"]))
if "observation.state" in self.config.input_shapes:
feat.append(self.state_enc_layers(obs_dict["observation.state"]))
@@ -833,9 +796,7 @@ class PretrainedImageEncoder(nn.Module):
def __init__(self, config):
super().__init__()
self.image_enc_layers, self.image_enc_out_shape = (
self._load_pretrained_vision_encoder(config)
)
self.image_enc_layers, self.image_enc_out_shape = self._load_pretrained_vision_encoder(config)
self.image_enc_proj = nn.Sequential(
nn.Linear(np.prod(self.image_enc_out_shape), config.latent_dim),
nn.LayerNorm(config.latent_dim),
@@ -846,21 +807,15 @@ class PretrainedImageEncoder(nn.Module):
"""Set up CNN encoder"""
from transformers import AutoModel
self.image_enc_layers = AutoModel.from_pretrained(
config.vision_encoder_name, trust_remote_code=True
)
self.image_enc_layers = AutoModel.from_pretrained(config.vision_encoder_name, trust_remote_code=True)
# self.image_enc_layers.pooler = Identity()
if hasattr(self.image_enc_layers.config, "hidden_sizes"):
self.image_enc_out_shape = self.image_enc_layers.config.hidden_sizes[
-1
] # Last channel dimension
self.image_enc_out_shape = self.image_enc_layers.config.hidden_sizes[-1] # Last channel dimension
elif hasattr(self.image_enc_layers, "fc"):
self.image_enc_out_shape = self.image_enc_layers.fc.in_features
else:
raise ValueError(
"Unsupported vision encoder architecture, make sure you are using a CNN"
)
raise ValueError("Unsupported vision encoder architecture, make sure you are using a CNN")
return self.image_enc_layers, self.image_enc_out_shape
def forward(self, x):
@@ -896,9 +851,7 @@ def _convert_normalization_params_to_tensor(normalization_params: dict) -> dict:
for key, value in inner_dict.items():
converted_params[outer_key][key] = torch.tensor(value)
if "image" in outer_key:
converted_params[outer_key][key] = converted_params[outer_key][
key
].view(3, 1, 1)
converted_params[outer_key][key] = converted_params[outer_key][key].view(3, 1, 1)
return converted_params

12
lerobot/common/policies/tdmpc/configuration_tdmpc.py
View File

@@ -183,13 +183,9 @@ class TDMPCConfig(PreTrainedConfig):
"If `n_action_steps > 1`, `n_action_repeats` must be left to its default value of 1."
)
if not self.use_mpc:
raise ValueError(
"If `n_action_steps > 1`, `use_mpc` must be set to `True`."
)
raise ValueError("If `n_action_steps > 1`, `use_mpc` must be set to `True`.")
if self.n_action_steps > self.horizon:
raise ValueError(
"`n_action_steps` must be less than or equal to `horizon`."
)
raise ValueError("`n_action_steps` must be less than or equal to `horizon`.")
def get_optimizer_preset(self) -> AdamConfig:
return AdamConfig(lr=self.optimizer_lr)
@@ -209,9 +205,7 @@ class TDMPCConfig(PreTrainedConfig):
if image_ft.shape[-2] != image_ft.shape[-1]:
# TODO(alexander-soare): This limitation is solely because of code in the random shift
# augmentation. It should be able to be removed.
raise ValueError(
f"Only square images are handled now. Got image shape {image_ft.shape}."
)
raise ValueError(f"Only square images are handled now. Got image shape {image_ft.shape}.")
@property
def observation_delta_indices(self) -> list:

130
lerobot/common/policies/tdmpc/modeling_tdmpc.py
View File

@@ -83,9 +83,7 @@ class TDMPCPolicy(PreTrainedPolicy):
config.validate_features()
self.config = config
self.normalize_inputs = Normalize(
config.input_features, config.normalization_mapping, dataset_stats
)
self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
self.normalize_targets = Normalize(
config.output_features, config.normalization_mapping, dataset_stats
)
@@ -110,9 +108,7 @@ class TDMPCPolicy(PreTrainedPolicy):
"""
self._queues = {
"observation.state": deque(maxlen=1),
"action": deque(
maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)
),
"action": deque(maxlen=max(self.config.n_action_steps, self.config.n_action_repeats)),
}
if self.config.image_features:
self._queues["observation.image"] = deque(maxlen=1)
@@ -127,9 +123,7 @@ class TDMPCPolicy(PreTrainedPolicy):
"""Select a single action given environment observations."""
batch = self.normalize_inputs(batch)
if self.config.image_features:
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.image"] = batch[next(iter(self.config.image_features))]
self._queues = populate_queues(self._queues, batch)
@@ -232,47 +226,35 @@ class TDMPCPolicy(PreTrainedPolicy):
self.config.action_feature.shape[0],
device=std.device,
)
gaussian_actions = torch.clamp(
mean.unsqueeze(1) + std.unsqueeze(1) * std_normal_noise, -1, 1
)
gaussian_actions = torch.clamp(mean.unsqueeze(1) + std.unsqueeze(1) * std_normal_noise, -1, 1)
# Compute elite actions.
actions = torch.cat([gaussian_actions, pi_actions], dim=1)
value = self.estimate_value(z, actions).nan_to_num_(0)
elite_idxs = torch.topk(
value, self.config.n_elites, dim=0
).indices # (n_elites, batch)
elite_idxs = torch.topk(value, self.config.n_elites, dim=0).indices # (n_elites, batch)
elite_value = value.take_along_dim(elite_idxs, dim=0) # (n_elites, batch)
# (horizon, n_elites, batch, action_dim)
elite_actions = actions.take_along_dim(
einops.rearrange(elite_idxs, "n b -> 1 n b 1"), dim=1
)
elite_actions = actions.take_along_dim(einops.rearrange(elite_idxs, "n b -> 1 n b 1"), dim=1)
# Update gaussian PDF parameters to be the (weighted) mean and standard deviation of the elites.
max_value = elite_value.max(0, keepdim=True)[0] # (1, batch)
# The weighting is a softmax over trajectory values. Note that this is not the same as the usage
# of Ω in eqn 4 of the TD-MPC paper. Instead it is the normalized version of it: s = Ω/ΣΩ. This
# makes the equations: μ = Σ(s⋅Γ), σ = Σ(s⋅(Γ-μ)²).
score = torch.exp(
self.config.elite_weighting_temperature * (elite_value - max_value)
)
score = torch.exp(self.config.elite_weighting_temperature * (elite_value - max_value))
score /= score.sum(axis=0, keepdim=True)
# (horizon, batch, action_dim)
_mean = torch.sum(
einops.rearrange(score, "n b -> n b 1") * elite_actions, dim=1
)
_mean = torch.sum(einops.rearrange(score, "n b -> n b 1") * elite_actions, dim=1)
_std = torch.sqrt(
torch.sum(
einops.rearrange(score, "n b -> n b 1")
* (elite_actions - einops.rearrange(_mean, "h b d -> h 1 b d"))
** 2,
* (elite_actions - einops.rearrange(_mean, "h b d -> h 1 b d")) ** 2,
dim=1,
)
)
# Update mean with an exponential moving average, and std with a direct replacement.
mean = (
self.config.gaussian_mean_momentum * mean
+ (1 - self.config.gaussian_mean_momentum) * _mean
self.config.gaussian_mean_momentum * mean + (1 - self.config.gaussian_mean_momentum) * _mean
)
std = _std.clamp_(self.config.min_std, self.config.max_std)
@@ -281,9 +263,7 @@ class TDMPCPolicy(PreTrainedPolicy):
# Randomly select one of the elite actions from the last iteration of MPPI/CEM using the softmax
# scores from the last iteration.
actions = elite_actions[
:, torch.multinomial(score.T, 1).squeeze(), torch.arange(batch_size)
]
actions = elite_actions[:, torch.multinomial(score.T, 1).squeeze(), torch.arange(batch_size)]
return actions
@@ -306,8 +286,7 @@ class TDMPCPolicy(PreTrainedPolicy):
# of the FOWM paper.
if self.config.uncertainty_regularizer_coeff > 0:
regularization = -(
self.config.uncertainty_regularizer_coeff
* self.model.Qs(z, actions[t]).std(0)
self.config.uncertainty_regularizer_coeff * self.model.Qs(z, actions[t]).std(0)
)
else:
regularization = 0
@@ -328,9 +307,7 @@ class TDMPCPolicy(PreTrainedPolicy):
G += (
running_discount
* torch.min(
terminal_values[
torch.randint(0, self.config.q_ensemble_size, size=(2,))
],
terminal_values[torch.randint(0, self.config.q_ensemble_size, size=(2,))],
dim=0,
)[0]
)
@@ -338,11 +315,7 @@ class TDMPCPolicy(PreTrainedPolicy):
G += running_discount * torch.min(terminal_values, dim=0)[0]
# Finally, also regularize the terminal value.
if self.config.uncertainty_regularizer_coeff > 0:
G -= (
running_discount
* self.config.uncertainty_regularizer_coeff
* terminal_values.std(0)
)
G -= running_discount * self.config.uncertainty_regularizer_coeff * terminal_values.std(0)
return G
def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
@@ -354,9 +327,7 @@ class TDMPCPolicy(PreTrainedPolicy):
batch = self.normalize_inputs(batch)
if self.config.image_features:
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.image"] = batch[next(iter(self.config.image_features))]
batch = self.normalize_targets(batch)
@@ -388,29 +359,21 @@ class TDMPCPolicy(PreTrainedPolicy):
current_observation[k] = observations[k][0]
next_observations[k] = observations[k][1:]
horizon, batch_size = next_observations[
"observation.image"
if self.config.image_features
else "observation.environment_state"
"observation.image" if self.config.image_features else "observation.environment_state"
].shape[:2]
# Run latent rollout using the latent dynamics model and policy model.
# Note this has shape `horizon+1` because there are `horizon` actions and a current `z`. Each action
# gives us a next `z`.
batch_size = batch["index"].shape[0]
z_preds = torch.empty(
horizon + 1, batch_size, self.config.latent_dim, device=device
)
z_preds = torch.empty(horizon + 1, batch_size, self.config.latent_dim, device=device)
z_preds[0] = self.model.encode(current_observation)
reward_preds = torch.empty_like(reward, device=device)
for t in range(horizon):
z_preds[t + 1], reward_preds[t] = self.model.latent_dynamics_and_reward(
z_preds[t], action[t]
)
z_preds[t + 1], reward_preds[t] = self.model.latent_dynamics_and_reward(z_preds[t], action[t])
# Compute Q and V value predictions based on the latent rollout.
q_preds_ensemble = self.model.Qs(
z_preds[:-1], action
) # (ensemble, horizon, batch)
q_preds_ensemble = self.model.Qs(z_preds[:-1], action) # (ensemble, horizon, batch)
v_preds = self.model.V(z_preds[:-1])
info.update({"Q": q_preds_ensemble.mean().item(), "V": v_preds.mean().item()})
@@ -424,14 +387,10 @@ class TDMPCPolicy(PreTrainedPolicy):
# actions (not actions estimated by π).
# Note: Here we do not use self.model_target, but self.model. This is to follow the original code
# and the FOWM paper.
q_targets = reward + self.config.discount * self.model.V(
self.model.encode(next_observations)
)
q_targets = reward + self.config.discount * self.model.V(self.model.encode(next_observations))
# From eqn 3 of FOWM. These appear as Q(z, a). Here we call them v_targets to emphasize that we
# are using them to compute loss for V.
v_targets = self.model_target.Qs(
z_preds[:-1].detach(), action, return_min=True
)
v_targets = self.model_target.Qs(z_preds[:-1].detach(), action, return_min=True)
# Compute losses.
# Exponentially decay the loss weight with respect to the timestep. Steps that are more distant in the
@@ -474,9 +433,7 @@ class TDMPCPolicy(PreTrainedPolicy):
temporal_loss_coeffs
* F.mse_loss(
q_preds_ensemble,
einops.repeat(
q_targets, "t b -> e t b", e=q_preds_ensemble.shape[0]
),
einops.repeat(q_targets, "t b -> e t b", e=q_preds_ensemble.shape[0]),
reduction="none",
).sum(0) # sum over ensemble
# `q_preds_ensemble` depends on the first observation and the actions.
@@ -514,14 +471,12 @@ class TDMPCPolicy(PreTrainedPolicy):
z_preds = z_preds.detach()
# Use stopgrad for the advantage calculation.
with torch.no_grad():
advantage = self.model_target.Qs(
z_preds[:-1], action, return_min=True
) - self.model.V(z_preds[:-1])
advantage = self.model_target.Qs(z_preds[:-1], action, return_min=True) - self.model.V(
z_preds[:-1]
)
info["advantage"] = advantage[0]
# (t, b)
exp_advantage = torch.clamp(
torch.exp(advantage * self.config.advantage_scaling), max=100.0
)
exp_advantage = torch.clamp(torch.exp(advantage * self.config.advantage_scaling), max=100.0)
action_preds = self.model.pi(z_preds[:-1]) # (t, b, a)
# Calculate the MSE between the actions and the action predictions.
# Note: FOWM's original code calculates the log probability (wrt to a unit standard deviation
@@ -575,9 +530,7 @@ class TDMPCPolicy(PreTrainedPolicy):
# Note a minor variation with respect to the original FOWM code. Here they do this based on an EMA
# update frequency parameter which is set to 2 (every 2 steps an update is done). To simplify the code
# we update every step and adjust the decay parameter `alpha` accordingly (0.99 -> 0.995)
update_ema_parameters(
self.model_target, self.model, self.config.target_model_momentum
)
update_ema_parameters(self.model_target, self.model, self.config.target_model_momentum)
class TDMPCTOLD(nn.Module):
@@ -588,9 +541,7 @@ class TDMPCTOLD(nn.Module):
self.config = config
self._encoder = TDMPCObservationEncoder(config)
self._dynamics = nn.Sequential(
nn.Linear(
config.latent_dim + config.action_feature.shape[0], config.mlp_dim
),
nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
nn.LayerNorm(config.mlp_dim),
nn.Mish(),
nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -601,9 +552,7 @@ class TDMPCTOLD(nn.Module):
nn.Sigmoid(),
)
self._reward = nn.Sequential(
nn.Linear(
config.latent_dim + config.action_feature.shape[0], config.mlp_dim
),
nn.Linear(config.latent_dim + config.action_feature.shape[0], config.mlp_dim),
nn.LayerNorm(config.mlp_dim),
nn.Mish(),
nn.Linear(config.mlp_dim, config.mlp_dim),
@@ -671,9 +620,7 @@ class TDMPCTOLD(nn.Module):
"Sanity check. The last linear layer needs 0 initialization on weights."
)
nn.init.zeros_(m[-1].weight)
nn.init.zeros_(
m[-1].bias
) # this has already been done, but keep this line here for good measure
nn.init.zeros_(m[-1].bias) # this has already been done, but keep this line here for good measure
def encode(self, obs: dict[str, Tensor]) -> Tensor:
"""Encodes an observation into its latent representation."""
@@ -812,9 +759,7 @@ class TDMPCObservationEncoder(nn.Module):
if config.robot_state_feature:
self.state_enc_layers = nn.Sequential(
nn.Linear(
config.robot_state_feature.shape[0], config.state_encoder_hidden_dim
),
nn.Linear(config.robot_state_feature.shape[0], config.state_encoder_hidden_dim),
nn.ELU(),
nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
nn.LayerNorm(config.latent_dim),
@@ -823,9 +768,7 @@ class TDMPCObservationEncoder(nn.Module):
if config.env_state_feature:
self.env_state_enc_layers = nn.Sequential(
nn.Linear(
config.env_state_feature.shape[0], config.state_encoder_hidden_dim
),
nn.Linear(config.env_state_feature.shape[0], config.state_encoder_hidden_dim),
nn.ELU(),
nn.Linear(config.state_encoder_hidden_dim, config.latent_dim),
nn.LayerNorm(config.latent_dim),
@@ -898,10 +841,7 @@ def update_ema_parameters(ema_net: nn.Module, net: nn.Module, alpha: float):
assert n_p_ema == n_p, "Parameter names don't match for EMA model update"
if isinstance(p, dict):
raise RuntimeError("Dict parameter not supported")
if (
isinstance(module, nn.modules.batchnorm._BatchNorm)
or not p.requires_grad
):
if isinstance(module, nn.modules.batchnorm._BatchNorm) or not p.requires_grad:
# Copy BatchNorm parameters, and non-trainable parameters directly.
p_ema.copy_(p.to(dtype=p_ema.dtype).data)
with torch.no_grad():
@@ -909,9 +849,7 @@ def update_ema_parameters(ema_net: nn.Module, net: nn.Module, alpha: float):
p_ema.add_(p.to(dtype=p_ema.dtype).data, alpha=1 - alpha)
def flatten_forward_unflatten(
fn: Callable[[Tensor], Tensor], image_tensor: Tensor
) -> Tensor:
def flatten_forward_unflatten(fn: Callable[[Tensor], Tensor], image_tensor: Tensor) -> Tensor:
"""Helper to temporarily flatten extra dims at the start of the image tensor.
Args:

5
lerobot/common/policies/vqbet/configuration_vqbet.py
View File

@@ -172,10 +172,7 @@ class VQBeTConfig(PreTrainedConfig):
if self.crop_shape is not None:
for key, image_ft in self.image_features.items():
if (
self.crop_shape[0] > image_ft.shape[1]
or self.crop_shape[1] > image_ft.shape[2]
):
if self.crop_shape[0] > image_ft.shape[1] or self.crop_shape[1] > image_ft.shape[2]:
raise ValueError(
f"`crop_shape` should fit within the images shapes. Got {self.crop_shape} "
f"for `crop_shape` and {image_ft.shape} for "

173
lerobot/common/policies/vqbet/modeling_vqbet.py
View File

@@ -64,9 +64,7 @@ class VQBeTPolicy(PreTrainedPolicy):
config.validate_features()
self.config = config
self.normalize_inputs = Normalize(
config.input_features, config.normalization_mapping, dataset_stats
)
self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
self.normalize_targets = Normalize(
config.output_features, config.normalization_mapping, dataset_stats
)
@@ -97,17 +95,11 @@ class VQBeTPolicy(PreTrainedPolicy):
if self.config.sequentially_select:
decay_params = (
decay_params
+ list(
self.vqbet.action_head.map_to_cbet_preds_primary_bin.parameters()
)
+ list(
self.vqbet.action_head.map_to_cbet_preds_secondary_bin.parameters()
)
+ list(self.vqbet.action_head.map_to_cbet_preds_primary_bin.parameters())
+ list(self.vqbet.action_head.map_to_cbet_preds_secondary_bin.parameters())
)
else:
decay_params = decay_params + list(
self.vqbet.action_head.map_to_cbet_preds_bin.parameters()
)
decay_params = decay_params + list(self.vqbet.action_head.map_to_cbet_preds_bin.parameters())
return [
{
@@ -145,12 +137,8 @@ class VQBeTPolicy(PreTrainedPolicy):
"""
batch = self.normalize_inputs(batch)
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack(
[batch[key] for key in self.config.image_features], dim=-4
)
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
# Note: It's important that this happens after stacking the images into a single key.
self._queues = populate_queues(self._queues, batch)
@@ -161,14 +149,8 @@ class VQBeTPolicy(PreTrainedPolicy):
)
if len(self._queues["action"]) == 0:
batch = {
k: torch.stack(list(self._queues[k]), dim=1)
for k in batch
if k in self._queues
}
actions = self.vqbet(batch, rollout=True)[
:, : self.config.action_chunk_size
]
batch = {k: torch.stack(list(self._queues[k]), dim=1) for k in batch if k in self._queues}
actions = self.vqbet(batch, rollout=True)[:, : self.config.action_chunk_size]
# the dimension of returned action is (batch_size, action_chunk_size, action_dim)
actions = self.unnormalize_outputs({"action": actions})["action"]
@@ -181,12 +163,8 @@ class VQBeTPolicy(PreTrainedPolicy):
def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict]:
"""Run the batch through the model and compute the loss for training or validation."""
batch = self.normalize_inputs(batch)
batch = dict(
batch
) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack(
[batch[key] for key in self.config.image_features], dim=-4
)
batch = dict(batch) # shallow copy so that adding a key doesn't modify the original
batch["observation.images"] = torch.stack([batch[key] for key in self.config.image_features], dim=-4)
batch = self.normalize_targets(batch)
# VQ-BeT discretizes action using VQ-VAE before training BeT (please refer to section 3.2 in the VQ-BeT paper https://arxiv.org/pdf/2403.03181)
if not self.vqbet.action_head.vqvae_model.discretized.item():
@@ -194,9 +172,7 @@ class VQBeTPolicy(PreTrainedPolicy):
# n_different_codes: how many of the total possible VQ codes are being used in single batch (how many of them have at least one encoder embedding as a nearest neighbor). This can be at most `vqvae_n_embed * number of layers of RVQ (=2)`.
# n_different_combinations: how many different code combinations are being used out of all possible combinations in single batch. This can be at most `vqvae_n_embed ^ number of layers of RVQ (=2)` (hint consider the RVQ as a decision tree).
loss, n_different_codes, n_different_combinations, recon_l1_error = (
self.vqbet.action_head.discretize(
self.config.n_vqvae_training_steps, batch["action"]
)
self.vqbet.action_head.discretize(self.config.n_vqvae_training_steps, batch["action"])
)
return loss, {
"n_different_codes": n_different_codes,
@@ -253,9 +229,7 @@ class SpatialSoftmax(nn.Module):
# we could use torch.linspace directly but that seems to behave slightly differently than numpy
# and causes a small degradation in pc_success of pre-trained models.
pos_x, pos_y = np.meshgrid(
np.linspace(-1.0, 1.0, self._in_w), np.linspace(-1.0, 1.0, self._in_h)
)
pos_x, pos_y = np.meshgrid(np.linspace(-1.0, 1.0, self._in_w), np.linspace(-1.0, 1.0, self._in_h))
pos_x = torch.from_numpy(pos_x.reshape(self._in_h * self._in_w, 1)).float()
pos_y = torch.from_numpy(pos_y.reshape(self._in_h * self._in_w, 1)).float()
# register as buffer so it's moved to the correct device.
@@ -370,12 +344,7 @@ class VQBeTModel(nn.Module):
num_tokens = self.config.n_action_pred_token + self.config.n_obs_steps - 1
self.register_buffer(
"select_target_actions_indices",
torch.row_stack(
[
torch.arange(i, i + self.config.action_chunk_size)
for i in range(num_tokens)
]
),
torch.row_stack([torch.arange(i, i + self.config.action_chunk_size) for i in range(num_tokens)]),
)
def forward(self, batch: dict[str, Tensor], rollout: bool) -> tuple[dict, dict]:
@@ -406,19 +375,13 @@ class VQBeTModel(nn.Module):
input_tokens.append(
self.state_projector(batch["observation.state"])
) # (batch, obs_step, projection dims)
input_tokens.append(
einops.repeat(
self.action_token, "1 1 d -> b n d", b=batch_size, n=n_obs_steps
)
)
input_tokens.append(einops.repeat(self.action_token, "1 1 d -> b n d", b=batch_size, n=n_obs_steps))
# Interleave tokens by stacking and rearranging.
input_tokens = torch.stack(input_tokens, dim=2)
input_tokens = einops.rearrange(input_tokens, "b n t d -> b (n t) d")
len_additional_action_token = self.config.n_action_pred_token - 1
future_action_tokens = self.action_token.repeat(
batch_size, len_additional_action_token, 1
)
future_action_tokens = self.action_token.repeat(batch_size, len_additional_action_token, 1)
# add additional action query tokens for predicting future action chunks
input_tokens = torch.cat([input_tokens, future_action_tokens], dim=1)
@@ -427,9 +390,9 @@ class VQBeTModel(nn.Module):
features = self.policy(input_tokens)
# len(self.config.input_features) is the number of different observation modes.
# this line gets the index of action prompt tokens.
historical_act_pred_index = np.arange(0, n_obs_steps) * (
len(self.config.input_features) + 1
) + len(self.config.input_features)
historical_act_pred_index = np.arange(0, n_obs_steps) * (len(self.config.input_features) + 1) + len(
self.config.input_features
)
# only extract the output tokens at the position of action query:
# Behavior Transformer (BeT), and VQ-BeT are both sequence-to-sequence prediction models,
@@ -449,15 +412,13 @@ class VQBeTModel(nn.Module):
action_head_output = self.action_head(features)
# if rollout, VQ-BeT don't calculate loss
if rollout:
return action_head_output["predicted_action"][
:, n_obs_steps - 1, :
].reshape(batch_size, self.config.action_chunk_size, -1)
return action_head_output["predicted_action"][:, n_obs_steps - 1, :].reshape(
batch_size, self.config.action_chunk_size, -1
)
# else, it calculate overall loss (bin prediction loss, and offset loss)
else:
output = batch["action"][:, self.select_target_actions_indices]
loss = self.action_head.loss_fn(
action_head_output, output, reduction="mean"
)
loss = self.action_head.loss_fn(action_head_output, output, reduction="mean")
return action_head_output, loss
@@ -492,9 +453,7 @@ class VQBeTHead(nn.Module):
else:
self.map_to_cbet_preds_bin = MLP(
in_channels=config.gpt_output_dim,
hidden_channels=[
self.vqvae_model.vqvae_num_layers * self.config.vqvae_n_embed
],
hidden_channels=[self.vqvae_model.vqvae_num_layers * self.config.vqvae_n_embed],
)
self.map_to_cbet_preds_offset = MLP(
in_channels=config.gpt_output_dim,
@@ -521,10 +480,7 @@ class VQBeTHead(nn.Module):
loss, metric = self.vqvae_model.vqvae_forward(actions)
n_different_codes = sum(
[
len(torch.unique(metric[2][:, i]))
for i in range(self.vqvae_model.vqvae_num_layers)
]
[len(torch.unique(metric[2][:, i])) for i in range(self.vqvae_model.vqvae_num_layers)]
)
n_different_combinations = len(torch.unique(metric[2], dim=0))
recon_l1_error = metric[0].detach().cpu().item()
@@ -585,18 +541,12 @@ class VQBeTHead(nn.Module):
cbet_secondary_logits / self.config.bet_softmax_temperature, dim=-1
)
sampled_secondary_centers = einops.rearrange(
torch.multinomial(
cbet_secondary_probs.view(-1, choices), num_samples=1
),
torch.multinomial(cbet_secondary_probs.view(-1, choices), num_samples=1),
"(NT) 1 -> NT",
NT=NT,
)
sampled_centers = torch.stack(
(sampled_primary_centers, sampled_secondary_centers), axis=1
)
cbet_logits = torch.stack(
[cbet_primary_logits, cbet_secondary_logits], dim=1
)
sampled_centers = torch.stack((sampled_primary_centers, sampled_secondary_centers), axis=1)
cbet_logits = torch.stack([cbet_primary_logits, cbet_secondary_logits], dim=1)
# if self.config.sequentially_select is False, bin prediction head samples primary and secondary code at once.
else:
cbet_logits = self.map_to_cbet_preds_bin(x)
@@ -605,9 +555,7 @@ class VQBeTHead(nn.Module):
"(NT) (G C) -> (NT) G C",
G=self.vqvae_model.vqvae_num_layers,
)
cbet_probs = torch.softmax(
cbet_logits / self.config.bet_softmax_temperature, dim=-1
)
cbet_probs = torch.softmax(cbet_logits / self.config.bet_softmax_temperature, dim=-1)
NT, G, choices = cbet_probs.shape
sampled_centers = einops.rearrange(
torch.multinomial(cbet_probs.view(-1, choices), num_samples=1),
@@ -627,17 +575,9 @@ class VQBeTHead(nn.Module):
sampled_offsets = sampled_offsets.sum(dim=1)
with torch.no_grad():
# Get the centroids (= vectors corresponding to the codes) of each layer to pass it through RVQ decoder
return_decoder_input = (
self.vqvae_model.get_embeddings_from_code(sampled_centers)
.clone()
.detach()
)
return_decoder_input = self.vqvae_model.get_embeddings_from_code(sampled_centers).clone().detach()
# pass the centroids through decoder to get actions.
decoded_action = (
self.vqvae_model.get_action_from_latent(return_decoder_input)
.clone()
.detach()
)
decoded_action = self.vqvae_model.get_action_from_latent(return_decoder_input).clone().detach()
# reshaped extracted offset to match with decoded centroids
sampled_offsets = einops.rearrange(
sampled_offsets, "NT (W A) -> NT W A", W=self.config.action_chunk_size
@@ -686,9 +626,7 @@ class VQBeTHead(nn.Module):
# Figure out the loss for the actions.
# First, we need to find the closest cluster center for each ground truth action.
with torch.no_grad():
state_vq, action_bins = self.vqvae_model.get_code(
action_seq
) # action_bins: NT, G
state_vq, action_bins = self.vqvae_model.get_code(action_seq) # action_bins: NT, G
# Now we can compute the loss.
@@ -711,12 +649,8 @@ class VQBeTHead(nn.Module):
+ cbet_loss2 * self.config.secondary_code_loss_weight
)
equal_primary_code_rate = torch.sum(
(action_bins[:, 0] == sampled_centers[:, 0]).int()
) / (NT)
equal_secondary_code_rate = torch.sum(
(action_bins[:, 1] == sampled_centers[:, 1]).int()
) / (NT)
equal_primary_code_rate = torch.sum((action_bins[:, 0] == sampled_centers[:, 0]).int()) / (NT)
equal_secondary_code_rate = torch.sum((action_bins[:, 1] == sampled_centers[:, 1]).int()) / (NT)
action_mse_error = torch.mean((action_seq - predicted_action) ** 2)
vq_action_error = torch.mean(torch.abs(action_seq - decoded_action))
@@ -730,9 +664,7 @@ class VQBeTHead(nn.Module):
"classification_loss": cbet_loss.detach().cpu().item(),
"offset_loss": offset_loss.detach().cpu().item(),
"equal_primary_code_rate": equal_primary_code_rate.detach().cpu().item(),
"equal_secondary_code_rate": equal_secondary_code_rate.detach()
.cpu()
.item(),
"equal_secondary_code_rate": equal_secondary_code_rate.detach().cpu().item(),
"vq_action_error": vq_action_error.detach().cpu().item(),
"offset_action_error": offset_action_error.detach().cpu().item(),
"action_error_max": action_error_max.detach().cpu().item(),
@@ -757,9 +689,7 @@ class VQBeTRgbEncoder(nn.Module):
# Always use center crop for eval
self.center_crop = torchvision.transforms.CenterCrop(config.crop_shape)
if config.crop_is_random:
self.maybe_random_crop = torchvision.transforms.RandomCrop(
config.crop_shape
)
self.maybe_random_crop = torchvision.transforms.RandomCrop(config.crop_shape)
else:
self.maybe_random_crop = self.center_crop
else:
@@ -780,9 +710,7 @@ class VQBeTRgbEncoder(nn.Module):
self.backbone = _replace_submodules(
root_module=self.backbone,
predicate=lambda x: isinstance(x, nn.BatchNorm2d),
func=lambda x: nn.GroupNorm(
num_groups=x.num_features // 16, num_channels=x.num_features
),
func=lambda x: nn.GroupNorm(num_groups=x.num_features // 16, num_channels=x.num_features),
)
# Set up pooling and final layers.
@@ -792,15 +720,11 @@ class VQBeTRgbEncoder(nn.Module):
# height and width from `config.image_features`.
images_shape = next(iter(config.image_features.values())).shape
dummy_shape_h_w = (
config.crop_shape if config.crop_shape is not None else images_shape[1:]
)
dummy_shape_h_w = config.crop_shape if config.crop_shape is not None else images_shape[1:]
dummy_shape = (1, images_shape[0], *dummy_shape_h_w)
feature_map_shape = get_output_shape(self.backbone, dummy_shape)[1:]
self.pool = SpatialSoftmax(
feature_map_shape, num_kp=config.spatial_softmax_num_keypoints
)
self.pool = SpatialSoftmax(feature_map_shape, num_kp=config.spatial_softmax_num_keypoints)
self.feature_dim = config.spatial_softmax_num_keypoints * 2
self.out = nn.Linear(config.spatial_softmax_num_keypoints * 2, self.feature_dim)
self.relu = nn.ReLU()
@@ -842,11 +766,7 @@ def _replace_submodules(
if predicate(root_module):
return func(root_module)
replace_list = [
k.split(".")
for k, m in root_module.named_modules(remove_duplicate=True)
if predicate(m)
]
replace_list = [k.split(".") for k, m in root_module.named_modules(remove_duplicate=True) if predicate(m)]
for *parents, k in replace_list:
parent_module = root_module
if len(parents) > 0:
@@ -861,9 +781,7 @@ def _replace_submodules(
else:
setattr(parent_module, k, tgt_module)
# verify that all BN are replaced
assert not any(
predicate(m) for _, m in root_module.named_modules(remove_duplicate=True)
)
assert not any(predicate(m) for _, m in root_module.named_modules(remove_duplicate=True))
return root_module
@@ -896,8 +814,7 @@ class VqVae(nn.Module):
)
self.encoder = MLP(
in_channels=self.config.action_feature.shape[0]
* self.config.action_chunk_size,
in_channels=self.config.action_feature.shape[0] * self.config.action_chunk_size,
hidden_channels=[
config.vqvae_enc_hidden_dim,
config.vqvae_enc_hidden_dim,
@@ -925,13 +842,9 @@ class VqVae(nn.Module):
# given latent vector, this function outputs the decoded action.
output = self.decoder(latent)
if self.config.action_chunk_size == 1:
return einops.rearrange(
output, "N (T A) -> N T A", A=self.config.action_feature.shape[0]
)
return einops.rearrange(output, "N (T A) -> N T A", A=self.config.action_feature.shape[0])
else:
return einops.rearrange(
output, "N (T A) -> N T A", A=self.config.action_feature.shape[0]
)
return einops.rearrange(output, "N (T A) -> N T A", A=self.config.action_feature.shape[0])
def get_code(self, state):
# in phase 2 of VQ-BeT training, we need a `ground truth labels of action data` to calculate the Focal loss for code prediction head. (please refer to section 3.3 in the paper https://arxiv.org/pdf/2403.03181)

181
lerobot/common/policies/vqbet/vqbet_utils.py
View File

@@ -123,15 +123,9 @@ class CausalSelfAttention(nn.Module):
# calculate query, key, values for all heads in batch and move head forward to be the batch dim
q, k, v = self.c_attn(x).split(self.gpt_hidden_dim, dim=2)
k = k.view(B, T, self.gpt_n_head, C // self.gpt_n_head).transpose(
1, 2
) # (B, nh, T, hs)
q = q.view(B, T, self.gpt_n_head, C // self.gpt_n_head).transpose(
1, 2
) # (B, nh, T, hs)
v = v.view(B, T, self.gpt_n_head, C // self.gpt_n_head).transpose(
1, 2
) # (B, nh, T, hs)
k = k.view(B, T, self.gpt_n_head, C // self.gpt_n_head).transpose(1, 2) # (B, nh, T, hs)
q = q.view(B, T, self.gpt_n_head, C // self.gpt_n_head).transpose(1, 2) # (B, nh, T, hs)
v = v.view(B, T, self.gpt_n_head, C // self.gpt_n_head).transpose(1, 2) # (B, nh, T, hs)
# causal self-attention; Self-attend: (B, nh, T, hs) x (B, nh, hs, T) -> (B, nh, T, T)
att = (q @ k.transpose(-2, -1)) * (1.0 / math.sqrt(k.size(-1)))
@@ -139,9 +133,7 @@ class CausalSelfAttention(nn.Module):
att = F.softmax(att, dim=-1)
att = self.attn_dropout(att)
y = att @ v # (B, nh, T, T) x (B, nh, T, hs) -> (B, nh, T, hs)
y = (
y.transpose(1, 2).contiguous().view(B, T, C)
) # re-assemble all head outputs side by side
y = y.transpose(1, 2).contiguous().view(B, T, C) # re-assemble all head outputs side by side
# output projection
y = self.resid_dropout(self.c_proj(y))
@@ -197,16 +189,12 @@ class GPT(nn.Module):
"ln_f": nn.LayerNorm(config.gpt_hidden_dim),
}
)
self.lm_head = nn.Linear(
config.gpt_hidden_dim, config.gpt_output_dim, bias=False
)
self.lm_head = nn.Linear(config.gpt_hidden_dim, config.gpt_output_dim, bias=False)
# init all weights, and apply a special scaled init to the residual projections, per GPT-2 paper
self.apply(self._init_weights)
for pn, p in self.named_parameters():
if pn.endswith("c_proj.weight"):
torch.nn.init.normal_(
p, mean=0.0, std=0.02 / math.sqrt(2 * config.gpt_n_layer)
)
torch.nn.init.normal_(p, mean=0.0, std=0.02 / math.sqrt(2 * config.gpt_n_layer))
# report number of parameters
n_params = sum(p.numel() for p in self.parameters())
@@ -220,17 +208,11 @@ class GPT(nn.Module):
)
# positional encodings that are added to the input embeddings
pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(
0
) # shape (1, t)
pos = torch.arange(0, t, dtype=torch.long, device=device).unsqueeze(0) # shape (1, t)
# forward the GPT model itself
tok_emb = self.transformer.wte(
input
) # token embeddings of shape (b, t, gpt_hidden_dim)
pos_emb = self.transformer.wpe(
pos
) # position embeddings of shape (1, t, gpt_hidden_dim)
tok_emb = self.transformer.wte(input) # token embeddings of shape (b, t, gpt_hidden_dim)
pos_emb = self.transformer.wpe(pos) # position embeddings of shape (1, t, gpt_hidden_dim)
x = self.transformer.drop(tok_emb + pos_emb)
for block in self.transformer.h:
x = block(x)
@@ -255,9 +237,7 @@ class GPT(nn.Module):
# but want to use a smaller block size for some smaller, simpler model
assert gpt_block_size <= self.config.gpt_block_size
self.config.gpt_block_size = gpt_block_size
self.transformer.wpe.weight = nn.Parameter(
self.transformer.wpe.weight[:gpt_block_size]
)
self.transformer.wpe.weight = nn.Parameter(self.transformer.wpe.weight[:gpt_block_size])
for block in self.transformer.h:
block.attn.bias = block.attn.bias[:, :, :gpt_block_size, :gpt_block_size]
@@ -290,10 +270,8 @@ class GPT(nn.Module):
param_dict = dict(self.named_parameters())
inter_params = decay & no_decay
union_params = decay | no_decay
assert len(inter_params) == 0, (
"parameters {} made it into both decay/no_decay sets!".format(
str(inter_params)
)
assert len(inter_params) == 0, "parameters {} made it into both decay/no_decay sets!".format(
str(inter_params)
)
assert len(param_dict.keys() - union_params) == 0, (
"parameters {} were not separated into either decay/no_decay set!".format(
@@ -390,12 +368,8 @@ class ResidualVQ(nn.Module):
codebook_input_dim = codebook_dim * heads
requires_projection = codebook_input_dim != dim
self.project_in = (
nn.Linear(dim, codebook_input_dim) if requires_projection else nn.Identity()
)
self.project_out = (
nn.Linear(codebook_input_dim, dim) if requires_projection else nn.Identity()
)
self.project_in = nn.Linear(dim, codebook_input_dim) if requires_projection else nn.Identity()
self.project_out = nn.Linear(codebook_input_dim, dim) if requires_projection else nn.Identity()
self.num_quantizers = num_quantizers
@@ -477,9 +451,7 @@ class ResidualVQ(nn.Module):
return all_codes
def forward(
self, x, indices=None, return_all_codes=False, sample_codebook_temp=None
):
def forward(self, x, indices=None, return_all_codes=False, sample_codebook_temp=None):
"""
For given input tensor x, this function will return the quantized output, the indices of the quantized output, and the loss.
First, the input tensor x is projected to the codebook dimension. Then, the input tensor x is passed through Nq layers of VectorQuantize.
@@ -508,17 +480,13 @@ class ResidualVQ(nn.Module):
)
ce_losses = []
should_quantize_dropout = (
self.training and self.quantize_dropout and not return_loss
)
should_quantize_dropout = self.training and self.quantize_dropout and not return_loss
# sample a layer index at which to dropout further residual quantization
# also prepare null indices and loss
if should_quantize_dropout:
rand_quantize_dropout_index = randrange(
self.quantize_dropout_cutoff_index, num_quant
)
rand_quantize_dropout_index = randrange(self.quantize_dropout_cutoff_index, num_quant)
if quant_dropout_multiple_of != 1:
rand_quantize_dropout_index = (
@@ -527,23 +495,14 @@ class ResidualVQ(nn.Module):
- 1
)
null_indices_shape = (
(x.shape[0], *x.shape[-2:])
if self.accept_image_fmap
else tuple(x.shape[:2])
)
null_indices = torch.full(
null_indices_shape, -1.0, device=device, dtype=torch.long
)
null_indices_shape = (x.shape[0], *x.shape[-2:]) if self.accept_image_fmap else tuple(x.shape[:2])
null_indices = torch.full(null_indices_shape, -1.0, device=device, dtype=torch.long)
null_loss = torch.full((1,), 0.0, device=device, dtype=x.dtype)
# go through the layers
for quantizer_index, layer in enumerate(self.layers):
if (
should_quantize_dropout
and quantizer_index > rand_quantize_dropout_index
):
if should_quantize_dropout and quantizer_index > rand_quantize_dropout_index:
all_indices.append(null_indices)
all_losses.append(null_loss)
continue
@@ -583,9 +542,7 @@ class ResidualVQ(nn.Module):
# stack all losses and indices
all_losses, all_indices = map(
partial(torch.stack, dim=-1), (all_losses, all_indices)
)
all_losses, all_indices = map(partial(torch.stack, dim=-1), (all_losses, all_indices))
ret = (quantized_out, all_indices, all_losses)
@@ -645,12 +602,8 @@ class VectorQuantize(nn.Module):
codebook_input_dim = codebook_dim * heads
requires_projection = codebook_input_dim != dim
self.project_in = (
nn.Linear(dim, codebook_input_dim) if requires_projection else nn.Identity()
)
self.project_out = (
nn.Linear(codebook_input_dim, dim) if requires_projection else nn.Identity()
)
self.project_in = nn.Linear(dim, codebook_input_dim) if requires_projection else nn.Identity()
self.project_out = nn.Linear(codebook_input_dim, dim) if requires_projection else nn.Identity()
self.eps = eps
self.commitment_weight = commitment_weight
@@ -664,14 +617,10 @@ class VectorQuantize(nn.Module):
self.orthogonal_reg_active_codes_only = orthogonal_reg_active_codes_only
self.orthogonal_reg_max_codes = orthogonal_reg_max_codes
assert not (ema_update and learnable_codebook), (
"learnable codebook not compatible with EMA update"
)
assert not (ema_update and learnable_codebook), "learnable codebook not compatible with EMA update"
assert 0 <= sync_update_v <= 1.0
assert not (sync_update_v > 0.0 and not learnable_codebook), (
"learnable codebook must be turned on"
)
assert not (sync_update_v > 0.0 and not learnable_codebook), "learnable codebook must be turned on"
self.sync_update_v = sync_update_v
@@ -683,9 +632,7 @@ class VectorQuantize(nn.Module):
)
if sync_codebook is None:
sync_codebook = (
distributed.is_initialized() and distributed.get_world_size() > 1
)
sync_codebook = distributed.is_initialized() and distributed.get_world_size() > 1
codebook_kwargs = {
"dim": codebook_dim,
@@ -850,17 +797,11 @@ class VectorQuantize(nn.Module):
# quantize again
quantize, embed_ind, distances = self._codebook(
x, **codebook_forward_kwargs
)
quantize, embed_ind, distances = self._codebook(x, **codebook_forward_kwargs)
if self.training:
# determine code to use for commitment loss
maybe_detach = (
torch.detach
if not self.learnable_codebook or freeze_codebook
else identity
)
maybe_detach = torch.detach if not self.learnable_codebook or freeze_codebook else identity
commit_quantize = maybe_detach(quantize)
@@ -870,9 +811,7 @@ class VectorQuantize(nn.Module):
if self.sync_update_v > 0.0:
# (21) in https://minyoungg.github.io/vqtorch/assets/draft_050523.pdf
quantize = quantize + self.sync_update_v * (
quantize - quantize.detach()
)
quantize = quantize + self.sync_update_v * (quantize - quantize.detach())
# function for calculating cross entropy loss to distance matrix
# used for (1) naturalspeech2 training residual vq latents to be close to the correct codes and (2) cross-entropy based commitment loss
@@ -905,9 +844,7 @@ class VectorQuantize(nn.Module):
embed_ind = rearrange(embed_ind, "1 (b h) n -> b n h", h=heads)
if self.accept_image_fmap:
embed_ind = rearrange(
embed_ind, "b (h w) ... -> b h w ...", h=height, w=width
)
embed_ind = rearrange(embed_ind, "b (h w) ... -> b h w ...", h=height, w=width)
if only_one:
embed_ind = rearrange(embed_ind, "b 1 -> b")
@@ -961,12 +898,8 @@ class VectorQuantize(nn.Module):
num_codes = codebook.shape[-2]
if (
self.orthogonal_reg_max_codes is not None
) and num_codes > self.orthogonal_reg_max_codes:
rand_ids = torch.randperm(num_codes, device=device)[
: self.orthogonal_reg_max_codes
]
if (self.orthogonal_reg_max_codes is not None) and num_codes > self.orthogonal_reg_max_codes:
rand_ids = torch.randperm(num_codes, device=device)[: self.orthogonal_reg_max_codes]
codebook = codebook[:, rand_ids]
orthogonal_reg_loss = orthogonal_loss_fn(codebook)
@@ -998,9 +931,7 @@ class VectorQuantize(nn.Module):
# if masking, only return quantized for where mask has True
if mask is not None:
quantize = torch.where(
rearrange(mask, "... -> ... 1"), quantize, orig_input
)
quantize = torch.where(rearrange(mask, "... -> ... 1"), quantize, orig_input)
return quantize, embed_ind, loss
@@ -1110,9 +1041,7 @@ def sample_vectors(samples, num):
def batched_sample_vectors(samples, num):
return torch.stack(
[sample_vectors(sample, num) for sample in samples.unbind(dim=0)], dim=0
)
return torch.stack([sample_vectors(sample, num) for sample in samples.unbind(dim=0)], dim=0)
def pad_shape(shape, size, dim=0):
@@ -1163,9 +1092,7 @@ def sample_vectors_distributed(local_samples, num):
all_num_samples = all_gather_sizes(local_samples, dim=0)
if rank == 0:
samples_per_rank = sample_multinomial(
num, all_num_samples / all_num_samples.sum()
)
samples_per_rank = sample_multinomial(num, all_num_samples / all_num_samples.sum())
else:
samples_per_rank = torch.empty_like(all_num_samples)
@@ -1278,9 +1205,7 @@ class EuclideanCodebook(nn.Module):
self.eps = eps
self.threshold_ema_dead_code = threshold_ema_dead_code
self.reset_cluster_size = (
reset_cluster_size
if (reset_cluster_size is not None)
else threshold_ema_dead_code
reset_cluster_size if (reset_cluster_size is not None) else threshold_ema_dead_code
)
assert callable(gumbel_sample)
@@ -1291,14 +1216,8 @@ class EuclideanCodebook(nn.Module):
"kmeans init is not compatible with multiple codebooks in distributed environment for now"
)
self.sample_fn = (
sample_vectors_distributed
if use_ddp and sync_kmeans
else batched_sample_vectors
)
self.kmeans_all_reduce_fn = (
distributed.all_reduce if use_ddp and sync_kmeans else noop
)
self.sample_fn = sample_vectors_distributed if use_ddp and sync_kmeans else batched_sample_vectors
self.kmeans_all_reduce_fn = distributed.all_reduce if use_ddp and sync_kmeans else noop
self.all_reduce_fn = distributed.all_reduce if use_ddp else noop
self.register_buffer("initted", torch.Tensor([not kmeans_init]))
@@ -1437,9 +1356,7 @@ class EuclideanCodebook(nn.Module):
distributed.all_reduce(variance_number)
batch_variance = variance_number / num_vectors
self.update_with_decay(
"batch_variance", batch_variance, self.affine_param_batch_decay
)
self.update_with_decay("batch_variance", batch_variance, self.affine_param_batch_decay)
def replace(self, batch_samples, batch_mask):
for ind, (samples, mask) in enumerate(
@@ -1448,9 +1365,7 @@ class EuclideanCodebook(nn.Module):
if not torch.any(mask):
continue
sampled = self.sample_fn(
rearrange(samples, "... -> 1 ..."), mask.sum().item()
)
sampled = self.sample_fn(rearrange(samples, "... -> 1 ..."), mask.sum().item())
sampled = rearrange(sampled, "1 ... -> ...")
self.embed.data[ind][mask] = sampled
@@ -1474,9 +1389,7 @@ class EuclideanCodebook(nn.Module):
def forward(self, x, sample_codebook_temp=None, mask=None, freeze_codebook=False):
needs_codebook_dim = x.ndim < 4
sample_codebook_temp = (
sample_codebook_temp
if (sample_codebook_temp is not None)
else self.sample_codebook_temp
sample_codebook_temp if (sample_codebook_temp is not None) else self.sample_codebook_temp
)
x = x.float()
@@ -1504,9 +1417,7 @@ class EuclideanCodebook(nn.Module):
if self.affine_param:
codebook_std = self.codebook_variance.clamp(min=1e-5).sqrt()
batch_std = self.batch_variance.clamp(min=1e-5).sqrt()
embed = (embed - self.codebook_mean) * (
batch_std / codebook_std
) + self.batch_mean
embed = (embed - self.codebook_mean) * (batch_std / codebook_std) + self.batch_mean
dist = -cdist(flatten, embed)
@@ -1524,9 +1435,7 @@ class EuclideanCodebook(nn.Module):
if self.training and self.ema_update and not freeze_codebook:
if self.affine_param:
flatten = (flatten - self.batch_mean) * (
codebook_std / batch_std
) + self.codebook_mean
flatten = (flatten - self.batch_mean) * (codebook_std / batch_std) + self.codebook_mean
if mask is not None:
embed_onehot[~mask] = 0.0
@@ -1549,9 +1458,7 @@ class EuclideanCodebook(nn.Module):
self.expire_codes_(x)
if needs_codebook_dim:
quantize, embed_ind = tuple(
rearrange(t, "1 ... -> ...") for t in (quantize, embed_ind)
)
quantize, embed_ind = tuple(rearrange(t, "1 ... -> ...") for t in (quantize, embed_ind))
dist = unpack_one(dist, ps, "h * d")