642 lines
28 KiB
Python
642 lines
28 KiB
Python
"""Action Chunking Transformer Policy
|
|
|
|
As per Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (https://arxiv.org/abs/2304.13705).
|
|
The majority of changes here involve removing unused code, unifying naming, and adding helpful comments.
|
|
"""
|
|
|
|
import math
|
|
import time
|
|
from collections import deque
|
|
from itertools import chain
|
|
from typing import Callable
|
|
|
|
import einops
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn.functional as F # noqa: N812
|
|
import torchvision
|
|
import torchvision.transforms as transforms
|
|
from torch import Tensor, nn
|
|
from torchvision.models._utils import IntermediateLayerGetter
|
|
from torchvision.ops.misc import FrozenBatchNorm2d
|
|
|
|
from lerobot.common.policies.act.configuration_act import ActionChunkingTransformerConfig
|
|
|
|
|
|
class ActionChunkingTransformerPolicy(nn.Module):
|
|
"""
|
|
Action Chunking Transformer Policy as per Learning Fine-Grained Bimanual Manipulation with Low-Cost
|
|
Hardware (paper: https://arxiv.org/abs/2304.13705, code: https://github.com/tonyzhaozh/act)
|
|
|
|
Note: In this code we use the terms `vae_encoder`, 'encoder', `decoder`. The meanings are as follows.
|
|
- The `vae_encoder` is, as per the literature around variational auto-encoders (VAE), the part of the
|
|
model that encodes the target data (a sequence of actions), and the condition (the robot
|
|
joint-space).
|
|
- A transformer with an `encoder` (not the VAE encoder) and `decoder` (not the VAE decoder) with
|
|
cross-attention is used as the VAE decoder. For these terms, we drop the `vae_` prefix because we
|
|
have an option to train this model without the variational objective (in which case we drop the
|
|
`vae_encoder` altogether, and nothing about this model has anything to do with a VAE).
|
|
|
|
Transformer
|
|
Used alone for inference
|
|
(acts as VAE decoder
|
|
during training)
|
|
┌───────────────────────┐
|
|
│ Outputs │
|
|
│ ▲ │
|
|
│ ┌─────►┌───────┐ │
|
|
┌──────┐ │ │ │Transf.│ │
|
|
│ │ │ ├─────►│decoder│ │
|
|
┌────┴────┐ │ │ │ │ │ │
|
|
│ │ │ │ ┌───┴───┬─►│ │ │
|
|
│ VAE │ │ │ │ │ └───────┘ │
|
|
│ encoder │ │ │ │Transf.│ │
|
|
│ │ │ │ │encoder│ │
|
|
└───▲─────┘ │ │ │ │ │
|
|
│ │ │ └───▲───┘ │
|
|
│ │ │ │ │
|
|
inputs └─────┼─────┘ │
|
|
│ │
|
|
└───────────────────────┘
|
|
"""
|
|
|
|
name = "act"
|
|
_multiple_obs_steps_not_handled_msg = (
|
|
"ActionChunkingTransformerPolicy does not handle multiple observation steps."
|
|
)
|
|
|
|
def __init__(self, cfg: ActionChunkingTransformerConfig | None = None):
|
|
"""
|
|
Args:
|
|
cfg: Policy configuration class instance or None, in which case the default instantiation of the
|
|
configuration class is used.
|
|
"""
|
|
super().__init__()
|
|
if cfg is None:
|
|
cfg = ActionChunkingTransformerConfig()
|
|
if cfg.n_obs_steps != 1:
|
|
raise ValueError(self._multiple_obs_steps_not_handled_msg)
|
|
self.cfg = cfg
|
|
|
|
# BERT style VAE encoder with input [cls, *joint_space_configuration, *action_sequence].
|
|
# The cls token forms parameters of the latent's distribution (like this [*means, *log_variances]).
|
|
if self.cfg.use_vae:
|
|
self.vae_encoder = _TransformerEncoder(cfg)
|
|
self.vae_encoder_cls_embed = nn.Embedding(1, cfg.d_model)
|
|
# Projection layer for joint-space configuration to hidden dimension.
|
|
self.vae_encoder_robot_state_input_proj = nn.Linear(cfg.state_dim, cfg.d_model)
|
|
# Projection layer for action (joint-space target) to hidden dimension.
|
|
self.vae_encoder_action_input_proj = nn.Linear(cfg.state_dim, cfg.d_model)
|
|
self.latent_dim = cfg.latent_dim
|
|
# Projection layer from the VAE encoder's output to the latent distribution's parameter space.
|
|
self.vae_encoder_latent_output_proj = nn.Linear(cfg.d_model, self.latent_dim * 2)
|
|
# Fixed sinusoidal positional embedding the whole input to the VAE encoder. Unsqueeze for batch
|
|
# dimension.
|
|
self.register_buffer(
|
|
"vae_encoder_pos_enc",
|
|
_create_sinusoidal_position_embedding(1 + 1 + cfg.chunk_size, cfg.d_model).unsqueeze(0),
|
|
)
|
|
|
|
# Backbone for image feature extraction.
|
|
self.image_normalizer = transforms.Normalize(
|
|
mean=cfg.image_normalization_mean, std=cfg.image_normalization_std
|
|
)
|
|
backbone_model = getattr(torchvision.models, cfg.vision_backbone)(
|
|
replace_stride_with_dilation=[False, False, cfg.replace_final_stride_with_dilation],
|
|
pretrained=cfg.use_pretrained_backbone,
|
|
norm_layer=FrozenBatchNorm2d,
|
|
)
|
|
# 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"})
|
|
|
|
# Transformer (acts as VAE decoder when training with the variational objective).
|
|
self.encoder = _TransformerEncoder(cfg)
|
|
self.decoder = _TransformerDecoder(cfg)
|
|
|
|
# Transformer encoder input projections. The tokens will be structured like
|
|
# [latent, robot_state, image_feature_map_pixels].
|
|
self.encoder_robot_state_input_proj = nn.Linear(cfg.state_dim, cfg.d_model)
|
|
self.encoder_latent_input_proj = nn.Linear(self.latent_dim, cfg.d_model)
|
|
self.encoder_img_feat_input_proj = nn.Conv2d(
|
|
backbone_model.fc.in_features, cfg.d_model, kernel_size=1
|
|
)
|
|
# Transformer encoder positional embeddings.
|
|
self.encoder_robot_and_latent_pos_embed = nn.Embedding(2, cfg.d_model)
|
|
self.encoder_cam_feat_pos_embed = _SinusoidalPositionEmbedding2D(cfg.d_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(cfg.chunk_size, cfg.d_model)
|
|
|
|
# Final action regression head on the output of the transformer's decoder.
|
|
self.action_head = nn.Linear(cfg.d_model, cfg.action_dim)
|
|
|
|
self._reset_parameters()
|
|
self._create_optimizer()
|
|
|
|
def _create_optimizer(self):
|
|
optimizer_params_dicts = [
|
|
{
|
|
"params": [
|
|
p for n, p in self.named_parameters() if not n.startswith("backbone") and p.requires_grad
|
|
]
|
|
},
|
|
{
|
|
"params": [
|
|
p for n, p in self.named_parameters() if n.startswith("backbone") and p.requires_grad
|
|
],
|
|
"lr": self.cfg.lr_backbone,
|
|
},
|
|
]
|
|
self.optimizer = torch.optim.AdamW(
|
|
optimizer_params_dicts, lr=self.cfg.lr, weight_decay=self.cfg.weight_decay
|
|
)
|
|
|
|
def _reset_parameters(self):
|
|
"""Xavier-uniform initialization of the transformer parameters as in the original code."""
|
|
for p in chain(self.encoder.parameters(), self.decoder.parameters()):
|
|
if p.dim() > 1:
|
|
nn.init.xavier_uniform_(p)
|
|
|
|
def reset(self):
|
|
"""This should be called whenever the environment is reset."""
|
|
if self.cfg.n_action_steps is not None:
|
|
self._action_queue = deque([], maxlen=self.cfg.n_action_steps)
|
|
|
|
def forward(self, batch: dict[str, Tensor], **_) -> Tensor:
|
|
return self.select_action(self, batch)
|
|
|
|
@torch.no_grad
|
|
def select_action(self, batch: dict[str, Tensor], **_) -> 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
|
|
environment. It works by managing the actions in a queue and only calling `select_actions` when the
|
|
queue is empty.
|
|
"""
|
|
if len(self._action_queue) == 0:
|
|
# `select_actions` returns a (batch_size, n_action_steps, *) tensor, but the queue effectively has
|
|
# shape (n_action_steps, batch_size, *), hence the transpose.
|
|
self._action_queue.extend(self._select_actions(batch).transpose(0, 1))
|
|
return self._action_queue.popleft()
|
|
|
|
@torch.no_grad
|
|
def _select_actions(self, batch: dict[str, Tensor]) -> Tensor:
|
|
"""Use the action chunking transformer to generate a sequence of actions."""
|
|
self.eval()
|
|
batch = self._reshape_batch(batch, add_obs_steps_dim=True)
|
|
actions, _ = self._forward(
|
|
batch["observation.state"], self.image_normalizer(batch["observation.images.top"])
|
|
)
|
|
return actions[: self.cfg.n_action_steps]
|
|
|
|
def _reshape_batch(self, batch: dict[str, Tensor], add_obs_steps_dim: bool = False) -> dict[str, Tensor]:
|
|
"""Reshapes the batch items to account for various requirements of this policy.
|
|
|
|
This function expects `batch` to have (at least):
|
|
{
|
|
"observation.state": (B, 1, J) OR (B, J) tensor of robot states (joint configuration).
|
|
"observation.images.top": (B, 1, C, H, W) OR (B, C, H, W) tensor of images.
|
|
}
|
|
|
|
TODO(alexander-soare): Right now this method does and undoes reshaping operations. This is just to
|
|
separate out the core logic from the temporary logic. See comments below.
|
|
"""
|
|
# Create a shallow copy.
|
|
batch = dict(batch)
|
|
|
|
# Add a dimension for observation steps.
|
|
if add_obs_steps_dim:
|
|
# Add a dimension for the observations steps. Since n_obs_steps > 1 is not supported right now,
|
|
# this just amounts to an unsqueeze.
|
|
for k in batch:
|
|
if k.startswith("observation."):
|
|
batch[k] = batch[k].unsqueeze(1)
|
|
|
|
# Temporary logic to remove the observation step dimension as the policy does not yet handle it.
|
|
# TODO(alexander-soare): generalize this to multiple observations steps.
|
|
# Check that there is only 1 observation step (policy does not yet handle more).
|
|
if not all(batch[k].shape[1] == 1 for k in batch if k.startswith("observation.")):
|
|
raise ValueError(self._multiple_obs_steps_not_handled_msg)
|
|
# Remove observation steps dimension.
|
|
for k in batch:
|
|
if k.startswith("observation."):
|
|
batch[k] = batch[k].squeeze(1)
|
|
|
|
# Temporary logic to add the multiple image dimension back in.
|
|
# TODO(alexander-soare): generalize this to multiple images. Once resolved, this logic will stack all
|
|
# images.
|
|
assert (
|
|
sum(k.startswith("observation.images.") and not k.endswith("is_pad") for k in batch) == 1
|
|
), f"{self.__class__.__name__} only handles one image for now."
|
|
# Since we only handle one image, just unsqueeze instead of stacking.
|
|
batch["observation.images.top"] = batch["observation.images.top"].unsqueeze(1)
|
|
|
|
return batch
|
|
|
|
def compute_loss(self, batch, **_) -> float:
|
|
batch = self._reshape_batch(batch)
|
|
|
|
self.train()
|
|
|
|
num_slices = self.cfg.batch_size
|
|
batch_size = self.cfg.chunk_size * num_slices
|
|
|
|
assert batch_size % self.cfg.chunk_size == 0
|
|
assert batch_size % num_slices == 0
|
|
|
|
actions_hat, (mu_hat, log_sigma_x2_hat) = self._forward(
|
|
batch["observation.state"],
|
|
self.image_normalizer(batch["observation.images.top"]),
|
|
batch["action"],
|
|
)
|
|
|
|
l1_loss = (
|
|
F.l1_loss(batch["action"], actions_hat, reduction="none") * ~batch["action_is_pad"].unsqueeze(-1)
|
|
).mean()
|
|
|
|
if self.cfg.use_vae:
|
|
# Calculate Dₖₗ(latent_pdf || standard_normal). Note: After computing the KL-divergence for
|
|
# each dimension independently, we sum over the latent dimension to get the total
|
|
# 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()
|
|
)
|
|
loss = l1_loss + mean_kld * self.cfg.kl_weight
|
|
else:
|
|
loss = l1_loss
|
|
|
|
return loss
|
|
|
|
def update(self, batch, **_) -> dict:
|
|
"""Run the model in train mode, compute the loss, and do an optimization step."""
|
|
start_time = time.time()
|
|
loss = self.compute_loss(batch)
|
|
loss.backward()
|
|
|
|
grad_norm = torch.nn.utils.clip_grad_norm_(
|
|
self.parameters(),
|
|
self.cfg.grad_clip_norm,
|
|
error_if_nonfinite=False,
|
|
)
|
|
|
|
self.optimizer.step()
|
|
self.optimizer.zero_grad()
|
|
|
|
info = {
|
|
"loss": loss.item(),
|
|
"grad_norm": float(grad_norm),
|
|
"lr": self.cfg.lr,
|
|
"update_s": time.time() - start_time,
|
|
}
|
|
|
|
return info
|
|
|
|
def _forward(
|
|
self, robot_state: Tensor, image: Tensor, actions: Tensor | None = None
|
|
) -> tuple[Tensor, tuple[Tensor | None, Tensor | None]]:
|
|
"""
|
|
Args:
|
|
robot_state: (B, J) batch of robot joint configurations.
|
|
image: (B, N, C, H, W) batch of N camera frames.
|
|
actions: (B, S, A) batch of actions from the target dataset which must be provided if the
|
|
VAE is enabled and the model is in training mode.
|
|
Returns:
|
|
(B, S, A) batch of action sequences
|
|
Tuple containing the latent PDF's parameters (mean, log(σ²)) both as (B, L) tensors where L is the
|
|
latent dimension.
|
|
"""
|
|
if self.cfg.use_vae and self.training:
|
|
assert (
|
|
actions is not None
|
|
), "actions must be provided when using the variational objective in training mode."
|
|
|
|
batch_size = robot_state.shape[0]
|
|
|
|
# Prepare the latent for input to the transformer encoder.
|
|
if self.cfg.use_vae and actions is not None:
|
|
# Prepare the input to the VAE encoder: [cls, *joint_space_configuration, *action_sequence].
|
|
cls_embed = einops.repeat(
|
|
self.vae_encoder_cls_embed.weight, "1 d -> b 1 d", b=batch_size
|
|
) # (B, 1, D)
|
|
robot_state_embed = self.vae_encoder_robot_state_input_proj(robot_state).unsqueeze(1) # (B, 1, D)
|
|
action_embed = self.vae_encoder_action_input_proj(actions) # (B, S, D)
|
|
vae_encoder_input = torch.cat([cls_embed, robot_state_embed, action_embed], axis=1) # (B, S+2, D)
|
|
|
|
# Prepare fixed positional embedding.
|
|
# Note: detach() shouldn't be necessary but leaving it the same as the original code just in case.
|
|
pos_embed = self.vae_encoder_pos_enc.clone().detach() # (1, S+2, D)
|
|
|
|
# Forward pass through VAE encoder to get the latent PDF parameters.
|
|
cls_token_out = self.vae_encoder(
|
|
vae_encoder_input.permute(1, 0, 2), pos_embed=pos_embed.permute(1, 0, 2)
|
|
)[0] # select the class token, with shape (B, D)
|
|
latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out)
|
|
mu = latent_pdf_params[:, : self.latent_dim]
|
|
# This is 2log(sigma). Done this way to match the original implementation.
|
|
log_sigma_x2 = latent_pdf_params[:, self.latent_dim :]
|
|
|
|
# Sample the latent with the reparameterization trick.
|
|
latent_sample = mu + log_sigma_x2.div(2).exp() * torch.randn_like(mu)
|
|
else:
|
|
# When not using the VAE encoder, we set the latent to be all zeros.
|
|
mu = log_sigma_x2 = None
|
|
latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=torch.float32).to(
|
|
robot_state.device
|
|
)
|
|
|
|
# Prepare all other transformer encoder inputs.
|
|
# Camera observation features and positional embeddings.
|
|
all_cam_features = []
|
|
all_cam_pos_embeds = []
|
|
for cam_id, _ in enumerate(self.cfg.camera_names):
|
|
cam_features = self.backbone(image[:, cam_id])["feature_map"]
|
|
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) # (B, C, h, w)
|
|
all_cam_features.append(cam_features)
|
|
all_cam_pos_embeds.append(cam_pos_embed)
|
|
# Concatenate camera observation feature maps and positional embeddings along the width dimension.
|
|
encoder_in = torch.cat(all_cam_features, axis=3)
|
|
cam_pos_embed = torch.cat(all_cam_pos_embeds, axis=3)
|
|
|
|
# Get positional embeddings for robot state and latent.
|
|
robot_state_embed = self.encoder_robot_state_input_proj(robot_state)
|
|
latent_embed = self.encoder_latent_input_proj(latent_sample)
|
|
|
|
# Stack encoder input and positional embeddings moving to (S, B, C).
|
|
encoder_in = torch.cat(
|
|
[
|
|
torch.stack([latent_embed, robot_state_embed], axis=0),
|
|
encoder_in.flatten(2).permute(2, 0, 1),
|
|
]
|
|
)
|
|
pos_embed = torch.cat(
|
|
[
|
|
self.encoder_robot_and_latent_pos_embed.weight.unsqueeze(1),
|
|
cam_pos_embed.flatten(2).permute(2, 0, 1),
|
|
],
|
|
axis=0,
|
|
)
|
|
|
|
# Forward pass through the transformer modules.
|
|
encoder_out = self.encoder(encoder_in, pos_embed=pos_embed)
|
|
decoder_in = torch.zeros(
|
|
(self.cfg.chunk_size, batch_size, self.cfg.d_model),
|
|
dtype=pos_embed.dtype,
|
|
device=pos_embed.device,
|
|
)
|
|
decoder_out = self.decoder(
|
|
decoder_in,
|
|
encoder_out,
|
|
encoder_pos_embed=pos_embed,
|
|
decoder_pos_embed=self.decoder_pos_embed.weight.unsqueeze(1),
|
|
)
|
|
|
|
# Move back to (B, S, C).
|
|
decoder_out = decoder_out.transpose(0, 1)
|
|
|
|
actions = self.action_head(decoder_out)
|
|
|
|
return actions, (mu, log_sigma_x2)
|
|
|
|
def save(self, fp):
|
|
torch.save(self.state_dict(), fp)
|
|
|
|
def load(self, fp):
|
|
d = torch.load(fp)
|
|
self.load_state_dict(d)
|
|
|
|
|
|
class _TransformerEncoder(nn.Module):
|
|
"""Convenience module for running multiple encoder layers, maybe followed by normalization."""
|
|
|
|
def __init__(self, cfg: ActionChunkingTransformerConfig):
|
|
super().__init__()
|
|
self.layers = nn.ModuleList([_TransformerEncoderLayer(cfg) for _ in range(cfg.n_encoder_layers)])
|
|
self.norm = nn.LayerNorm(cfg.d_model) if cfg.pre_norm else nn.Identity()
|
|
|
|
def forward(self, x: Tensor, pos_embed: Tensor | None = None) -> Tensor:
|
|
for layer in self.layers:
|
|
x = layer(x, pos_embed=pos_embed)
|
|
x = self.norm(x)
|
|
return x
|
|
|
|
|
|
class _TransformerEncoderLayer(nn.Module):
|
|
def __init__(self, cfg: ActionChunkingTransformerConfig):
|
|
super().__init__()
|
|
self.self_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
|
|
|
|
# Feed forward layers.
|
|
self.linear1 = nn.Linear(cfg.d_model, cfg.dim_feedforward)
|
|
self.dropout = nn.Dropout(cfg.dropout)
|
|
self.linear2 = nn.Linear(cfg.dim_feedforward, cfg.d_model)
|
|
|
|
self.norm1 = nn.LayerNorm(cfg.d_model)
|
|
self.norm2 = nn.LayerNorm(cfg.d_model)
|
|
self.dropout1 = nn.Dropout(cfg.dropout)
|
|
self.dropout2 = nn.Dropout(cfg.dropout)
|
|
|
|
self.activation = _get_activation_fn(cfg.feedforward_activation)
|
|
self.pre_norm = cfg.pre_norm
|
|
|
|
def forward(self, x, pos_embed: Tensor | None = None) -> Tensor:
|
|
skip = x
|
|
if self.pre_norm:
|
|
x = self.norm1(x)
|
|
q = k = x if pos_embed is None else x + pos_embed
|
|
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
|
|
x = self.norm2(x)
|
|
else:
|
|
x = self.norm1(x)
|
|
skip = x
|
|
x = self.linear2(self.dropout(self.activation(self.linear1(x))))
|
|
x = skip + self.dropout2(x)
|
|
if not self.pre_norm:
|
|
x = self.norm2(x)
|
|
return x
|
|
|
|
|
|
class _TransformerDecoder(nn.Module):
|
|
def __init__(self, cfg: ActionChunkingTransformerConfig):
|
|
"""Convenience module for running multiple decoder layers followed by normalization."""
|
|
super().__init__()
|
|
self.layers = nn.ModuleList([_TransformerDecoderLayer(cfg) for _ in range(cfg.n_decoder_layers)])
|
|
self.norm = nn.LayerNorm(cfg.d_model)
|
|
|
|
def forward(
|
|
self,
|
|
x: Tensor,
|
|
encoder_out: Tensor,
|
|
decoder_pos_embed: Tensor | None = None,
|
|
encoder_pos_embed: Tensor | None = None,
|
|
) -> Tensor:
|
|
for layer in self.layers:
|
|
x = layer(
|
|
x, encoder_out, decoder_pos_embed=decoder_pos_embed, encoder_pos_embed=encoder_pos_embed
|
|
)
|
|
if self.norm is not None:
|
|
x = self.norm(x)
|
|
return x
|
|
|
|
|
|
class _TransformerDecoderLayer(nn.Module):
|
|
def __init__(self, cfg: ActionChunkingTransformerConfig):
|
|
super().__init__()
|
|
self.self_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
|
|
self.multihead_attn = nn.MultiheadAttention(cfg.d_model, cfg.n_heads, dropout=cfg.dropout)
|
|
|
|
# Feed forward layers.
|
|
self.linear1 = nn.Linear(cfg.d_model, cfg.dim_feedforward)
|
|
self.dropout = nn.Dropout(cfg.dropout)
|
|
self.linear2 = nn.Linear(cfg.dim_feedforward, cfg.d_model)
|
|
|
|
self.norm1 = nn.LayerNorm(cfg.d_model)
|
|
self.norm2 = nn.LayerNorm(cfg.d_model)
|
|
self.norm3 = nn.LayerNorm(cfg.d_model)
|
|
self.dropout1 = nn.Dropout(cfg.dropout)
|
|
self.dropout2 = nn.Dropout(cfg.dropout)
|
|
self.dropout3 = nn.Dropout(cfg.dropout)
|
|
|
|
self.activation = _get_activation_fn(cfg.feedforward_activation)
|
|
self.pre_norm = cfg.pre_norm
|
|
|
|
def maybe_add_pos_embed(self, tensor: Tensor, pos_embed: Tensor | None) -> Tensor:
|
|
return tensor if pos_embed is None else tensor + pos_embed
|
|
|
|
def forward(
|
|
self,
|
|
x: Tensor,
|
|
encoder_out: Tensor,
|
|
decoder_pos_embed: Tensor | None = None,
|
|
encoder_pos_embed: Tensor | None = None,
|
|
) -> Tensor:
|
|
"""
|
|
Args:
|
|
x: (Decoder Sequence, Batch, Channel) tensor of input tokens.
|
|
encoder_out: (Encoder Sequence, B, C) output features from the last layer of the encoder we are
|
|
cross-attending with.
|
|
decoder_pos_embed: (ES, 1, C) positional embedding for keys (from the encoder).
|
|
encoder_pos_embed: (DS, 1, C) Positional_embedding for the queries (from the decoder).
|
|
Returns:
|
|
(DS, B, C) tensor of decoder output features.
|
|
"""
|
|
skip = x
|
|
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 = skip + self.dropout1(x)
|
|
if self.pre_norm:
|
|
skip = x
|
|
x = self.norm2(x)
|
|
else:
|
|
x = self.norm1(x)
|
|
skip = x
|
|
x = self.multihead_attn(
|
|
query=self.maybe_add_pos_embed(x, decoder_pos_embed),
|
|
key=self.maybe_add_pos_embed(encoder_out, encoder_pos_embed),
|
|
value=encoder_out,
|
|
)[0] # select just the output, not the attention weights
|
|
x = skip + self.dropout2(x)
|
|
if self.pre_norm:
|
|
skip = x
|
|
x = self.norm3(x)
|
|
else:
|
|
x = self.norm2(x)
|
|
skip = x
|
|
x = self.linear2(self.dropout(self.activation(self.linear1(x))))
|
|
x = skip + self.dropout3(x)
|
|
if not self.pre_norm:
|
|
x = self.norm3(x)
|
|
return x
|
|
|
|
|
|
def _create_sinusoidal_position_embedding(num_positions: int, dimension: int) -> Tensor:
|
|
"""1D sinusoidal positional embeddings as in Attention is All You Need.
|
|
|
|
Args:
|
|
num_positions: Number of token positions required.
|
|
Returns: (num_positions, dimension) position embeddings (the first dimension is the batch dimension).
|
|
|
|
"""
|
|
|
|
def get_position_angle_vec(position):
|
|
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[:, 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()
|
|
|
|
|
|
class _SinusoidalPositionEmbedding2D(nn.Module):
|
|
"""2D sinusoidal positional embeddings similar to what's presented in Attention Is All You Need.
|
|
|
|
The variation is that the position indices are normalized in [0, 2π] (not quite: the lower bound is 1/H
|
|
for the vertical direction, and 1/W for the horizontal direction.
|
|
"""
|
|
|
|
def __init__(self, dimension: int):
|
|
"""
|
|
Args:
|
|
dimension: The desired dimension of the embeddings.
|
|
"""
|
|
super().__init__()
|
|
self.dimension = dimension
|
|
self._two_pi = 2 * math.pi
|
|
self._eps = 1e-6
|
|
# Inverse "common ratio" for the geometric progression in sinusoid frequencies.
|
|
self._temperature = 10000
|
|
|
|
def forward(self, x: Tensor) -> Tensor:
|
|
"""
|
|
Args:
|
|
x: A (B, C, H, W) batch of 2D feature map to generate the embeddings for.
|
|
Returns:
|
|
A (1, C, H, W) batch of corresponding sinusoidal positional embeddings.
|
|
"""
|
|
not_mask = torch.ones_like(x[0, :1]) # (1, H, W)
|
|
# Note: These are like range(1, H+1) and range(1, W+1) respectively, but in most implementations
|
|
# they would be range(0, H) and range(0, W). Keeping it at as is to match the original code.
|
|
y_range = not_mask.cumsum(1, dtype=torch.float32)
|
|
x_range = not_mask.cumsum(2, dtype=torch.float32)
|
|
|
|
# "Normalize" the position index such that it ranges in [0, 2π].
|
|
# Note: Adding epsilon on the denominator should not be needed as all values of y_embed and x_range
|
|
# are non-zero by construction. This is an artifact of the original code.
|
|
y_range = y_range / (y_range[:, -1:, :] + self._eps) * self._two_pi
|
|
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
|
|
)
|
|
|
|
x_range = x_range.unsqueeze(-1) / inverse_frequency # (1, H, W, 1)
|
|
y_range = y_range.unsqueeze(-1) / inverse_frequency # (1, H, W, 1)
|
|
|
|
# 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)
|
|
|
|
return pos_embed
|
|
|
|
|
|
def _get_activation_fn(activation: str) -> Callable:
|
|
"""Return an activation function given a string."""
|
|
if activation == "relu":
|
|
return F.relu
|
|
if activation == "gelu":
|
|
return F.gelu
|
|
if activation == "glu":
|
|
return F.glu
|
|
raise RuntimeError(f"activation should be relu/gelu/glu, not {activation}.")
|