import einops import numpy as np import torch from torch import nn from .backbone import build_backbone from .transformer import Transformer, TransformerEncoder def get_sinusoid_encoding_table(n_position, d_hid): def get_position_angle_vec(position): return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)] sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)]) 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.FloatTensor(sinusoid_table).unsqueeze(0) class ActionChunkingTransformer(nn.Module): """ Action Chunking Transformer 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 symbols `vae_encoder`, 'encoder', `decoder`. The meanings are as follows. - The `vae_encoder` is, as per the literature around conditional variational auto-encoders (cVAE), the part of the model that encodes the target data (here, a sequence of actions), and the condition (here, we include the robot joint-space state as an input to the encoder). - The `transformer` is the cVAE's decoder. But since we have an option to train this model without the variational objective (in which case we drop the `vae_encoder` altogether), we don't call it the `vae_decoder`. # TODO(now): remove the following - The `encoder` is actually a component of the cVAE's "decoder". But we refer to it as an "encoder" because, in terms of the transformer with cross-attention that forms the cVAE's decoder, it is the "encoder" part. 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). - The `decoder` is a building block of the VAE decoder, and is just the "decoder" part of a transformer with cross-attention. For the same reasoning behind the naming of `encoder`, we make this term agnostic to the option to use a variational objective for training. """ def __init__( self, backbones, transformer, vae_encoder, state_dim, action_dim, horizon, camera_names, use_vae ): """Initializes the model. Parameters: backbones: torch module of the backbone to be used. See backbone.py transformer: torch module of the transformer architecture. See transformer.py state_dim: robot state dimension of the environment horizon: number of object queries, ie detection slot. This is the maximal number of objects DETR can detect in a single image. For COCO, we recommend 100 queries. Args: state_dim: Robot positional state dimension. action_dim: Action dimension. horizon: The number of actions to generate in one forward pass. use_vae: Whether to use the variational objective. TODO(now): Give more details. """ super().__init__() self.camera_names = camera_names self.transformer = transformer self.vae_encoder = vae_encoder self.use_vae = use_vae hidden_dim = transformer.d_model # 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 use_vae: self.cls_embed = nn.Embedding(1, hidden_dim) # Projection layer for joint-space configuration to hidden dimension. self.vae_encoder_robot_state_input_proj = nn.Linear(state_dim, hidden_dim) # Projection layer for action (joint-space target) to hidden dimension. self.vae_encoder_action_input_proj = nn.Linear(state_dim, hidden_dim) # Final size of latent z. TODO(now): Add to hyperparams. self.latent_dim = 32 # Projection layer from the VAE encoder's output to the latent distribution's parameter space. self.vae_encoder_latent_output_proj = nn.Linear(hidden_dim, self.latent_dim * 2) # Fixed sinusoidal positional embedding the whole input to the VAE encoder. self.register_buffer( "vae_encoder_pos_enc", get_sinusoid_encoding_table(1 + 1 + horizon, hidden_dim) ) # Transformer encoder input projections. The tokens will be structured like # [latent, robot_state, image_feature_map_pixels]. self.backbones = nn.ModuleList(backbones) self.encoder_img_feat_input_proj = nn.Conv2d(backbones[0].num_channels, hidden_dim, kernel_size=1) self.encoder_robot_state_input_proj = nn.Linear(state_dim, hidden_dim) self.encoder_latent_input_proj = nn.Linear(self.latent_dim, hidden_dim) # TODO(now): Fix this nonsense. One positional embedding is needed. We should extract the image # feature dimension with a dry run. self.additional_pos_embed = nn.Embedding( 2, hidden_dim ) # learned position embedding for proprio and latent # Transformer decoder. # Learnable positional embedding for the transformer's decoder (in the style of DETR object queries). self.decoder_pos_embed = nn.Embedding(horizon, hidden_dim) # Final action regression head on the output of the transformer's decoder. self.action_head = nn.Linear(hidden_dim, action_dim) def forward(self, robot_state, image, actions=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. """ if self.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 # Prepare the latent for input to the transformer. if self.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.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) # Note: detach() shouldn't be necessary but leaving it the same as the original code just in case. # Prepare fixed positional embedding. pos_embed = self.vae_encoder_pos_enc.clone().detach() # (1, S+2, D) # Forward pass through VAE encoder and sample the latent with the reparameterization trick. cls_token_out = self.vae_encoder( vae_encoder_input.permute(1, 0, 2), pos=pos_embed.permute(1, 0, 2) )[0] # (B, D) latent_pdf_params = self.vae_encoder_latent_output_proj(cls_token_out) mu = latent_pdf_params[:, : self.latent_dim] logvar = latent_pdf_params[:, self.latent_dim :] # Use reparameterization trick to sample from the latent's PDF. latent_sample = mu + logvar.div(2).exp() * torch.randn_like(mu) else: # When not using the VAE encoder, we set the latent to be all zeros. mu = logvar = None latent_sample = torch.zeros([batch_size, self.latent_dim], dtype=robot_state.dtype).to( robot_state.device ) # Prepare all other transformer inputs. # Image observation features and position embeddings. all_cam_features = [] all_cam_pos = [] for cam_id, _ in enumerate(self.camera_names): # TODO(now): remove the positional embedding from the backbones. cam_features, pos = self.backbones[0](image[:, cam_id]) # HARDCODED cam_features = cam_features[0] # take the last layer feature pos = pos[0] cam_features = self.encoder_img_feat_input_proj(cam_features) # (B, C, h, w) all_cam_features.append(cam_features) all_cam_pos.append(pos) # Concatenate image observation feature maps along the width dimension. transformer_input = torch.cat(all_cam_features, axis=3) # TODO(now): remove the positional embedding from the backbones. pos = torch.cat(all_cam_pos, axis=3) robot_state_embed = self.encoder_robot_state_input_proj(robot_state) latent_embed = self.encoder_latent_input_proj(latent_sample) # TODO(now): Explain all of this madness. transformer_input = torch.cat( [ torch.stack([latent_embed, robot_state_embed], axis=0), transformer_input.flatten(2).permute(2, 0, 1), ] ) pos_embed = torch.cat( [self.additional_pos_embed.weight.unsqueeze(1), pos.flatten(2).permute(2, 0, 1)], axis=0 ) # Run the transformer and project the outputs to the action space. transformer_output = self.transformer( transformer_input, encoder_pos=pos_embed, decoder_pos=self.decoder_pos_embed.weight.unsqueeze(1), ).transpose(0, 1) # back to (B, S, C) actions = self.action_head(transformer_output) return actions, [mu, logvar] def build(args): # From state # backbone = None # from state for now, no need for conv nets # From image backbones = [] backbone = build_backbone(args) backbones.append(backbone) transformer = Transformer( d_model=args.hidden_dim, dropout=args.dropout, nhead=args.nheads, dim_feedforward=args.dim_feedforward, num_encoder_layers=args.enc_layers, num_decoder_layers=args.dec_layers, normalize_before=args.pre_norm, ) # TODO(now): args.enc_layers shouldn't be shared with the transformer decoder vae_encoder = TransformerEncoder( num_layers=args.enc_layers, d_model=args.hidden_dim, nhead=args.nheads, dim_feedforward=args.dim_feedforward, dropout=args.dropout, activation="relu", normalize_before=args.pre_norm, ) model = ActionChunkingTransformer( backbones, transformer, vae_encoder, state_dim=args.state_dim, action_dim=args.action_dim, horizon=args.num_queries, camera_names=args.camera_names, use_vae=args.vae, ) n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print("number of parameters: {:.2f}M".format(n_parameters / 1e6)) return model