Adding Action Chunking with Transformers (ACT) to baselines

examples/baselines/act/README.md

@@ -0,0 +1,96 @@
+# Action Chunking with Transformers (ACT)
+
+Code for running the ACT algorithm based on ["Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware"](https://arxiv.org/pdf/2304.13705). It is adapted from the [original code](https://github.com/tonyzhaozh/act).
+
+## Installation
+
+To get started, we recommend using conda/mamba to create a new environment and install the dependencies
+
+```bash
+conda create -n act-ms python=3.9
+conda activate act-ms
+pip install -e .
+```
+
+## Demonstration Download and Preprocessing
+
+By default for fast downloads and smaller file sizes, ManiSkill demonstrations are stored in a highly reduced/compressed format which includes not keeping any observation data. Run the command to download the demonstration and convert it to a format that includes observation data and the desired action space.
+
+```bash
+python -m mani_skill.utils.download_demo "PickCube-v1"
+```
+
+```bash
+python -m mani_skill.trajectory.replay_trajectory \
+  --traj-path ~/.maniskill/demos/PickCube-v1/motionplanning/trajectory.h5 \
+  --use-first-env-state -c pd_ee_delta_pos -o state \
+  --save-traj --num-procs 10
+```
+
+Set -o to rgbd for RGBD observations. Note that the control mode can heavily influence how well Behavior Cloning performs. In the paper, they reported a degraded performance when using delta joint positions as actions instead of target joint positions. By default, we recommend using `pd_joint_delta_pos` for control mode as all tasks can be solved with that control mode, although it is harder to learn with BC than `pd_ee_delta_pos` or `pd_ee_delta_pose` for robots that have those control modes. Finally, the type of demonstration data used can also impact performance, with typically neural network generated demonstrations being easier to learn from than human/motion planning generated demonstrations.
+
+## Training
+
+We provide scripts to train ACT on demonstrations. Make sure to use the same sim backend as the backend the demonstrations were collected with.
+
+
+Note that some demonstrations are slow (e.g. motion planning or human teleoperated) and can exceed the default max episode steps which can be an issue as imitation learning algorithms learn to solve the task at the same speed the demonstrations solve it. In this case, you can use the `--max-episode-steps` flag to set a higher value so that the policy can solve the task in time. General recommendation is to set `--max-episode-steps` to about 2x the length of the mean demonstrations length you are using for training. We provide recommended numbers for demonstrations in the examples.sh script.
+
+Example training, learning from 100 demonstrations generated via motionplanning in the PickCube-v1 task
+```bash
+python train.py --env-id PickCube-v1 \
+  --demo-path ~/.maniskill/demos/PickCube-v1/motionplanning/trajectory.state.pd_ee_delta_pos.cuda.h5 \
+  --control-mode "pd_ee_delta_pos" --sim-backend "cpu" --num-demos 100 --max_episode_steps 100 \
+  --total_iters 30000 
+```
+
+
+## Train and Evaluate with GPU Simulation
+
+You can also choose to train on trajectories generated in the GPU simulation and evaluate much faster with the GPU simulation. However as most demonstrations are usually generated in the CPU simulation (via motionplanning or teleoperation), you may observe worse performance when evaluating on the GPU simulation vs the CPU simulation. This can be partially alleviated by using the replay trajectory tool to try and replay trajectories back in the GPU simulation.
+
+It is also recommended to not save videos if you are using a lot of parallel environments as the video size can get very large.
+
+To replay trajectories in the GPU simulation, you can use the following command. Note that this can be a bit slow as the replay trajectory tool is currently not optimized for GPU parallelized environments.
+
+```bash
+python -m mani_skill.trajectory.replay_trajectory \
+  --traj-path ~/.maniskill/demos/PickCube-v1/motionplanning/trajectory.h5 \
+  --use-first-env-state -c pd_ee_delta_pos -o state \
+  --save-traj --num-procs 1 -b gpu --count 100 # process only 100 trajectories
+```
+
+Once our GPU backend demonstration dataset is ready, you can use the following command to train and evaluate on the GPU simulation.
+
+```bash
+python train.py --env-id PickCube-v1 \
+  --demo-path ~/.maniskill/demos/PickCube-v1/motionplanning/trajectory.state.pd_ee_delta_pos.cuda.h5 \
+  --control-mode "pd_ee_delta_pos" --sim-backend "gpu" --num-demos 100 --max_episode_steps 100 \
+  --total_iters 30000 \
+  --num-eval-envs 100 --no-capture-video
+```
+
+## Citation
+
+If you use this baseline please cite the following
+```
+@inproceedings{DBLP:conf/rss/ZhaoKLF23,
+  author       = {Tony Z. Zhao and
+                  Vikash Kumar and
+                  Sergey Levine and
+                  Chelsea Finn},
+  editor       = {Kostas E. Bekris and
+                  Kris Hauser and
+                  Sylvia L. Herbert and
+                  Jingjin Yu},
+  title        = {Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware},
+  booktitle    = {Robotics: Science and Systems XIX, Daegu, Republic of Korea, July
+                  10-14, 2023},
+  year         = {2023},
+  url          = {https://doi.org/10.15607/RSS.2023.XIX.016},
+  doi          = {10.15607/RSS.2023.XIX.016},
+  timestamp    = {Thu, 20 Jul 2023 15:37:49 +0200},
+  biburl       = {https://dblp.org/rec/conf/rss/ZhaoKLF23.bib},
+  bibsource    = {dblp computer science bibliography, https://dblp.org}
+}
+```

examples/baselines/act/act/detr/backbone.py

@@ -0,0 +1,121 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+Backbone modules.
+"""
+from collections import OrderedDict
+
+import torch
+import torch.nn.functional as F
+import torchvision
+from torch import nn
+from torchvision.models._utils import IntermediateLayerGetter
+from typing import Dict, List
+
+from examples.baselines.act.act.utils import NestedTensor, is_main_process
+from examples.baselines.act.act.detr.position_encoding import build_position_encoding
+
+import IPython
+e = IPython.embed
+
+class FrozenBatchNorm2d(torch.nn.Module):
+    """
+    BatchNorm2d where the batch statistics and the affine parameters are fixed.
+
+    Copy-paste from torchvision.misc.ops with added eps before rqsrt,
+    without which any other policy_models than torchvision.policy_models.resnet[18,34,50,101]
+    produce nans.
+    """
+
+    def __init__(self, n):
+        super(FrozenBatchNorm2d, self).__init__()
+        self.register_buffer("weight", torch.ones(n))
+        self.register_buffer("bias", torch.zeros(n))
+        self.register_buffer("running_mean", torch.zeros(n))
+        self.register_buffer("running_var", torch.ones(n))
+
+    def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict,
+                              missing_keys, unexpected_keys, error_msgs):
+        num_batches_tracked_key = prefix + 'num_batches_tracked'
+        if num_batches_tracked_key in state_dict:
+            del state_dict[num_batches_tracked_key]
+
+        super(FrozenBatchNorm2d, self)._load_from_state_dict(
+            state_dict, prefix, local_metadata, strict,
+            missing_keys, unexpected_keys, error_msgs)
+
+    def forward(self, x):
+        # move reshapes to the beginning
+        # to make it fuser-friendly
+        w = self.weight.reshape(1, -1, 1, 1)
+        b = self.bias.reshape(1, -1, 1, 1)
+        rv = self.running_var.reshape(1, -1, 1, 1)
+        rm = self.running_mean.reshape(1, -1, 1, 1)
+        eps = 1e-5
+        scale = w * (rv + eps).rsqrt()
+        bias = b - rm * scale
+        return x * scale + bias
+
+
+class BackboneBase(nn.Module):
+
+    def __init__(self, backbone: nn.Module, train_backbone: bool, num_channels: int, return_interm_layers: bool):
+        super().__init__()
+        # for name, parameter in backbone.named_parameters(): # only train later layers # TODO do we want this?
+        #     if not train_backbone or 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
+        #         parameter.requires_grad_(False)
+        if return_interm_layers:
+            return_layers = {"layer1": "0", "layer2": "1", "layer3": "2", "layer4": "3"}
+        else:
+            return_layers = {'layer4': "0"}
+        self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)
+        self.num_channels = num_channels
+
+    def forward(self, tensor):
+        xs = self.body(tensor)
+        return xs
+        # out: Dict[str, NestedTensor] = {}
+        # for name, x in xs.items():
+        #     m = tensor_list.mask
+        #     assert m is not None
+        #     mask = F.interpolate(m[None].float(), size=x.shape[-2:]).to(torch.bool)[0]
+        #     out[name] = NestedTensor(x, mask)
+        # return out
+
+
+class Backbone(BackboneBase):
+    """ResNet backbone with frozen BatchNorm."""
+    def __init__(self, name: str,
+                 train_backbone: bool,
+                 return_interm_layers: bool,
+                 dilation: bool):
+        backbone = getattr(torchvision.models, name)(
+            replace_stride_with_dilation=[False, False, dilation],
+            pretrained=is_main_process(), norm_layer=FrozenBatchNorm2d) # pretrained # TODO do we want frozen batch_norm??
+        num_channels = 512 if name in ('resnet18', 'resnet34') else 2048
+        super().__init__(backbone, train_backbone, num_channels, return_interm_layers)
+
+
+class Joiner(nn.Sequential):
+    def __init__(self, backbone, position_embedding):
+        super().__init__(backbone, position_embedding)
+
+    def forward(self, tensor_list: NestedTensor):
+        xs = self[0](tensor_list)
+        out: List[NestedTensor] = []
+        pos = []
+        for name, x in xs.items():
+            out.append(x)
+            # position encoding
+            pos.append(self[1](x).to(x.dtype))
+
+        return out, pos
+
+
+def build_backbone(args):
+    position_embedding = build_position_encoding(args)
+    train_backbone = args.lr_backbone > 0
+    return_interm_layers = args.masks
+    backbone = Backbone(args.backbone, train_backbone, return_interm_layers, args.dilation)
+    model = Joiner(backbone, position_embedding)
+    model.num_channels = backbone.num_channels
+    return model

examples/baselines/act/act/detr/detr_vae.py

@@ -0,0 +1,139 @@
+# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+"""
+DETR model and criterion classes.
+"""
+import torch
+from torch import nn
+from torch.autograd import Variable
+from examples.baselines.act.act.detr.transformer import build_transformer, TransformerEncoder, TransformerEncoderLayer
+
+import numpy as np
+
+import IPython
+e = IPython.embed
+
+
+def reparametrize(mu, logvar):
+    std = logvar.div(2).exp()
+    eps = Variable(std.data.new(std.size()).normal_())
+    return mu + std * eps
+
+
+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 DETRVAE(nn.Module):
+    """ This is the DETR module that performs object detection """
+    def __init__(self, backbones, transformer, encoder, state_dim, action_dim, num_queries):
+        super().__init__()
+        self.num_queries = num_queries
+        self.transformer = transformer
+        self.encoder = encoder
+        hidden_dim = transformer.d_model
+        self.action_head = nn.Linear(hidden_dim, action_dim)
+        self.query_embed = nn.Embedding(num_queries, hidden_dim)
+        if backbones is not None:
+            self.input_proj = nn.Conv2d(backbones[0].num_channels, hidden_dim, kernel_size=1)
+            self.backbones = nn.ModuleList(backbones)
+            self.input_proj_robot_state = nn.Linear(state_dim, hidden_dim)
+        else:
+            self.input_proj_robot_state = nn.Linear(state_dim, hidden_dim)
+            self.backbones = None
+
+        # encoder extra parameters
+        self.latent_dim = 32 # size of latent z
+        self.cls_embed = nn.Embedding(1, hidden_dim) # extra cls token embedding
+        self.encoder_state_proj = nn.Linear(state_dim, hidden_dim)  # project state to embedding
+        self.encoder_action_proj = nn.Linear(action_dim, hidden_dim) # project action to embedding
+        self.latent_proj = nn.Linear(hidden_dim, self.latent_dim*2) # project hidden state to latent std, var
+        self.register_buffer('pos_table', get_sinusoid_encoding_table(1+1+num_queries, hidden_dim)) # [CLS], state, actions
+
+        # decoder extra parameters
+        self.latent_out_proj = nn.Linear(self.latent_dim, hidden_dim) # project latent sample to embedding
+        self.additional_pos_embed = nn.Embedding(2, hidden_dim) # learned position embedding for state and proprio
+
+    def forward(self, obs, actions=None):
+        is_training = actions is not None
+        state = obs['state'] if self.backbones is not None else obs
+        bs = state.shape[0]
+
+        if is_training:
+            # project CLS token, state sequence, and action sequence to embedding dim
+            cls_embed = self.cls_embed.weight  # (1, hidden_dim)
+            cls_embed = torch.unsqueeze(cls_embed, axis=0).repeat(bs, 1, 1)  # (bs, 1, hidden_dim)
+            state_embed = self.encoder_state_proj(state) # (bs, hidden_dim)
+            state_embed = torch.unsqueeze(state_embed, axis=1)  # (bs, 1, hidden_dim)
+            action_embed = self.encoder_action_proj(actions)  # (bs, seq, hidden_dim)
+            # concat them together to form an input to the CVAE encoder
+            encoder_input = torch.cat([cls_embed, state_embed, action_embed], axis=1) # (bs, seq+2, hidden_dim)
+            encoder_input = encoder_input.permute(1, 0, 2) # (seq+2, bs, hidden_dim)
+            # no masking is applied to all parts of the CVAE encoder input
+            is_pad = torch.full((bs, encoder_input.shape[0]), False).to(state.device) # False: not a padding
+            # obtain position embedding
+            pos_embed = self.pos_table.clone().detach()
+            pos_embed = pos_embed.permute(1, 0, 2)  # (seq+2, 1, hidden_dim)
+            # query CVAE encoder
+            encoder_output = self.encoder(encoder_input, pos=pos_embed, src_key_padding_mask=is_pad)
+            encoder_output = encoder_output[0] # take cls output only
+            latent_info = self.latent_proj(encoder_output)
+            mu = latent_info[:, :self.latent_dim]
+            logvar = latent_info[:, self.latent_dim:]
+            latent_sample = reparametrize(mu, logvar)
+            latent_input = self.latent_out_proj(latent_sample)
+        else:
+            mu = logvar = None
+            latent_sample = torch.zeros([bs, self.latent_dim], dtype=torch.float32).to(state.device)
+            latent_input = self.latent_out_proj(latent_sample)
+
+        # CVAE decoder
+        if self.backbones is not None:
+            vis_data = obs['rgb'] if "rgb" in obs else obs['rgbd']
+            num_cams = vis_data.shape[1]
+
+            # Image observation features and position embeddings
+            all_cam_features = []
+            all_cam_pos = []
+            for cam_id in range(num_cams):
+                features, pos = self.backbones[0](vis_data[:, cam_id]) # HARDCODED
+                features = features[0] # take the last layer feature # (batch, hidden_dim, H, W)
+                pos = pos[0] # (1, hidden_dim, H, W)
+                all_cam_features.append(self.input_proj(features))
+                all_cam_pos.append(pos)
+
+            # proprioception features (state)
+            proprio_input = self.input_proj_robot_state(state)
+            # fold camera dimension into width dimension
+            src = torch.cat(all_cam_features, axis=3) # (batch, hidden_dim, 4, 8)
+            pos = torch.cat(all_cam_pos, axis=3) # (batch, hidden_dim, 4, 8)
+            hs = self.transformer(src, None, self.query_embed.weight, pos, latent_input, proprio_input, self.additional_pos_embed.weight)[0] # (batch, num_queries, hidden_dim)
+        else:
+            state = self.input_proj_robot_state(state)
+            hs = self.transformer(None, None, self.query_embed.weight, None, latent_input, state, self.additional_pos_embed.weight)[0]
+
+        a_hat = self.action_head(hs)
+        return a_hat, [mu, logvar]
+
+
+def build_encoder(args):
+    d_model = args.hidden_dim # 256
+    dropout = args.dropout # 0.1
+    nhead = args.nheads # 8
+    dim_feedforward = args.dim_feedforward # 2048
+    num_encoder_layers = args.enc_layers # 4 # TODO shared with VAE decoder
+    normalize_before = args.pre_norm # False
+    activation = "relu"
+
+    encoder_layer = TransformerEncoderLayer(d_model, nhead, dim_feedforward,
+                                            dropout, activation, normalize_before)
+    encoder_norm = nn.LayerNorm(d_model) if normalize_before else None
+    encoder = TransformerEncoder(encoder_layer, num_encoder_layers, encoder_norm)
+
+    return encoder