Merge branch 'main' of github.com:MarkFzp/act-plus-plus into main

这个提交包含在:
Zipeng Fu
2023-11-14 14:43:05 -08:00
当前提交 b7e387a9b9
共有 4 个文件被更改,包括 258 次插入 和 11 次删除

查看文件

@ -58,7 +58,7 @@ CUDA_VISIBLE_DEVICES=0 python3 imitate_episodes.py \
--ckpt_dir /scr/tonyzhao/train_logs/cube_scripted \ --ckpt_dir /scr/tonyzhao/train_logs/cube_scripted \
--policy_class ACT --kl_weight 10 --chunk_size 50 \ --policy_class ACT --kl_weight 10 --chunk_size 50 \
--hidden_dim 512 --batch_size 12 --dim_feedforward 3200 --lr 1e-5 --seed 0 \ --hidden_dim 512 --batch_size 12 --dim_feedforward 3200 --lr 1e-5 --seed 0 \
--num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000 --num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000 --no_encoder
# launch experiment on all data # launch experiment on all data
@ -70,8 +70,55 @@ CUDA_VISIBLE_DEVICES=0 python3 imitate_episodes.py \
--ckpt_dir /scr/tonyzhao/train_logs/cube_scripted_mirror \ --ckpt_dir /scr/tonyzhao/train_logs/cube_scripted_mirror \
--policy_class ACT --kl_weight 10 --chunk_size 50 \ --policy_class ACT --kl_weight 10 --chunk_size 50 \
--hidden_dim 512 --batch_size 12 --dim_feedforward 3200 --lr 1e-5 --seed 0 \ --hidden_dim 512 --batch_size 12 --dim_feedforward 3200 --lr 1e-5 --seed 0 \
--num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000 --no_encoder
####### DIFFUSION POLICY
- first install https://github.com/ARISE-Initiative/robomimic/tree/r2d2 (note the r2d2 branch)
- on top of it pip install the current repo requirements
conda activate mobile
export MUJOCO_GL=egl
cd /home/tonyzhao/Research/act-plus-plus
CUDA_VISIBLE_DEVICES=0 python3 imitate_episodes.py \
--task_name sim_transfer_cube_scripted \
--ckpt_dir /scr/tonyzhao/train_logs/cube_scripted_diffusion_sweep_0 \
--policy_class Diffusion --chunk_size 32 \
--batch_size 32 --lr 1e-5 --seed 0 \
--num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000
conda activate mobile
export MUJOCO_GL=egl
cd /home/tonyzhao/Research/act-plus-plus
CUDA_VISIBLE_DEVICES=1 python3 imitate_episodes.py \
--task_name sim_transfer_cube_scripted \
--ckpt_dir /scr/tonyzhao/train_logs/cube_scripted_diffusion_sweep_1 \
--policy_class Diffusion --chunk_size 16 \
--batch_size 32 --lr 1e-5 --seed 0 \
--num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000
# above are all 100 train diffusion steps, 1e-5
conda activate mobile
export MUJOCO_GL=egl
cd /home/tonyzhao/Research/act-plus-plus
CUDA_VISIBLE_DEVICES=1 python3 imitate_episodes.py \
--task_name sim_transfer_cube_scripted \
--ckpt_dir /scr/tonyzhao/train_logs/cube_scripted_diffusion_sweep_2_50step_1e-4 \
--policy_class Diffusion --chunk_size 32 \
--batch_size 32 --lr 1e-4 --seed 0 \
--num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000 --num_steps 100000 --eval_every 2000 --validate_every 2000 --save_every 2000
---------------------------------------------------------------------------------------
NOTE: chunk size cannot be any number, try before launching
TODO: Add history, EMA at test time

查看文件

@ -15,7 +15,7 @@ from constants import PUPPET_GRIPPER_JOINT_OPEN
from utils import load_data # data functions from utils import load_data # data functions
from utils import sample_box_pose, sample_insertion_pose # robot functions from utils import sample_box_pose, sample_insertion_pose # robot functions
from utils import compute_dict_mean, set_seed, detach_dict, calibrate_linear_vel, postprocess_base_action # helper functions from utils import compute_dict_mean, set_seed, detach_dict, calibrate_linear_vel, postprocess_base_action # helper functions
from policy import ACTPolicy, CNNMLPPolicy from policy import ACTPolicy, CNNMLPPolicy, DiffusionPolicy
from visualize_episodes import save_videos from visualize_episodes import save_videos
from detr.models.latent_model import Latent_Model_Transformer from detr.models.latent_model import Latent_Model_Transformer
@ -80,6 +80,19 @@ def main(args):
'action_dim': 16, 'action_dim': 16,
'no_encoder': args['no_encoder'], 'no_encoder': args['no_encoder'],
} }
elif policy_class == 'Diffusion':
policy_config = {'lr': args['lr'],
'camera_names': camera_names,
'action_dim': 16,
'observation_horizon': 1,
'action_horizon': 8,
'prediction_horizon': args['chunk_size'],
'num_queries': args['chunk_size'],
'num_inference_timesteps': 10,
'ema_power': 0.75,
'vq': False,
}
elif policy_class == 'CNNMLP': elif policy_class == 'CNNMLP':
policy_config = {'lr': args['lr'], 'lr_backbone': lr_backbone, 'backbone' : backbone, 'num_queries': 1, policy_config = {'lr': args['lr'], 'lr_backbone': lr_backbone, 'backbone' : backbone, 'num_queries': 1,
'camera_names': camera_names,} 'camera_names': camera_names,}
@ -129,7 +142,7 @@ def main(args):
print() print()
exit() exit()
train_dataloader, val_dataloader, stats, _ = load_data(dataset_dir, name_filter, camera_names, batch_size_train, batch_size_val, args['chunk_size'], args['skip_mirrored_data'], config['load_pretrain']) train_dataloader, val_dataloader, stats, _ = load_data(dataset_dir, name_filter, camera_names, batch_size_train, batch_size_val, args['chunk_size'], args['skip_mirrored_data'], config['load_pretrain'], policy_class)
# save dataset stats # save dataset stats
stats_path = os.path.join(ckpt_dir, f'dataset_stats.pkl') stats_path = os.path.join(ckpt_dir, f'dataset_stats.pkl')
@ -151,6 +164,8 @@ def make_policy(policy_class, policy_config):
policy = ACTPolicy(policy_config) policy = ACTPolicy(policy_config)
elif policy_class == 'CNNMLP': elif policy_class == 'CNNMLP':
policy = CNNMLPPolicy(policy_config) policy = CNNMLPPolicy(policy_config)
elif policy_class == 'Diffusion':
policy = DiffusionPolicy(policy_config)
else: else:
raise NotImplementedError raise NotImplementedError
return policy return policy
@ -161,6 +176,8 @@ def make_optimizer(policy_class, policy):
optimizer = policy.configure_optimizers() optimizer = policy.configure_optimizers()
elif policy_class == 'CNNMLP': elif policy_class == 'CNNMLP':
optimizer = policy.configure_optimizers() optimizer = policy.configure_optimizers()
elif policy_class == 'Diffusion':
optimizer = policy.configure_optimizers()
else: else:
raise NotImplementedError raise NotImplementedError
return optimizer return optimizer
@ -214,7 +231,10 @@ def eval_bc(config, ckpt_name, save_episode=True, num_rollouts=50):
stats = pickle.load(f) stats = pickle.load(f)
pre_process = lambda s_qpos: (s_qpos - stats['qpos_mean']) / stats['qpos_std'] pre_process = lambda s_qpos: (s_qpos - stats['qpos_mean']) / stats['qpos_std']
post_process = lambda a: a * stats['action_std'] + stats['action_mean'] if policy_class == 'Diffusion':
post_process = lambda a: ((a + 1) / 2) * (stats['action_max'] - stats['action_min']) + stats['action_min']
else:
post_process = lambda a: a * stats['action_std'] + stats['action_mean']
# load environment # load environment
if real_robot: if real_robot:
@ -306,6 +326,10 @@ def eval_bc(config, ckpt_name, save_episode=True, num_rollouts=50):
raw_action = (actions_for_curr_step * exp_weights).sum(dim=0, keepdim=True) raw_action = (actions_for_curr_step * exp_weights).sum(dim=0, keepdim=True)
else: else:
raw_action = all_actions[:, t % query_frequency] raw_action = all_actions[:, t % query_frequency]
elif config['policy_class'] == "Diffusion":
if t % query_frequency == 0:
all_actions = policy(qpos, curr_image)
raw_action = all_actions[:, t % query_frequency]
elif config['policy_class'] == "CNNMLP": elif config['policy_class'] == "CNNMLP":
raw_action = policy(qpos, curr_image) raw_action = policy(qpos, curr_image)
else: else:
@ -416,7 +440,7 @@ def train_bc(train_dataloader, val_dataloader, config):
if epoch_val_loss < min_val_loss: if epoch_val_loss < min_val_loss:
min_val_loss = epoch_val_loss min_val_loss = epoch_val_loss
best_ckpt_info = (step, min_val_loss, deepcopy(policy.state_dict())) best_ckpt_info = (step, min_val_loss, deepcopy(policy.state_dict()))
for k in validation_summary.keys(): for k in list(validation_summary.keys()):
validation_summary[f'val_{k}'] = validation_summary.pop(k) validation_summary[f'val_{k}'] = validation_summary.pop(k)
wandb.log(validation_summary, step=step) wandb.log(validation_summary, step=step)
print(f'Val loss: {epoch_val_loss:.5f}') print(f'Val loss: {epoch_val_loss:.5f}')

165
policy.py
查看文件

@ -2,11 +2,174 @@ import torch.nn as nn
from torch.nn import functional as F from torch.nn import functional as F
import torchvision.transforms as transforms import torchvision.transforms as transforms
import torch import torch
import numpy as np
from detr.main import build_ACT_model_and_optimizer, build_CNNMLP_model_and_optimizer from detr.main import build_ACT_model_and_optimizer, build_CNNMLP_model_and_optimizer
import IPython import IPython
e = IPython.embed e = IPython.embed
from collections import OrderedDict
from robomimic.models.base_nets import ResNet18Conv, SpatialSoftmax
from robomimic.algo.diffusion_policy import replace_bn_with_gn, ConditionalUnet1D
from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
from diffusers.schedulers.scheduling_ddim import DDIMScheduler
from diffusers.training_utils import EMAModel
class DiffusionPolicy(nn.Module):
def __init__(self, args_override):
super().__init__()
self.camera_names = args_override['camera_names']
self.observation_horizon = args_override['observation_horizon'] ### TODO TODO TODO DO THIS
self.action_horizon = args_override['action_horizon'] # apply chunk size
self.prediction_horizon = args_override['prediction_horizon'] # chunk size
self.num_inference_timesteps = args_override['num_inference_timesteps']
self.ema_power = args_override['ema_power']
self.lr = args_override['lr']
self.weight_decay = 0
self.num_kp = 32
self.feature_dimension = 64
self.ac_dim = args_override['action_dim'] # 14 + 2
self.obs_dim = self.feature_dimension * len(self.camera_names) + 14 # camera features and proprio
backbones = []
pools = []
linears = []
for _ in self.camera_names:
backbones.append(ResNet18Conv(**{'input_channel': 3, 'pretrained': False, 'input_coord_conv': False}))
pools.append(SpatialSoftmax(**{'input_shape': [512, 15, 20], 'num_kp': self.num_kp, 'temperature': 1.0, 'learnable_temperature': False, 'noise_std': 0.0}))
linears.append(torch.nn.Linear(int(np.prod([self.num_kp, 2])), self.feature_dimension))
self.backbones = nn.ModuleList(backbones)
self.pools = nn.ModuleList(pools)
self.linears = nn.ModuleList(linears)
self.backbones = replace_bn_with_gn(self.backbones) # TODO
self.noise_pred_net = ConditionalUnet1D(
input_dim=self.ac_dim,
global_cond_dim=self.obs_dim*self.observation_horizon
)
nets = nn.ModuleDict({
'policy': nn.ModuleDict({
'backbones': self.backbones,
'pools': self.pools,
'linears': self.linears,
'noise_pred_net': self.noise_pred_net
})
})
nets = nets.float()
self.ema = EMAModel(model=nets, power=self.ema_power)
self.nets = nets
# setup noise scheduler
self.noise_scheduler = DDIMScheduler(
num_train_timesteps=50,
beta_schedule='squaredcos_cap_v2',
clip_sample=True,
set_alpha_to_one=True,
steps_offset=0,
prediction_type='epsilon'
)
def configure_optimizers(self):
optimizer = torch.optim.AdamW(self.nets.parameters(), lr=self.lr, weight_decay=self.weight_decay)
return optimizer
def __call__(self, qpos, image, actions=None, is_pad=None):
B = qpos.shape[0]
if actions is not None: # training time
all_features = []
for cam_id in range(len(self.camera_names)):
cam_image = image[:, cam_id]
cam_features = self.backbones[cam_id](cam_image)
pool_features = self.pools[cam_id](cam_features)
pool_features = torch.flatten(pool_features, start_dim=1)
out_features = self.linears[cam_id](pool_features)
all_features.append(out_features)
obs_cond = torch.cat(all_features + [qpos], dim=1)
# sample noise to add to actions
noise = torch.randn(actions.shape, device=obs_cond.device)
# sample a diffusion iteration for each data point
timesteps = torch.randint(
0, self.noise_scheduler.config.num_train_timesteps,
(B,), device=obs_cond.device
).long()
# add noise to the clean actions according to the noise magnitude at each diffusion iteration
# (this is the forward diffusion process)
noisy_actions = self.noise_scheduler.add_noise(
actions, noise, timesteps)
# predict the noise residual
noise_pred = self.noise_pred_net(noisy_actions, timesteps, global_cond=obs_cond)
# L2 loss
loss = F.mse_loss(noise_pred, noise)
loss_dict = {}
loss_dict['l2_loss'] = loss
loss_dict['loss'] = loss
return loss_dict
else: # inference time
To = self.observation_horizon
Ta = self.action_horizon
Tp = self.prediction_horizon
action_dim = self.ac_dim
# select network ### TODO TODO
# nets = self.nets
# if self.ema is not None:
# nets = self.ema.averaged_model
all_features = []
for cam_id in range(len(self.camera_names)):
cam_image = image[:, cam_id]
cam_features = self.backbones[cam_id](cam_image)
pool_features = self.pools[cam_id](cam_features)
pool_features = torch.flatten(pool_features, start_dim=1)
out_features = self.linears[cam_id](pool_features)
all_features.append(out_features)
obs_cond = torch.cat(all_features + [qpos], dim=1)
# initialize action from Guassian noise
noisy_action = torch.randn(
(B, Tp, action_dim), device=obs_cond.device)
naction = noisy_action
# init scheduler
self.noise_scheduler.set_timesteps(self.num_inference_timesteps)
for k in self.noise_scheduler.timesteps:
# predict noise
noise_pred = self.noise_pred_net(
sample=naction,
timestep=k,
global_cond=obs_cond
)
# inverse diffusion step (remove noise)
naction = self.noise_scheduler.step(
model_output=noise_pred,
timestep=k,
sample=naction
).prev_sample
return naction
class ACTPolicy(nn.Module): class ACTPolicy(nn.Module):
def __init__(self, args_override): def __init__(self, args_override):
super().__init__() super().__init__()

查看文件

@ -11,7 +11,7 @@ import IPython
e = IPython.embed e = IPython.embed
class EpisodicDataset(torch.utils.data.Dataset): class EpisodicDataset(torch.utils.data.Dataset):
def __init__(self, dataset_path_list, camera_names, norm_stats, episode_ids, episode_len, chunk_size): def __init__(self, dataset_path_list, camera_names, norm_stats, episode_ids, episode_len, chunk_size, policy_class):
super(EpisodicDataset).__init__() super(EpisodicDataset).__init__()
self.episode_ids = episode_ids self.episode_ids = episode_ids
self.dataset_path_list = dataset_path_list self.dataset_path_list = dataset_path_list
@ -21,6 +21,7 @@ class EpisodicDataset(torch.utils.data.Dataset):
self.chunk_size = chunk_size self.chunk_size = chunk_size
self.cumulative_len = np.cumsum(self.episode_len) self.cumulative_len = np.cumsum(self.episode_len)
self.max_episode_len = max(episode_len) self.max_episode_len = max(episode_len)
self.policy_class = policy_class
self.__getitem__(0) # initialize self.is_sim self.__getitem__(0) # initialize self.is_sim
self.is_sim = False self.is_sim = False
@ -101,7 +102,14 @@ class EpisodicDataset(torch.utils.data.Dataset):
# normalize image and change dtype to float # normalize image and change dtype to float
image_data = image_data / 255.0 image_data = image_data / 255.0
action_data = (action_data - self.norm_stats["action_mean"]) / self.norm_stats["action_std"]
if self.policy_class == 'Diffusion':
# normalize to [-1, 1]
action_data = ((action_data - self.norm_stats["action_min"]) / (self.norm_stats["action_max"] - self.norm_stats["action_min"])) * 2 - 1
else:
# normalize to mean 0 std 1
action_data = (action_data - self.norm_stats["action_mean"]) / self.norm_stats["action_std"]
qpos_data = (qpos_data - self.norm_stats["qpos_mean"]) / self.norm_stats["qpos_std"] qpos_data = (qpos_data - self.norm_stats["qpos_mean"]) / self.norm_stats["qpos_std"]
except: except:
@ -150,7 +158,12 @@ def get_norm_stats(dataset_path_list):
qpos_std = all_qpos_data.std(dim=[0]).float() qpos_std = all_qpos_data.std(dim=[0]).float()
qpos_std = torch.clip(qpos_std, 1e-2, np.inf) # clipping qpos_std = torch.clip(qpos_std, 1e-2, np.inf) # clipping
action_min = all_action_data.min(dim=0).values.float()
action_max = all_action_data.max(dim=0).values.float()
eps = 0.0001
stats = {"action_mean": action_mean.numpy(), "action_std": action_std.numpy(), stats = {"action_mean": action_mean.numpy(), "action_std": action_std.numpy(),
"action_min": action_min.numpy() - eps,"action_max": action_max.numpy() + eps,
"qpos_mean": qpos_mean.numpy(), "qpos_std": qpos_std.numpy(), "qpos_mean": qpos_mean.numpy(), "qpos_std": qpos_std.numpy(),
"example_qpos": qpos} "example_qpos": qpos}
@ -167,7 +180,7 @@ def find_all_hdf5(dataset_dir, skip_mirrored_data):
return hdf5_files return hdf5_files
def load_data(dataset_dir, name_filter, camera_names, batch_size_train, batch_size_val, chunk_size, skip_mirrored_data=False, load_pretrain=False): def load_data(dataset_dir, name_filter, camera_names, batch_size_train, batch_size_val, chunk_size, skip_mirrored_data=False, load_pretrain=False, policy_class=None):
dataset_path_list = find_all_hdf5(dataset_dir, skip_mirrored_data) dataset_path_list = find_all_hdf5(dataset_dir, skip_mirrored_data)
dataset_path_list = [n for n in dataset_path_list if name_filter(n)] dataset_path_list = [n for n in dataset_path_list if name_filter(n)]
num_episodes = len(dataset_path_list) num_episodes = len(dataset_path_list)
@ -189,8 +202,8 @@ def load_data(dataset_dir, name_filter, camera_names, batch_size_train, batch_si
val_episode_len = [all_episode_len[i] for i in val_episode_ids] val_episode_len = [all_episode_len[i] for i in val_episode_ids]
# construct dataset and dataloader # construct dataset and dataloader
train_dataset = EpisodicDataset(dataset_path_list, camera_names, norm_stats, train_episode_ids, train_episode_len, chunk_size) train_dataset = EpisodicDataset(dataset_path_list, camera_names, norm_stats, train_episode_ids, train_episode_len, chunk_size, policy_class)
val_dataset = EpisodicDataset(dataset_path_list, camera_names, norm_stats, val_episode_ids, val_episode_len, chunk_size) val_dataset = EpisodicDataset(dataset_path_list, camera_names, norm_stats, val_episode_ids, val_episode_len, chunk_size, policy_class)
train_dataloader = DataLoader(train_dataset, batch_size=batch_size_train, shuffle=True, pin_memory=True, num_workers=1, prefetch_factor=1) train_dataloader = DataLoader(train_dataset, batch_size=batch_size_train, shuffle=True, pin_memory=True, num_workers=1, prefetch_factor=1)
val_dataloader = DataLoader(val_dataset, batch_size=batch_size_val, shuffle=True, pin_memory=True, num_workers=1, prefetch_factor=1) val_dataloader = DataLoader(val_dataset, batch_size=batch_size_val, shuffle=True, pin_memory=True, num_workers=1, prefetch_factor=1)