-
Notifications
You must be signed in to change notification settings - Fork 3
/
imitate_diffusion.py
240 lines (195 loc) · 8.4 KB
/
imitate_diffusion.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
import binascii
import os
import random
from copy import deepcopy
import numpy as np
import optuna
import torch
import torch.nn as nn
from diffusers.optimization import get_scheduler
from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
from diffusers.training_utils import EMAModel
from torch.utils.data import DataLoader
from bimanual_imitation.algorithms.abstract_base.imitate_bc_base import ImitateBcBase
from bimanual_imitation.algorithms.configs import ALG, DiffusionParamConfig
from bimanual_imitation.algorithms.core.diffusion.diffusion import ConditionalUnet1D
from bimanual_imitation.algorithms.core.shared.util import Timer
from irl_data.data_chunking import create_chunking_dataset
class ImitateDiffusion(ImitateBcBase):
@property
def alg(self):
return ALG.DIFFUSION
@property
def train_dataset(self):
return self._train_dataset
@property
def bc_policy(self):
return self._ema_noise_pred_net
@property
def device(self):
return torch.device("cuda" if torch.cuda.is_available() else "cpu")
def get_bc_policy_weights(self):
# Get a copy of the current ema model
ema_noise_pred_net = deepcopy(self._noise_pred_net)
self._ema.copy_to(ema_noise_pred_net.parameters())
return ema_noise_pred_net.state_dict()
def pre_eval(self):
# Note: these models are equal by reference. This may result in more EMA
# smoothing compared to only performing pre_eval once (after training)
# Alternatively, use deepcopy
self._ema_noise_pred_net = self._noise_pred_net
self._ema.copy_to(self._ema_noise_pred_net.parameters())
super().pre_eval()
def bc_policy_fn(self, obs_cond):
with torch.no_grad():
B = 1
# initialize action from Guassian noise
noisy_action = torch.randn(
(B, self.train_dataset.pred_horizon, self._action_dim_single), device=self.device
)
naction = noisy_action
# init scheduler
self._noise_scheduler.set_timesteps(self.cfg.num_diffusion_iters)
for k in self._noise_scheduler.timesteps:
# predict noise
noise_pred = self._ema_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
def init_params(self, cfg: DiffusionParamConfig):
random.seed(os.urandom(4))
seed = int(binascii.hexlify(os.urandom(4)), 16)
np.random.seed(seed)
torch.manual_seed(seed)
self._train_dataset = create_chunking_dataset(
environment=self.env_name,
stage="train",
pred_horizon=cfg.pred_horizon,
obs_horizon=cfg.obs_horizon,
action_horizon=cfg.action_horizon,
limit_trajs=self.limit_trajs,
normalize=True,
)
self._train_dataloader = DataLoader(
dataset=self._train_dataset,
batch_size=cfg.batch_size,
shuffle=True,
pin_memory=True,
# num_workers=1,
# prefetch_factor=1,
)
tmp_env = self.create_env(verbose=True)
self._obs_dim = tmp_env.observation_space.shape[0]
self._action_dim_single = tmp_env.action_space.shape[0] // cfg.pred_horizon
del tmp_env
# configure model
self._noise_pred_net = ConditionalUnet1D(
input_dim=self._action_dim_single, global_cond_dim=self._obs_dim
)
self._noise_scheduler = DDPMScheduler(
num_train_timesteps=cfg.num_diffusion_iters,
# the choise of beta schedule has big impact on performance
# we found squared cosine works the best
beta_schedule="squaredcos_cap_v2",
# clip output to [-1,1] to improve stability
clip_sample=True,
# our network predicts noise (instead of denoised action)
prediction_type="epsilon",
)
# device transfer
_ = self._noise_pred_net.to(self.device)
self._ema = EMAModel(parameters=self._noise_pred_net.parameters())
self._optimizer = torch.optim.AdamW(
params=self._noise_pred_net.parameters(),
lr=cfg.opt_learning_rate,
weight_decay=cfg.opt_weight_decay,
)
self._lr_scheduler = get_scheduler(
name=cfg.lr_scheduler,
optimizer=self._optimizer,
num_warmup_steps=cfg.lr_warmup_steps,
num_training_steps=len(self._train_dataloader) * self.max_iter,
)
self._total_time = 0
self._global_step = 0
self._epoch = 0
self.cfg = cfg
def suggest_hyperparams(self, trial: optuna.Trial):
param_cfg = DiffusionParamConfig(
batch_size=trial.suggest_categorical("batch_size", [128, 256, 512]),
num_diffusion_iters=trial.suggest_categorical("num_diffusion_iters", [50, 100]),
opt_learning_rate=trial.suggest_categorical("opt_learning_rate", [1e-4, 5e-5, 1e-5]),
opt_weight_decay=trial.suggest_categorical("opt_weight_decay", [1e-3, 1e-6]),
lr_warmup_steps=trial.suggest_categorical("lr_warmup_steps", [500, 1000]),
)
return param_cfg
def step_optimizer(self):
epoch_loss = []
with Timer() as t_train:
for batch_idx, nbatch in enumerate(self._train_dataloader):
# data normalized in dataset
# device transfer
nobs = nbatch["obs"].to(self.device)
naction = nbatch["action"].to(self.device)
B = nobs.shape[0]
# observation as FiLM conditioning
# (B, obs_horizon, obs_dim)
obs_cond = nobs[:, : self.train_dataset.obs_horizon, :]
# (B, obs_horizon * obs_dim)
obs_cond = obs_cond.flatten(start_dim=1)
# sample noise to add to actions
noise = torch.randn(naction.shape, device=self.device)
# sample a diffusion iteration for each data point
timesteps = torch.randint(
0, self._noise_scheduler.config.num_train_timesteps, (B,), device=self.device
).long()
# add noise to the clean images according to the noise magnitude at each diffusion iteration
# (this is the forward diffusion process)
noisy_actions = self._noise_scheduler.add_noise(naction, noise, timesteps)
# predict the noise residual
noise_pred = self._noise_pred_net(noisy_actions, timesteps, global_cond=obs_cond)
# L2 loss
loss = nn.functional.mse_loss(noise_pred, noise)
# optimize
loss.backward()
self._optimizer.step()
self._optimizer.zero_grad()
# step lr scheduler every batch
# this is different from standard pytorch behavior
self._lr_scheduler.step()
# update Exponential Moving Average of the model weights
self._ema.step(self._noise_pred_net.parameters())
# logging
loss_cpu = loss.item()
epoch_loss.append(loss_cpu)
self._global_step += 1
avg_loss = np.mean(epoch_loss)
self._total_time += t_train.dt
self._epoch += 1
iter_info = [
("iter", self._epoch, int),
("gstep", int(self._global_step), int),
("loss", avg_loss, float),
("lr", self._lr_scheduler.get_last_lr()[0], float),
("ttrain", t_train.dt, float),
("ttotal", self._total_time, float),
]
return iter_info
def get_default_run_options(self):
diffusion_defaults = {
"--mode": ("train_policy", str),
"--max_iter": (30, int),
"--num_evals": (10, int),
"--num_rollouts_per_eval": (10, int),
"--snapshot_save_freq": (3, int),
"--print_freq": (1, int),
"--limit_trajs": (200, int),
"--export_data": (False, bool),
}
return diffusion_defaults
if __name__ == "__main__":
ImitateDiffusion().run()