-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathimitation_learning.py
168 lines (141 loc) · 6.52 KB
/
imitation_learning.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
from agent import PolicyNet
import logging
import hydra
import torch
import numpy as np
import ligent
import h5py
from torch.utils.data import TensorDataset, DataLoader
import torch.nn as nn
import torch.optim as optim
# from torch import multiprocessing as mp
from omegaconf import OmegaConf
from dotmap import DotMap
from hydra.utils import instantiate
from pyvirtualdisplay import Display
from tqdm import tqdm
# from gymnasium.wrappers import RecordEpisodeStatistics, ClipAction, \
# NormalizeObservation, TransformObservation, NormalizeReward, \
# TransformReward, RecordVideo
# from gym.wrappers import RecordVideo
import other_utils
# from agent.ppo import PPOAgent
# from buffer import ReplayBuffer, PrioritizedReplayBuffer, PPOReplayBuffer, get_buffer
# from taskEnv import ComeHereEnv
import time
logger = logging.getLogger(__name__)
device = "cuda" if torch.cuda.is_available() else "cpu"
torch.set_float32_matmul_precision('high')
# device='cpu'
def eval_env(env, agent, episodes, seed, action_decoder, env_decoder):
returns = []
distance_info_s = []
ligent.set_scenes_dir("./custom_scenes")
# eval_start = time.time()
# print("Start eval!", flush=True)
for episode in range(episodes):
# state, _ = env.reset(seed=np.random.randint(0, 10000) + seed)
state_img, state_text = env.reset()
env_decoder.reset()
done, blocked = False, False
while not (done or blocked):
state_img = np.expand_dims(state_img, 0)
state_text = np.expand_dims(state_text, 0)
action = agent.get_action((state_img,state_text), sample=False)
action_env = action_decoder.decode(action)
# state, _, _, info = env.step(action_env)
(state_img, state_text), _, _, info = env.step(**action_env)
reward, done, blocked, cumulate_reward, elpased_step, distance_info = env_decoder.step(info)
# returns.append(info['episode']['r'].item())
returns.append(cumulate_reward)
distance_info_s.append(distance_info)
# print(f"eval costs {time.time()-eval_start} s!", flush=True)
ligent.set_scenes_dir("")
return np.mean(returns), np.std(returns), distance_info_s
@torch.no_grad()
def eval(model, data_loader, criterion):
model.eval()
total_loss = 0
total_correct = 0
total_samples = 0
for obs_Vs, labels in data_loader:
obs_T = torch.zeros((len(obs_Vs), 520), device=device)
logits = model(obs_Vs, obs_T)
loss = criterion(logits, labels)
total_loss += loss.item()
_, predicted = torch.max(logits,-1)
total_correct += (predicted == labels).sum().item()
total_samples += labels.size(0)
accuracy = total_correct/total_samples
average_loss = total_loss/len(data_loader)
return average_loss, accuracy
def train(cfg, seed: int, log_dict: dict, logger: logging.Logger, train_loader, eval_loader):
# env = ligent.Environment(path="/home/liuan/workspace/drl_project/ligent-linux-server/LIGENT.x86_64")
# env_decoder = ComeHereEnv(distance_reward=10, success_reward=200, distance_min=1.2, step_penalty=1, episode_len=500, is_debug=True)
# action_decoder = instantiate(cfg.action_decoder, device=device)
other_utils.set_seed_everywhere("", seed)
# eval_env_decoder = ComeHereEnv(distance_reward=10, success_reward=200, distance_min=1.2, step_penalty=1, episode_len=100, is_debug=True)
# state_size = other_utils.get_space_shape(env.observation_space, is_vector_env=cfg.vec_envs > 1)
# action_size = other_utils.get_space_shape(env.action_space, is_vector_env=cfg.vec_envs > 1)
feature_net = instantiate(cfg.feature_net, device=device)
agent = instantiate(cfg.agent, preprocess_net=feature_net, device=device)
cfg = DotMap(OmegaConf.to_container(cfg.train, resolve=True))
model = PolicyNet(feature_net=agent.get_feature_net(), actor_net=agent.get_actor_net())
# model = PolicyNet(feature_net=feature_net, )
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=3e-4)
best_eval_acc = 0
not_ascending_epoch = 0
epoch_cnt = 0
while True:
epoch_cnt += 1
running_loss = 0.0
for obs_V, labels in tqdm(train_loader):
optimizer.zero_grad()
obs_T = torch.zeros((len(obs_V), 520), device=device)
outputs = model(obs_V, obs_T)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
eval_loss, eval_acc = eval(model, eval_loader, criterion=criterion)
logger.info(f'epoch [{epoch_cnt}] training_loss: {running_loss / len(train_loader):.3f}, eval_loss: {eval_loss}, eval_acc: {eval_acc}')
running_loss = 0.0
model.train()
agent.save(f'epoch_{epoch_cnt}_')
if eval_acc > best_eval_acc:
best_eval_acc = eval_acc
not_ascending_epoch = 0
agent.save(f'best_acc_')
else:
not_ascending_epoch += 1
if not_ascending_epoch >= 5:
break
def get_dataloader(f_path="/home/liuan/workspace/drl_project/ligentAgent/dataset/Episode1000.h5"):
with h5py.File(f_path, 'r') as f:
# Get the datasets
obs_V_dataset = f['obs_V']
action_dataset = f['action']
# Convert the datasets to numpy arrays
obs_V_dataset = obs_V_dataset[:]
action_dataset = action_dataset[:,0]
# Convert the numpy arrays to PyTorch Tensors
obs_V_tensor = torch.from_numpy(obs_V_dataset).to(device)
action_tensor = torch.from_numpy(action_dataset).to(device).type(torch.long)
random_idx = torch.randperm(action_tensor.size(0), device=device)
train_len = int(len(random_idx)*0.7)
train_dataset = TensorDataset(obs_V_tensor[random_idx[:train_len]], action_tensor[random_idx[:train_len]])
eval_dataset = TensorDataset(obs_V_tensor[random_idx[train_len:]], action_tensor[random_idx[train_len:]])
# Create a DataLoader from the TensorDataset
train_dataloader = DataLoader(train_dataset, batch_size=128, shuffle=True)
eval_dataloader = DataLoader(eval_dataset, batch_size=128, shuffle=False)
return train_dataloader, eval_dataloader
@hydra.main(config_path="cfgs", config_name="config", version_base="1.3")
def main(cfg):
log_dict = other_utils.get_log_dict(cfg.agent._target_)
for seed in cfg.seeds:
with torch.autograd.set_detect_anomaly(True):
train(cfg, seed, log_dict, logger, *(get_dataloader()))
if __name__=="__main__":
with Display(visible=False) as disp:
main()