-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain.py
207 lines (172 loc) · 9.55 KB
/
train.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
'''
Author: Jiaheng Hu
Train the Network
'''
import torch
from utils import generate_true_data, calc_gradient_penalty, int_to_onehot, generate_true_regress_data, calc_reward_from_rnet
from params import get_params
from Networks.Generator import AllocationGenerator
from Networks.Discriminator import Discriminator
from torch.utils.tensorboard import SummaryWriter
import numpy as np
import os
import torch.nn.functional as F
from Networks.RewardNet import RewardNet
def train(
training_steps: int = 1500000,
learning_rate: float = 0.001,
print_output_every_n_steps: int = 5000,
n_critic: int = 5,
):
params = get_params()
batch_size = params['batch_size']
if params['sim_env']:
from MAETF.simulator import MultiAgentEnv
env_name = 'simenv'
else:
from toy_env import MultiAgentEnv
env_name = 'toyenv'
# environment for getting hand-crafted rewards
env = MultiAgentEnv(n_num_grids=params['env_grid_num'],
n_num_agents=params['n_agent_types'],
n_env_types=params['n_env_types'],
agent_num=params['agent_num'])
worker_device = torch.device("cuda:0")
# Models
generator = AllocationGenerator(
n_agent=params['n_agent_types'],
n_env_grid=params['env_grid_num'],
env_input_len=params['env_input_len'],
design_input_len=params['design_input_len'],
norm=params['gen_norm'],
layer_size=params['design_layer_size']).to(worker_device)
discriminator = Discriminator(alloc_length=params['env_grid_num'] * params['n_agent_types'],
env_size=params["env_input_len"],
norm=params['dis_norm']).to(worker_device)
if params['load_from_file']:
# load weight
exit("err: change the path (line 56 train.py)")
out_dir = "./gan_logs/"+params['load_from_file']
generator.load_state_dict(torch.load("./gan_logs/"+params['load_from_file']+"/generator_weight"))
discriminator.load_state_dict(torch.load("./gan_logs/"+params['load_from_file']+"/discriminator_weight"))
else:
out_dir = os.path.join(params['gan_loc'], "")
out_dir += '%s_rnet:%i' % (params['vary_env'], params['use_regress_net'])
out_dir += '_%s_latent:%d_%s' % (params['data_method'], params['design_input_len'], env_name)
# out_dir += '_%s_%s' % (params['gen_norm'], params['dis_norm'])
# out_dir += '_nsamp:%d' % (params['n_samples'])
# out_dir += '_%s_gpl:%g' % (params['dis_norm'], params['gp_lambda'])
# out_dir += '_atypes:%d_enum:%d_etypes:%d' % (
# params['n_agent_types'], params['env_grid_num'], params['n_env_types'])
if os.path.exists(out_dir):
cmd = 'rm %s/*' % out_dir
os.system(cmd)
if params['use_regress_net']:
reward_net = RewardNet(params['n_agent_types'],
env_length=params['n_env_types'],
norm=params['reward_norm'],
n_hidden_layers=5,
hidden_layer_size=256).to(worker_device)
reward_net.load_state_dict(torch.load(params['regress_net_loc']))
reward_net.eval()
# Optimizers
generator_optimizer = torch.optim.Adam(generator.parameters(), lr=learning_rate)
discriminator_optimizer = torch.optim.Adam(
discriminator.parameters(), lr=learning_rate
)
writer = SummaryWriter(log_dir=out_dir)
for i in range(training_steps):
# TODO: env type should be the actual env type: env_type pass into onehot
if params['vary_env'] == 'static':
env_type = [0, 1, 2, 3]
# env_type = [0, 1, 1, 0, 2, 1, 3, 2, 3]
elif params['vary_env'] == 'discrete':
env_type_list = [[0, 1, 2, 3], [1, 2, 3, 0], [2, 3, 0, 3]]
env_dex = np.random.randint(len(env_type_list))
env_type = env_type_list[env_dex]
elif params['vary_env'] == 'random':
env_type = np.random.choice(4, 4)
else:
exit("error train.py line 71")
env_onehot = torch.tensor(int_to_onehot(env_type, params['n_env_types']),
dtype=torch.float, device=worker_device)
env_onehot = env_onehot.reshape(1, -1).repeat(batch_size, 1)
# Create noisy input for generator
# noise = torch.rand((batch_size, params['design_input_len']), device=worker_device)
noise = torch.normal(0, 1, size=(batch_size, params['design_input_len']), device=worker_device)
generated_data_logits = generator(noise, env_onehot)
generated_data_raw = F.softmax(generated_data_logits, dim=-1)
generated_data = generated_data_raw.reshape(batch_size, -1)
# generated random data based on reward net
if params['use_regress_net']:
true_data, true_avg_r, true_raw_r = generate_true_regress_data(env, params['n_samples'],
env_type, reward_net,
data_method=params['data_method'],
fake_data=generated_data_logits)
else:
# if we use an idealized dataset
true_data, true_avg_r, true_raw_r = generate_true_data(env, params['n_samples'], env_type,
data_method=params['data_method'],
fake_data=generated_data_logits)
true_data = torch.tensor(true_data).float().to(worker_device)
if i % n_critic == 0:
# zero the gradients on each iteration
generator_optimizer.zero_grad()
# Train the generator
# We invert the labels here and don't train the discriminator because we want the generator
# to make things the discriminator classifies as true.
generator_discriminator_out = discriminator(generated_data, env_onehot)
# generator_loss = loss(generator_discriminator_out, true_labels)
generator_loss = - torch.mean(generator_discriminator_out)
generator_loss.backward()
generator_optimizer.step()
# Train the discriminator on the true/generated data
discriminator_optimizer.zero_grad()
true_discriminator_loss = discriminator(true_data, env_onehot).mean()
# true_discriminator_loss = loss(true_discriminator_out, true_labels)
# add .detach() here think about this
generator_discriminator_loss = discriminator(generated_data.detach(), env_onehot).mean()
gp = calc_gradient_penalty(discriminator, true_data, generated_data.detach(),
env_onehot, worker_device) * params['gp_lambda']
discriminator_loss_log = generator_discriminator_loss.detach() - true_discriminator_loss.detach()
discriminator_loss = generator_discriminator_loss - true_discriminator_loss + gp
discriminator_loss.backward()
discriminator_optimizer.step()
# log interval
if i % (2 * n_critic) == 0:
writer.add_scalar('Train' + '/generator_loss', generator_loss.mean(), i)
if params['vary_env'] == 'static' or params['vary_env'] == 'random':
writer.add_scalar('Disc' + '/generator_discriminator_loss',
generator_discriminator_loss.detach().mean(), i)
writer.add_scalar('Disc' + '/true_discriminator_loss',
true_discriminator_loss.detach().mean(), i)
writer.add_scalar('Train' + '/discriminator_loss',
discriminator_loss_log.mean(), i)
elif params['vary_env'] == 'discrete':
writer.add_scalar('Disc' + '/generator_discriminator_loss' + str(env_dex),
generator_discriminator_loss.detach().mean(), i)
writer.add_scalar('Disc' + '/true_discriminator_loss' + str(env_dex),
true_discriminator_loss.detach().mean(), i)
writer.add_scalar('Train' + '/discriminator_loss' + str(env_dex),
discriminator_loss_log.mean(), i)
if i % print_output_every_n_steps == 0:
print(f"episode: {i}")
print(f"env type is: {env_type}")
int_alloc = np.array([env.get_integer(alloc) for alloc in generated_data_raw.detach().cpu().numpy()])
for alloc in int_alloc[:5]:
print(alloc)
if params['sim_env']:
generated_rewards = calc_reward_from_rnet(env, reward_net, int_alloc, env_onehot, batch_size)
else:
generated_rewards = np.array([env.get_integer_reward(alloc, env_type) for alloc in int_alloc])
print(f"fake data average reward: {generated_rewards.mean()}")
writer.add_scalar('R' + '/fake_avg_reward', generated_rewards.mean(), i)
print(f"real data average reward: {true_avg_r}")
print(f"random sample average reward: {true_raw_r}")
writer.add_scalar('R' + '/true_avg_reward', true_avg_r, i)
writer.add_scalar('R' + '/true_raw_reward', true_raw_r, i)
torch.save(generator.state_dict(), os.path.join(out_dir, "generator_weight"))
torch.save(discriminator.state_dict(), os.path.join(out_dir, "discriminator_weight"))
return generator, discriminator
if __name__ == '__main__':
train()