-
Notifications
You must be signed in to change notification settings - Fork 0
/
a2c_torch.py
348 lines (285 loc) · 13.7 KB
/
a2c_torch.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
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
"""
Created by arvindsrikantan on 2018-04-23
"""
import os
import gym_minigrid
gym_minigrid
from reinforce_torch import Reinforce
import argparse
import sys
import gym
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.utils.data
from torch.autograd import Variable
import numpy as np
import pickle
def to_categorical(y, num_classes):
""" 1-hot encodes a tensor """
return np.eye(num_classes, dtype='float32')[y]
class Actor(nn.Module):
def __init__(self, env_name, vocab, max_len, with_language=True):
super(Actor, self).__init__()
self.temp_env = gym.make(env_name)
self.inp_shape = self.temp_env.observation_space.spaces['image'].shape
self.with_language = with_language
if self.inp_shape is None:
self.inp_shape = self.temp_env.observation_space.spaces['image'].spaces['image'].shape
self.conv = nn.Conv2d(self.inp_shape[-1], 4, (2, 2))
self.linear = nn.Linear(36, 24)
if self.with_language:
self.embedding = nn.Embedding(len(vocab), 24)
self.lstm = nn.LSTM(24, 12, batch_first=True)
self.linear = nn.Linear(36 + (max_len * 12), 24)
self.dense = nn.Linear(24, self.temp_env.action_space.n)
def forward(self, inputs):
# might run into problems cos of np resize
vis_input, txt_input = Variable(torch.from_numpy(inputs[0].transpose((0, 3, 1, 2)))), \
Variable(torch.from_numpy(np.array(inputs[1]).reshape((np.array(inputs[1]).shape[0],-1))))
cnn_outputs = F.max_pool2d(F.relu(self.conv(vis_input)), (2, 2))
flat_cnn_out = cnn_outputs.view(cnn_outputs.shape[0], -1)
linear_in = flat_cnn_out
if self.with_language:
embed_out = self.embedding(txt_input)
lstm_out, _ = self.lstm(embed_out)
concat_out = torch.cat((flat_cnn_out, lstm_out.contiguous().view(txt_input.shape[0], -1)), dim=1)
linear_in = concat_out
linear1_out = F.relu(self.linear(linear_in))
out = F.softmax(self.dense(linear1_out))
return out
class Critic(nn.Module):
def __init__(self, env_name, vocab, max_len, with_language=True):
super(Critic, self).__init__()
self.temp_env = gym.make(env_name)
self.with_language = with_language
self.inp_shape = self.temp_env.observation_space.spaces['image'].shape
if self.inp_shape is None:
self.inp_shape = self.temp_env.observation_space.spaces['image'].spaces['image'].shape
self.conv = nn.Conv2d(self.inp_shape[-1], 4, (2, 2))
self.linear1 = nn.Linear(36, 24)
if self.with_language:
self.embedding = nn.Embedding(len(vocab), 24)
self.lstm = nn.LSTM(24, 12, batch_first=True)
self.linear1 = nn.Linear(36 + (max_len * 12), 24)
self.linear2 = nn.Linear(24, 30)
self.dense = nn.Linear(30, 1)
def forward(self, inputs):
# might run into problems cos of np resize
vis_input, txt_input = Variable(torch.from_numpy(inputs[0].transpose((0, 3, 1, 2)))), \
Variable(torch.from_numpy(inputs[1].reshape((inputs[1].shape[0],-1))))
cnn_outputs = F.max_pool2d(F.relu(self.conv(vis_input)), (2, 2))
flat_cnn_out = cnn_outputs.view(cnn_outputs.shape[0], -1)
linear_in = flat_cnn_out
if self.with_language:
embed_out = self.embedding(txt_input)
lstm_out, _ = self.lstm(embed_out)
concat_out = torch.cat((flat_cnn_out, lstm_out.contiguous().view(txt_input.shape[0], -1)), dim=1)
linear_in = concat_out
linear1_out = F.relu(self.linear1(linear_in))
linear2_out = F.relu(self.linear2(linear1_out))
out = self.dense(linear2_out)
return out
class A2C(Reinforce):
# Implementation of N-step Advantage Actor Critic.
# This class inherits the Reinforce class, so for example, you can reuse
# generate_episode() here.
def __init__(self, model, lr, critic_model, critic_lr, vocab, max_len, n=20):
"""
Initializes A2C.
:param model: The actor model.
:param lr: Learning rate for the actor model.
:param critic_model: The critic model.
:param critic_lr: Learning rate for the critic model.
:param n: The value of N in N-step A2C.
"""
self.critic_model = critic_model
self.max_len = max_len
self.n = n
self.lr = lr
super().__init__(model, lr, vocab, max_len)
self.critic_optimizer = optim.Adam(self.critic_model.parameters(), lr=critic_lr)
def mse_loss_func(pred, target):
loss = torch.sum((pred - target) ** 2)
return loss
self.critic_loss = nn.MSELoss(size_average=True) # mse_loss_func
def train(self, env, episodes, env_name, gamma=1.0, render=False, reward_scale=1.0, without_mission=False):
checkpointing = 500
test_rewards = []
train_rewards = []
power_gamma = {k: gamma ** k for k in range(10000)}
for episode in range(episodes + 1):
if episode % checkpointing == 0:
# Checkpoint
self.save_weights("../pickles/a2c/%s/checkpoint/%s_n_%s_iter_%s.h5" % (env_name, "%s", self.n, episode))
test_reward = []
for _ in range(100):
_, _, rewards = self.generate_episode(env, reward_scale)
test_reward += [sum(rewards) * reward_scale]
test_rewards.append((np.array(test_reward).mean(), np.array(test_reward).std()))
print("Average test rewards = %s" % (str(test_rewards[-1])))
np.save("../pickles/a2c/%s/test-rewards/n_%s_iter_%s.npy" % (env_name, self.n, episode),
np.array(test_rewards))
states, actions, rewards = self.generate_episode(env, reward_scale, render=render)
r = np.zeros(len(rewards))
g = np.zeros(len(rewards))
T = len(rewards)
if without_mission:
states_transformed = np.array(states)
else:
im, descr = zip(*states)
# descr = self.padding(descr)
states_transformed = [np.array(im), np.array(descr)]
self.critic_model.eval()
v = self.critic_model(states_transformed).data.numpy().flatten()
for t in reversed(range(T)):
v_end = 0 if (t + self.n >= T) else v[t + self.n]
r[t] = power_gamma[self.n] * v_end + sum(
[(power_gamma[k] * rewards[t + k] if (t + k < T) else 0) for k in range(self.n)])
g[t] = r[t] - v[t]
self.optimizer.zero_grad()
self.model.train()
model_out = self.model(states_transformed)
loss = self.custom_loss(
Variable(torch.from_numpy(to_categorical(actions, num_classes=env.action_space.n))),
model_out,
Variable(torch.from_numpy(g.astype("float32")))
)
loss.backward()
self.optimizer.step()
self.critic_optimizer.zero_grad()
self.critic_model.train()
critic_model_out = self.critic_model(states_transformed)
critic_loss = self.critic_loss(critic_model_out, Variable(torch.from_numpy(r.astype("float32"))))
critic_loss.backward()
self.critic_optimizer.step()
print("Episode %6d's, Steps = %3d, loss = %+.5f, critic_loss = %+.5f, cumulative reward:%+5.5f" % (
episode, len(states), loss.data[0], critic_loss.data[0],
sum(rewards) * reward_scale))
train_rewards.append(sum(rewards) * reward_scale)
np.save("../pickles/a2c/%s/n_%s_train-rewards.npy" % (env_name, self.n), np.array(train_rewards))
def padding(self, instructions):
out = []
for instr in instructions:
if len(instr) < self.max_len:
padding = list(np.zeros(self.max_len - len(instr)))
instr.extend(padding)
out.append(instr)
return out
def save_weights(self, name):
torch.save(self.model.state_dict(), name % "actor")
torch.save(self.critic_model.state_dict(), name % "critic")
def load_weights(self, name):
checkpoint = torch.load(name % "actor")
if isinstance(checkpoint, dict) and 'state_dict' in checkpoint:
self.model.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint)
checkpoint = torch.load(name % "critic")
if isinstance(checkpoint, dict) and 'state_dict' in checkpoint:
self.critic_model.load_state_dict(checkpoint['state_dict'])
else:
self.critic_model.load_state_dict(checkpoint)
def createDirectories(l):
for l in l:
if not os.path.exists(l):
os.makedirs(l)
def max_len(envs):
max_length = 0
for env_name in envs:
env = gym.make(env_name)
for i in range(100000):
state = env.reset()
max_length = max(max_length, len(state['mission']))
return max_length
def parse_arguments():
# Command-line flags are defined here.
parser = argparse.ArgumentParser()
parser.add_argument('--environment-name', dest='environment_name',
type=str, default='MiniGrid-Fetch-6x6-N2-v0',
help="Path to the actor model config file.")
parser.add_argument('--num-episodes', dest='num_episodes', type=int,
default=100000, help="Number of episodes to train on.")
parser.add_argument('--reward-scale', dest='reward_scale', type=float,
default=1, help="The scale factor for rewards")
parser.add_argument('--lr', dest='lr', type=float,
default=1e-3, help="The actor's learning rate.")
parser.add_argument('--critic-lr', dest='critic_lr', type=float,
default=1e-3, help="The critic's learning rate.") # 5e-4 before
parser.add_argument('--n', dest='n', type=int,
default=20, help="The value of N in N-step A2C.")
parser.add_argument('--gamma', dest='gamma', type=float,
default=0.99, help="The value of gamma in A2C.")
# https://stackoverflow.com/questions/15008758/parsing-boolean-values-with-argparse
parser_group = parser.add_mutually_exclusive_group(required=False)
parser_group.add_argument('--render', dest='render',
action='store_true',
help="Whether to render the environment.")
parser_group.add_argument('--no-render', dest='render',
action='store_false',
help="Whether to render the environment.")
parser.set_defaults(render=False)
parser_group = parser.add_mutually_exclusive_group(required=False)
parser_group.add_argument('--with-mission', dest='without_mission',
action='store_false',
help="Whether to use the mission string.")
parser_group.add_argument('--without-mission', dest='without_mission',
action='store_true',
help="Whether to use the mission string.")
parser.set_defaults(without_mission=False)
return parser.parse_args()
def main(args, load_models=None):
# Parse command-line arguments.
args = parse_arguments()
environment_name = args.environment_name
print("Running env: %s, with reward scaling of: %s" % (environment_name, args.reward_scale))
# Create the environment.
env = gym.make(environment_name)
dirs = [
# "../pickles/a2c/weights",
"../pickles/a2c/%s/checkpoint/" % environment_name,
"../pickles/a2c/%s/test-rewards/" % environment_name,
"../pickles/a2c/%s/test-rewards-lists/" % environment_name
]
createDirectories(dirs)
num_episodes = args.num_episodes
lr = args.lr
critic_lr = args.critic_lr
n = 20 # args.n
render = args.render
print(
"Training args: episodes=num_episodes, env_name=%s, render=%s, reward_scale=%s, without_mission=%s, gamma=%s" %
(environment_name, render, args.reward_scale, args.without_mission, args.gamma))
vocab = pickle.load(open('../data/vocab.p', 'rb'))
envs = ['MiniGrid-Empty-6x6-v0', 'MiniGrid-DoorKey-5x5-v0', 'MiniGrid-MultiRoom-N2-S4-v0',
'MiniGrid-Fetch-5x5-N2-v0',
'MiniGrid-GoToDoor-5x5-v0', 'MiniGrid-PutNear-6x6-N2-v0', 'MiniGrid-LockedRoom-v0']
max_length = 30 # max_len(envs)
# Load the actor model from file.
model = Actor(environment_name, vocab, max_len=max_length)
if torch.cuda.is_available():
model = model.cuda()
# Critic model
critic_model = Critic(environment_name, vocab, max_len=max_length)
if torch.cuda.is_available():
critic_model = critic_model.cuda()
# critic_model.summary()
# exit()
# TODO: Train the model using A2C and plot the learning curves.
a2c = A2C(model, lr, critic_model, critic_lr, vocab, max_len=max_length, n=n)
if load_models is not None:
a2c.load_weights(load_models)
print("Loaded")
a2c.train(env, episodes=num_episodes, env_name=environment_name, render=render, reward_scale=args.reward_scale,
without_mission=args.without_mission, gamma=args.gamma)
# for _n in [1, 20, 50, 100]:
# print("Starting for n=%s" % _n)
# get_test_rewards(env, a2c, n=_n)
#
# for _n in [1, 20, 50, 100]:
# print("Starting for n=%s" % _n)
# print(list(zip(*test_reward_with_error(n=_n))))
# plot()
if __name__ == '__main__':
main(sys.argv)