forked from aaravpandya/RL-GFootball
-
Notifications
You must be signed in to change notification settings - Fork 0
/
core.py
215 lines (172 loc) · 7.51 KB
/
core.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
import os
# import gym
import torch
import torch.nn as nn
from stable_baselines3.common.policies import ActorCriticPolicy
from stable_baselines3.common.vec_env import SubprocVecEnv
from stable_baselines3 import PPO2, logger
from stable_baselines3.common import monitor
import gfootball.env as football_env
def residual_block(inputs: torch.Tensor, depth: int) -> torch.Tensor:
out = nn.Conv2D(depth, 3, 1, padding='same')(inputs)
out = nn.ReLU()(out)
out = nn.Conv2D(depth, 3, 1, padding='same')(out)
return out + inputs
def split_heads(x, num_heads):
"""Split the last dimension into (num_heads, depth).
Transpose the result such that the shape is (batch_size, num_heads, seq_len, depth)
"""
c = torch.broadcast_to(x, [num_heads, torch.shape(x)[0], x.shape[1], x.shape[2]])
return torch.transpose(c, perm=[1, 0, 2, 3])
def mlp(x):
y = nn.Dense(384, activation='relu')(x)
y = nn.Dense(384, activation='relu')(y)
y = nn.Dense(54, activation='relu')(y)
return y
# def layer_norm(x):
# return LayerNormalization()(x)
# def conv_lstm(x):
# tf.keras.layers.ConvLSTM2D(96,3,1,padding='same',)
def mhdpa(v, k, q, num_heads):
# batch_size = tf.shape(q)[0]
# q = self.wq(q) # (batch_size, seq_len, d_model)
# k = self.wk(k) # (batch_size, seq_len, d_model)
# v = self.wv(v) # (batch_size, seq_len, d_model)
wq = nn.Dense(18)(q)
wk = nn.Dense(18)(k)
wv = nn.Dense(18)(v)
# wq = LayerNormalization()(wq)
# wk = LayerNormalization()(wk)
# wv = LayerNormalization()(wv)
print("wq" + str(wq.shape))
wq = split_heads(wq, num_heads) # (batch_size, num_heads, seq_len_q, depth)
wk = split_heads(wk, num_heads) # (batch_size, num_heads, seq_len_k, depth)
wv = split_heads(wv, num_heads) # (batch_size, num_heads, seq_len_v, depth)
# scaled_attention.shape == (batch_size, num_heads, seq_len_q, depth)
# attention_weights.shape == (batch_size, num_heads, seq_len_q, seq_len_k)
scaled_attention = nn.functional.scaled_dot_product_attention(
wq, wk, wv)
print("sca " + str(scaled_attention))
# (batch_size, seq_len_q, num_heads, depth)
scaled_attention = torch.transpose(scaled_attention, perm=[0, 2, 1, 3])
print("sca_ " + str(scaled_attention))
concat_attention = torch.reshape(scaled_attention,
(torch.shape(scaled_attention)[0], scaled_attention.shape[1], scaled_attention.shape[2]*scaled_attention.shape[3])) # (batch_size, seq_len_q, d_model)
print("concat" + str(concat_attention))
# output = mlp(concat_attention)
# output = output + q
output = nn.Dense(scaled_attention.shape[2]*scaled_attention.shape[3])(concat_attention)
# output = LayerNorma1lization()(output)
return output
def build_impala_cnn(unscaled_images, depths=[16, 32, 32], **conv_kwargs):
"""
Model used in the paper "IMPALA: Scalable Distributed Deep-RL with
Importance Weighted Actor-Learner Architectures" https://arxiv.org/abs/1802.01561
"""
layer_num = 0
def get_layer_num_str():
nonlocal layer_num
num_str = str(layer_num)
layer_num += 1
return num_str
def conv_layer(out, depth):
return nn.functinal.conv2d(out, depth, 3, padding='same', name='layer_' + get_layer_num_str())
def residual_block(inputs):
depth = inputs.get_shape()[-1].value
out = nn.relu(inputs)
out = conv_layer(out, depth)
out = nn.relu(out)
out = conv_layer(out, depth)
return out + inputs
def conv_sequence(inputs, depth):
out = conv_layer(inputs, depth)
out = nn.functional.max_pool2d(out, pool_size=3, strides=2, padding='same')
out = residual_block(out)
out = residual_block(out)
return out
out = torch.cast(unscaled_images, torch.float32) / 255.
for depth in depths:
out = conv_sequence(out, depth)
out = torch.reshape(
out, (torch.shape(out)[0], out.shape[1]*out.shape[2], out.shape[3]))
print(out.shape)
out = mhdpa(out, out, out, 3)
out = nn.Flatten()(out)
out = nn.functional.relu(out)
out = nn.functional.dense(out, 256, activation=nn.ReLU)
return out
# Custom MLP policy of three layers of size 128 each for the actor and 2 layers of 32 for the critic,
# with a nature_cnn feature extractor
class CustomPolicy(ActorCriticPolicy):
def __init__(self, sess, ob_space, ac_space, n_env, n_steps, n_batch, reuse=False, **kwargs):
super(CustomPolicy, self).__init__(sess, ob_space, ac_space,
n_env, n_steps, n_batch, reuse=reuse, scale=True)
activation = nn.ReLU()
print("processed obs" + str(self.processed_obs.shape))
extracted_features = build_impala_cnn(self.processed_obs, **kwargs)
print("extracted_features"+str(extracted_features.shape))
extracted_features = nn.flatten(extracted_features)
print("ex2 " + str(extracted_features.shape))
pi_h = extracted_features
for i, layer_size in enumerate([128, 128, 128]):
pi_h = activation(nn.Dense(pi_h, layer_size, name='pi_fc' + str(i)))
pi_latent = pi_h
vf_h = extracted_features
for i, layer_size in enumerate([32, 32]):
vf_h = activation(nn.Dense(vf_h, layer_size, name='vf_fc' + str(i)))
value_fn = nn.Dense(vf_h, 1, name='vf')
vf_latent = vf_h
self._proba_distribution, self._policy, self.q_value = \
self.pdtype.proba_distribution_from_latent(
pi_latent, vf_latent, init_scale=0.01)
self._value_fn = value_fn
self._setup_init()
def step(self, obs, state=None, mask=None, deterministic=False):
if deterministic:
action, value, neglogp = self.sess.run([self.deterministic_action, self.value_flat, self.neglogp],
{self.obs_ph: obs})
else:
action, value, neglogp = self.sess.run([self.action, self.value_flat, self.neglogp],
{self.obs_ph: obs})
return action, value, self.initial_state, neglogp
def proba_step(self, obs, state=None, mask=None):
return self.sess.run(self.policy_proba, {self.obs_ph: obs})
def value(self, obs, state=None, mask=None):
return self.sess.run(self.value_flat, {self.obs_ph: obs})
def create_single_football_env(seed):
"""Creates gfootball environment."""
env = football_env.create_environment(env_name="academy_3_vs_1_with_keeper", stacked=True,
representation='extracted', render=False and (seed == 0), channel_dimensions=(64, 64), rewards='scoring,checkpoints')
env = monitor.Monitor(env, logger.get_dir()
and os.path.join(logger.get_dir(), str(seed)))
return env
def callback(_, _t):
model.save("Atari.pkl")
print("model saved")
return True
vec_env = SubprocVecEnv([
(lambda _i=i: create_single_football_env(_i))
for i in range(8)])
# env = make_atari_env('SeaquestNoFrameskip-v0', num_env=4, seed=0)
# # Frame-stacking with 4 frames
# env = VecFrameStack(env, n_stack=4)
model = PPO2(CustomPolicy, vec_env, verbose=1,tensorboard_log="Atari",full_tensorboard_log=True,
learning_rate=0.000343,n_steps=512,nminibatches=4,gamma=0.993,cliprange=0.08,
ent_coef=0.03,lam=0.95,noptepochs=4,max_grad_norm=0.64)
# # model.setup_model()
# # model = PPO2.load("3_vs_1_Sep_3_full_mhdpa.pkl")
# # model.set_env(vec_env)
# # model.setup_model()
model.learn(total_timesteps=20000000, callback=callback,
tb_log_name="Atari", reset_num_timesteps=False)
# model = PPO2.load("Hopefully_this_works_Continued_2.pkl")
# env = football_env.create_environment(env_name="academy_3_vs_1_with_keeper",stacked=False,render=True
# ,channel_dimensions=(96,96), logdir="",rewards='scoring,checkpoints')
# env.
# obs = env.reset()
# while True:
# action, _states = model.predict(obs)
# obs, rewards, dones, info = env.step(action)
# # env.render()
# if(dones):
# obs = env.reset()