-
Notifications
You must be signed in to change notification settings - Fork 2
/
run_train_cmap.py
287 lines (230 loc) · 14.1 KB
/
run_train_cmap.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
import numpy as np
import tensorflow as tf
from tensorflow.contrib import slim
import environment
import expert
from model import CMAP
import os
import copy
import time
import cv2
flags = tf.app.flags
flags.DEFINE_string('maps', 'training-09x09-0127', 'Comma separated game environment list')
flags.DEFINE_string('logdir', './output/dummy', 'Log directory')
flags.DEFINE_boolean('debug', False, 'Save debugging information')
flags.DEFINE_boolean('multiproc', False, 'Multiproc environment')
flags.DEFINE_boolean('random_goal', True, 'Allow random goal')
flags.DEFINE_boolean('random_spawn', True, 'Allow random spawn')
flags.DEFINE_integer('max_steps_per_episode', 10 ** 100, 'Max steps per episode')
flags.DEFINE_integer('num_games', 10 ** 8, 'Number of games to play')
flags.DEFINE_integer('batch_size', 1, 'Number of environments to run')
flags.DEFINE_float('learning_rate', 0.001, 'ADAM learning rate')
flags.DEFINE_float('decay', 0.99, 'DAGGER decay')
FLAGS = flags.FLAGS
def DAGGER_train_step(sess, train_op, global_step, train_step_kwargs):
env = train_step_kwargs['env']
exp = train_step_kwargs['exp']
net = train_step_kwargs['net']
summary_writer = train_step_kwargs['summary_writer']
step_history = train_step_kwargs['step_history']
step_history_op = train_step_kwargs['step_history_op']
gradient_names = train_step_kwargs['gradient_names']
gradient_summary_op = train_step_kwargs['gradient_summary_op']
update_global_step_op = train_step_kwargs['update_global_step_op']
estimate_maps = train_step_kwargs['estimate_maps']
value_maps = train_step_kwargs['value_maps']
def _build_map_summary(estimate_maps, value_maps):
def _to_image(img):
return (np.expand_dims(np.squeeze(img), axis=2) * 255).astype(np.uint8)
est_maps = [tf.Summary.Value(tag='losses/free_space_estimates_{}'.format(scale),
image=tf.Summary.Image(
encoded_image_string=cv2.imencode('.png', image)[1].tostring(),
height=image.shape[0],
width=image.shape[1]))
for scale, map in enumerate(estimate_maps[-1])
for image in (_to_image(map),)]
val_maps = [tf.Summary.Value(tag='losses/values_{}'.format(scale),
image=tf.Summary.Image(
encoded_image_string=cv2.imencode('.png', image)[1].tostring(),
height=image.shape[0],
width=image.shape[1]))
for scale, map in enumerate(value_maps[-1])
for image in (_to_image(map),)]
return tf.Summary(value=est_maps + val_maps)
def _build_trajectory_summary(rate, loss, rewards_history, info_history, exp):
image = np.ones((28 + exp._width * 100, 28 + exp._height * 100, 3), dtype=np.uint8) * 255
def _node_to_game_coordinate(node):
row, col = node
return 14 + int((col - 0.5) * 100), 14 + int((row - 0.5) * 100)
def _pose_to_game_coordinate(pose):
x, y = pose[:2]
return 14 + int(x), 14 + image.shape[1] - int(y)
cv2.putText(image, exp._env_name, (0, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0))
for row, col in exp._walls:
loc = np.array([col, row])
points = [loc, loc + np.array([0, 1]),
loc + np.array([1, 1]), loc + np.array([1, 0])]
points = np.array([pts * 100 + np.array([14, 14]) for pts in points])
cv2.fillConvexPoly(image, points, (224, 172, 52))
for info in info_history:
cv2.circle(image, _node_to_game_coordinate(info['GOAL.LOC']), 10, (82, 82, 255), -1)
cv2.circle(image, _node_to_game_coordinate(info['SPAWN.LOC']), 10, (211, 111, 112), -1)
cv2.circle(image, _pose_to_game_coordinate(info['POSE']), 4, (63, 121, 255), -1)
encoded = cv2.imencode('.png', image)[1].tostring()
return tf.Summary(value=[tf.Summary.Value(tag='losses/trajectory',
image=tf.Summary.Image(encoded_image_string=encoded,
height=image.shape[0],
width=image.shape[1])),
tf.Summary.Value(tag='losses/supervision_rate', simple_value=rate),
tf.Summary.Value(tag='losses/average_loss_per_step', simple_value=loss),
tf.Summary.Value(tag='losses/reward', simple_value=sum(rewards_history))])
def _build_walltime_summary(begin, data, end):
return tf.Summary(value=[tf.Summary.Value(tag='time/DAGGER_eval_walltime', simple_value=(data - begin)),
tf.Summary.Value(tag='time/DAGGER_train_walltime', simple_value=(end - data)),
tf.Summary.Value(tag='time/DAGGER_complete_walltime', simple_value=(end - begin))])
def _build_gradient_summary(gradient_names, gradient_collections):
gradient_means = np.array(gradient_collections).mean(axis=0).tolist()
return tf.Summary(value=[tf.Summary.Value(tag='gradient/{}'.format(var), simple_value=val)
for var, val in zip(gradient_names, gradient_means)])
train_step_start = time.time()
np_global_step = sess.run(global_step)
random_rate = FLAGS.decay ** np_global_step
env.reset()
obs, info = env.observations()
optimal_action_history = [exp.get_optimal_action(info)]
observation_history = [obs]
egomotion_history = [[0., 0.]]
rewards_history = [0.]
estimate_maps_history = [[np.zeros((1, 64, 64, 3))] * net._estimate_scale]
info_history = [info]
estimate_maps_images = []
value_maps_images = []
# Dataset aggregation
terminal = False
while not terminal and len(info_history) < FLAGS.max_steps_per_episode:
_, previous_info = env.observations()
previous_info = copy.deepcopy(previous_info)
feed_dict = prepare_feed_dict(net.input_tensors, {'sequence_length': np.array([1]),
'visual_input': np.array([[observation_history[-1]]]),
'egomotion': np.array([[egomotion_history[-1]]]),
'reward': np.array([[rewards_history[-1]]]),
'estimate_map_list': estimate_maps_history[-1],
'is_training': False})
results = sess.run([net.output_tensors['action']] +
estimate_maps +
value_maps +
net.intermediate_tensors['estimate_map_list'], feed_dict=feed_dict)
predict_action = np.squeeze(results[0])
optimal_action = exp.get_optimal_action(previous_info)
dagger_action = random_rate * optimal_action + (1 - random_rate) * predict_action
action = np.argmax(dagger_action)
obs, reward, terminal, info = env.step(action)
optimal_action_history.append(copy.deepcopy(optimal_action))
observation_history.append(copy.deepcopy(obs))
egomotion_history.append(environment.calculate_egomotion(previous_info['POSE'], info['POSE']))
rewards_history.append(copy.deepcopy(reward))
estimate_maps_history.append([tensor[:, 0, :, :, :]
for tensor in results[1 + len(estimate_maps) + len(value_maps):]])
info_history.append(copy.deepcopy(info))
estimate_maps_images.append(results[1:1 + len(estimate_maps)])
value_maps_images.append(results[1 + len(estimate_maps):1 + len(estimate_maps) + len(value_maps)])
train_step_eval = time.time()
assert len(optimal_action_history) == len(observation_history) == len(egomotion_history) == len(rewards_history)
# Training
gradient_collections = []
cumulative_loss = 0
for i in xrange(0, len(optimal_action_history), FLAGS.batch_size):
batch_end_index = min(len(optimal_action_history), i + FLAGS.batch_size)
batch_size = batch_end_index - i
concat_observation_history = [observation_history[:batch_end_index]] * batch_size
concat_egomotion_history = [egomotion_history[:batch_end_index]] * batch_size
concat_reward_history = [rewards_history[:batch_end_index]] * batch_size
concat_optimal_action_history = optimal_action_history[i:batch_end_index]
concat_estimate_map_list = [np.zeros((batch_size, 64, 64, 3))] * net._estimate_scale
feed_dict = prepare_feed_dict(net.input_tensors, {'sequence_length': np.arange(i, batch_end_index) + 1,
'visual_input': np.array(concat_observation_history),
'egomotion': np.array(concat_egomotion_history),
'reward': np.array(concat_reward_history),
'optimal_action': np.array(concat_optimal_action_history),
'estimate_map_list': concat_estimate_map_list,
'is_training': True})
train_ops = [net.output_tensors['loss'], train_op] + gradient_summary_op
results = sess.run(train_ops, feed_dict=feed_dict)
cumulative_loss += results[0]
gradient_collections.append(results[2:])
cumulative_loss /= len(optimal_action_history)
train_step_end = time.time()
summary_text = ','.join('{}[{}]-{}={}'.format(key, idx, step, value)
for step, info in enumerate(info_history)
for key in ('GOAL.LOC', 'SPAWN.LOC', 'POSE', 'env_name')
for idx, value in enumerate(info[key]))
step_history_summary, new_global_step = sess.run([step_history_op, update_global_step_op],
feed_dict={step_history: summary_text})
summary_writer.add_summary(step_history_summary, global_step=np_global_step)
summary_writer.add_summary(_build_map_summary(estimate_maps_images, value_maps_images),
global_step=np_global_step)
summary_writer.add_summary(_build_gradient_summary(gradient_names, gradient_collections),
global_step=np_global_step)
summary_writer.add_summary(_build_trajectory_summary(random_rate, cumulative_loss,
rewards_history, info_history, exp),
global_step=np_global_step)
summary_writer.add_summary(_build_walltime_summary(train_step_start, train_step_eval, train_step_end),
global_step=np_global_step)
should_stop = new_global_step >= FLAGS.num_games
return cumulative_loss, should_stop
def prepare_feed_dict(tensors, data):
feed_dict = {}
for k, v in tensors.iteritems():
if k not in data:
continue
if not isinstance(v, list):
if isinstance(data[k], np.ndarray):
feed_dict[v] = data[k].astype(v.dtype.as_numpy_dtype)
else:
feed_dict[v] = data[k]
else:
for t, d in zip(v, data[k]):
feed_dict[t] = d.astype(t.dtype.as_numpy_dtype)
return feed_dict
def main(_):
def _readout(target):
max_axis = tf.reduce_max(target, [0, 1], keep_dims=True)
min_axis = tf.reduce_min(target, [0, 1], keep_dims=True)
image = (target - min_axis) / (max_axis - min_axis)
return image
tf.reset_default_graph()
env = environment.get_game_environment(FLAGS.maps,
multiproc=FLAGS.multiproc,
random_goal=FLAGS.random_goal,
random_spawn=FLAGS.random_spawn)
exp = expert.Expert()
net = CMAP()
estimate_images = [_readout(estimate[0, -1, :, :, 0])
for estimate in net.intermediate_tensors['estimate_map_list']]
value_images = [_readout(value[0, :, :, 0]) for value in tf.unstack(net.intermediate_tensors['value_map'], axis=1)]
step_history = tf.placeholder(tf.string, name='step_history')
step_history_op = tf.summary.text('game/step_history', step_history, collections=['game'])
global_step = slim.get_or_create_global_step()
update_global_step_op = tf.assign_add(global_step, 1)
optimizer = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate)
gradients = optimizer.compute_gradients(net.output_tensors['loss'])
gradient_names = [v.name for _, v in gradients]
gradient_summary_op = [tf.reduce_mean(tf.abs(g)) for g, _ in gradients]
train_op = optimizer.apply_gradients(gradients)
slim.learning.train(train_op=train_op,
logdir=FLAGS.logdir,
global_step=global_step,
train_step_fn=DAGGER_train_step,
train_step_kwargs=dict(env=env, exp=exp, net=net,
update_global_step_op=update_global_step_op,
step_history=step_history,
step_history_op=step_history_op,
gradient_names=gradient_names,
gradient_summary_op=gradient_summary_op,
estimate_maps=estimate_images,
value_maps=value_images),
number_of_steps=FLAGS.num_games,
save_interval_secs=300 if not FLAGS.debug else 60,
save_summaries_secs=300 if not FLAGS.debug else 60)
if __name__ == '__main__':
tf.app.run()