-
Notifications
You must be signed in to change notification settings - Fork 0
/
util.py
434 lines (357 loc) · 14.1 KB
/
util.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
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
# Environment
import gym
#from gym.spaces import prng
# Utility
import cv2
import numpy as np
from collections import deque
import argparse
from functools import partial
import pickle
from sklearn.metrics import confusion_matrix
from sklearn.preprocessing import normalize
import matplotlib.pyplot as plt
import tensorflow as tf
from models import clipped_mse
class MaxAndSkipEnv(gym.Wrapper):
def __init__(self, env, skip=4):
"""Return only every `skip`-th frame"""
gym.Wrapper.__init__(self, env)
# most recent raw observations (for max pooling across time steps)
self._obs_buffer = np.zeros((2,) + env.observation_space.shape, dtype=np.uint8)
self._skip = skip
def step(self, action):
"""Repeat action, sum reward, and max over last observations."""
total_reward = 0.0
done = None
for i in range(self._skip):
obs, reward, done, info = self.env.step(action)
if i == self._skip - 2: self._obs_buffer[0] = obs
if i == self._skip - 1: self._obs_buffer[1] = obs
total_reward += reward
if done:
break
# Note that the observation on the done=True frame
# doesn't matter
max_frame = self._obs_buffer.max(axis=0)
return max_frame, total_reward, done, info
def reset(self, **kwargs):
return self.env.reset(**kwargs)
# Rescales the Image width by k and the height by l.
def repeat_upsample(rgb_array, k=1, l=1, err=[]):
# repeat kinda crashes if k/l are zero
if k <= 0 or l <= 0:
if not err:
print("Number of repeats must be larger than 0, k: {}, l: {}, returning default array!".format(k, l))
err.append('logged')
return rgb_array
# repeat the pixels k times along the y axis and l times along the x axis
# if the input image is of shape (m,n,3), the output image will be of shape (k*m, l*n, 3)
return np.repeat(np.repeat(rgb_array, k, axis=0), l, axis=1)
# Makes an environment for simulation
# Should be exact same every time now
def make_environment(game):
env = gym.make(game)
env = MaxAndSkipEnv(env, 2)
env.seed(0)
env.action_space.np_random.seed(0)
#prng.seed(0)
return env
# Makes a simple catalog for an episode ~ [observation,action,reward*,done*,info*]
# *Unless initial state
# Action Effects: 0 ~ None, 1 ~ Fire, 2 ~ Right, 3 ~ Left
def record_episode(env, num=1):
episodes = []
for i in range(num):
done = False
lives = 5
t_lives = 5
episode = []
episode.append([env.reset(), 0, done, None])
while not done:
if lives - t_lives is True:
action = 1
episode[-1].insert(1, action)
episode.append(list(env.step(action)))
done = episode[-1][2]
lives = t_lives
t_lives = episode[-1][-1]['ale.lives']
else:
action = env.action_space.sample()
episode[-1].insert(1, action)
episode.append(list(env.step(action)))
done = episode[-1][2]
t_lives = episode[-1][-1]['ale.lives']
episodes.append(episode)
env.close()
return episodes, env.action_space.n
# Record a human playing the game Breakout
# Same format as record_episode
def record_human(ep, eps, obs_before, obs_after, action, rew, done, info):
if done is False:
ep.append((obs_before, action, rew, done))
else:
eps.append(ep)
ep = []
return
# Save episodes into directory sorted by episode number
# Caution: Will overwrite without hesitation
def save_episodes(dir, eps, num=0):
for i in range(num, num + len(eps)):
pickle.dump(eps[i], open(dir + str(i) + '.dump', 'wb'))
return
# Load episodes from directory. You'll have to which ones you want.
def load_episodes(dir, nums):
eps = []
for i in nums:
eps.append(pickle.load(open(dir + str(i) + '.dump', 'rb')))
return eps
# Plays back the episode
def playback(frames):
for i in range(len(frames)):
cv2.imshow('frame', repeat_upsample(frames[i][:, :, ::-1], 3, 3))
if cv2.waitKey(30) & 0xFF == ord('q'):
break
# Prepares the data by splitting frames into observation and target
# Makes images black and white and rescales to be half of the original size
def forward_data(episodes, n_actions=4):
stacks = []
actions = []
targets = []
for j in range(len(episodes)):
frames, inputs, _, _ = zip(*episodes[j])
frames = list(frames)
for i in range(len(episodes[j])):
frames[i] = frames[i][::2, ::2]
inputs = np.array(inputs)
inputs[inputs == 1] = 0
stack = []
action = []
target = []
for i in range(len(episodes[j]) - 4):
stack.append(np.concatenate(np.array(frames[i:i + 2]), axis=2))
action.append(inputs[i + 1])
target.append(np.concatenate(np.array(frames[i + 2:i + 4]), axis=2))
stacks += stack
actions += action
targets += target
return np.array(stacks), np.array(actions), np.array(targets)
def modal_data(episodes, n_actions=4):
stacks = []
actions = []
targets = []
for j in range(len(episodes)):
frames, inputs, _, _ = zip(*episodes[j])
frames = list(frames)
for i in range(len(episodes[j])):
frames[i] = frames[i][::2, ::2]
inputs = np.array(inputs)
inputs[inputs == 1] = 0
stack = []
action = []
target = []
for i in range(len(episodes[j]) - 6):
stack.append(np.concatenate(np.array(frames[i:i + 4]), axis=2))
action.append(inputs[i + 3])
temp = np.copy(frames[i+4:i+6])
temp[0] = temp[0] - frames[i+3]
temp[1] = temp[1] - frames[i+4]
target.append(np.concatenate(np.array(temp),axis = 2))
stacks += stack
actions += action
targets += target
return np.array(stacks), np.array(actions), np.array(targets)
def inverse_data(episodes, n_actions=4):
stacks = []
actions = []
for j in range(len(episodes)):
frames, inputs, _, _ = zip(*episodes[j])
frames = list(frames)
for i in range(len(episodes[j])):
frames[i] = frames[i]
inputs = np.array(inputs)
#inputs[inputs == 1] = 0
stack = []
action = []
for i in range(len(episodes[j]) - 3):
stack.append(np.concatenate(np.array(frames[i:i + 4]), axis=2))
action.append(inputs[i + 1])
stacks += stack
actions += action
return np.array(stacks), np.array(actions)
def linear_vector_data(episodes, frame_num = 4, action_num = 2, use_images = False):
stacks = []
actions = []
for j in range(len(episodes)):
t_f = []
frames, inputs, _, _ = zip(*episodes[j])
for i in range(len(episodes[j])-frame_num + 1):
if use_images:
t_f.append(np.concatenate(np.array(frames[i:i + frame_num]), axis=2))
else:
t_f.append(np.concatenate(np.array(frames[i:i+frame_num])))
stacks += t_f[:-1]
actions += inputs[frame_num-1:-1]
return np.array(stacks), np.array(actions)
def inverse_vector_data(episodes, frame_num = 4, action_num = 2, use_images = False):
stacks = []
actions = []
for j in range(len(episodes)):
t_f = []
frames, inputs, _, _ = zip(*episodes[j])
for i in range(len(episodes[j]) - frame_num + 1):
if use_images:
t_f.append(np.concatenate(np.array(frames[i:i + frame_num]), axis=2))
else:
t_f.append(np.concatenate(np.array(frames[i:i+frame_num])))
stacks += t_f
#actions += inputs[frame_num-2:-1]
actions += inputs[0:-1 - frame_num + 2]
return np.array(stacks), np.array(actions)
def vector_modal_data(episodes, frame_num = 4, latent_action_num = 2, use_images = False):
stacks = []
targets = []
actions = []
for j in range(len(episodes)):
t_s = []
t_f = []
frames, inputs, _, _ = zip(*episodes[j])
for i in range(len(episodes[j]) - frame_num):
if use_images:
t_s.append(np.concatenate(np.array(frames[i:i + frame_num]), axis=2))
t_f.append(np.concatenate(np.array(frames[i + frame_num]), axis=2))
else:
t_s.append(np.concatenate(np.array(frames[i:i + frame_num])))
t_f.append(np.array(frames[i + frame_num]))
stacks += t_s
targets += t_f
actions += inputs[frame_num - 1:-1]
return np.array(stacks), np.array(actions), np.array(targets)
# Data integrity tool. Shows that observations and target are logically constructed
def validate_data(d, a, t):
for i in range(4):
cv2.imshow('frame', repeat_upsample(np.array(d[:, :, 3*i:3*i+3]), 3, 3))
if cv2.waitKey(3000) & 0xFF == ord('q'):
break
print(a)
for i in range(2):
cv2.imshow('frame', repeat_upsample(np.array(t[:, :, 3*i:3*i+3]), 3, 3))
if cv2.waitKey(3000) & 0xFF == ord('q'):
break
print('close')
print('done')
if cv2.waitKey(3000) & 0xFF == ord('q'):
return
def make_confusion_plot(i_model,f_model,c_model,random_episodes,human_episodes):
print("Computing i_model confusion")
fig, ax = plt.subplots(nrows = 2, ncols = 3, constrained_layout = True)
# Make Confusion Matrix
data,actions = inverse_data(random_episodes)
r_score = np.argmax(i_model.predict([(data - np.mean(data, axis=0)) / 255.0]), axis=1)
r_score[r_score == 1] = 0
r_conf = confusion_matrix(actions, r_score)
r_conf = normalize(r_conf,norm = 'l1')
print("Random Score is " + str(np.trace(r_conf) / np.sum(r_conf)))
print(r_conf)
data, actions = inverse_data(human_episodes)
h_score = np.argmax(i_model.predict([(data - np.mean(data, axis=0)) / 255.0]), axis=1)
h_conf = confusion_matrix(actions, h_score)
h_conf = normalize(h_conf, norm = 'l1')
print("Human Score is " + str(np.trace(h_conf) / np.sum(h_conf)))
print(h_conf)
im = ax[0,0].imshow(r_conf)
ax[0,0].set_xticks(np.arange(3))
ax[0,0].set_yticks(np.arange(3))
ax[0,0].set_xticklabels(['None', 'Right', 'Left'])
ax[0,0].set_yticklabels(['None', 'Right', 'Left'])
ax[0,0].set_title("Random Inverse")
ax[1, 0].imshow(h_conf)
ax[1, 0].set_xticks(np.arange(3))
ax[1, 0].set_yticks(np.arange(3))
ax[1, 0].set_xticklabels(['None', 'Right', 'Left'])
ax[1, 0].set_yticklabels(['None', 'Right', 'Left'])
ax[1, 0].set_title("Human Inverse")
print("Computing f_model confusion")
data, actions, targets = forward_data(random_episodes)
sess = tf.Session()
r_score = []
pick = [0, 0, 0, 0]
idx_actions = np.random.randint(0,len(actions), size = 100)
actions = actions[idx_actions]
for j in range(100):
frames = (data[j] - np.mean(data, axis=0)) / 255.0
frames = np.tile(frames, (4, 1, 1, 1))
pick = np.sum(clipped_mse(f_model.predict([frames, np.arange(0, 4)])[:,:,:,3:6], targets[j][:,:,3:6]).eval(session=sess), axis=(1, 2))
r_score.append(np.argmin(pick))
r_score = np.array(r_score)
r_score[r_score == 1] = 0
actions[actions == 1] = 0
r_conf = confusion_matrix(actions, r_score)
r_conf = normalize(r_conf,norm = 'l1')
print("Random Score is " + str(np.trace(r_conf) / np.sum(r_conf)))
print(r_conf)
data, actions, targets = forward_data(human_episodes)
sess = tf.Session()
h_score = []
pick = [0, 0, 0, 0]
idx_actions = np.random.randint(0, len(actions), size=100)
actions = actions[idx_actions]
for j in range(100):
frames = (data[j] - np.mean(data, axis=0)) / 255.0
frames = np.tile(frames, (4, 1, 1, 1))
pick = np.sum(clipped_mse(f_model.predict([frames, np.arange(0, 4)])[:,:,:,3:6], targets[j][:,:,3:6]).eval(session=sess),axis=(1, 2))
h_score.append(np.argmin(pick))
h_score = np.array(h_score)
h_score[h_score == 1] = 0
actions[actions == 1] = 0
h_conf = confusion_matrix(actions, h_score)
h_conf = normalize(h_conf, norm = 'l1')
print("Human Score is " + str(np.trace(h_conf) / np.sum(h_conf)))
print(h_conf)
ax[0, 1].imshow(r_conf)
ax[0, 1].set_xticks(np.arange(3))
ax[0, 1].set_yticks(np.arange(3))
ax[0, 1].set_xticklabels(['None', 'Right', 'Left'])
ax[0, 1].set_yticklabels(['None', 'Right', 'Left'])
ax[0, 1].set_title("Random Forward")
ax[1, 1].imshow(h_conf)
ax[1, 1].set_xticks(np.arange(3))
ax[1, 1].set_yticks(np.arange(3))
ax[1, 1].set_xticklabels(['None', 'Right', 'Left'])
ax[1, 1].set_yticklabels(['None', 'Right', 'Left'])
ax[1, 1].set_title("Human Forward")
print("Computing c_model confusion")
# Make Confusion Matrix
data, actions = inverse_data(random_episodes)
data = data[:-2]
actions = actions[2:]
r_score = np.argmax(c_model.predict([(data - np.mean(data, axis=0)) / 255.0]), axis=1)
r_score[r_score == 1] = 0
r_conf = confusion_matrix(actions, r_score)
r_conf = normalize(r_conf)
print("Random Score is " + str(np.trace(r_conf) / np.sum(r_conf)))
print(r_conf)
data, actions = inverse_data(human_episodes)
data = data[:-2]
actions = actions[2:]
h_score = np.argmax(c_model.predict([(data - np.mean(data, axis=0)) / 255.0]), axis=1)
h_score[h_score == 1] = 0
h_conf = confusion_matrix(actions, h_score)
h_conf = normalize(h_conf,norm = 'l1')
print("Human Score is " + str(np.trace(h_conf) / np.sum(h_conf)))
print(h_conf)
ax[0, 2].imshow(r_conf)
ax[0, 2].set_xticks(np.arange(3))
ax[0, 2].set_yticks(np.arange(3))
ax[0, 2].set_xticklabels(['None', 'Right', 'Left'])
ax[0, 2].set_yticklabels(['None', 'Right', 'Left'])
ax[0, 2].set_title("Random Clone")
ax[1, 2].imshow(h_conf)
ax[1, 2].set_xticks(np.arange(3))
ax[1, 2].set_yticks(np.arange(3))
ax[1, 2].set_xticklabels(['None', 'Right', 'Left'])
ax[1, 2].set_yticklabels(['None', 'Right', 'Left'])
ax[1, 2].set_title("Human Clone")
fig.colorbar(im, ax=list(ax.flatten()))
plt.show()
return