-
Notifications
You must be signed in to change notification settings - Fork 3
/
small_env_lang_4x4.py
387 lines (307 loc) · 13 KB
/
small_env_lang_4x4.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
"""
Modified training enviornment for grounded meta-rl agent for the GSP task distribution used in Kumar et al. 2022. We modified it to be amenable to agents with grounding losses.
lang_group: human (human language), synth (synthetic language), auto (autoencoder), program_pen_library (program with library learning), program_pen_nolibrary (program without library learning)
rules: gsp_4x4 (human tasks), gsp_4x4_null (control tasks)
"""
import gym
from gym.utils import seeding
from PIL import Image as PILImage
from gym.spaces import Box
from gym.spaces import Discrete
from gym.spaces import Dict
import numpy as np
from itertools import product
from itertools import permutations
import pickle
class BattleshipEnv(gym.Env):
reward_range = (-float('inf'), float('inf'))
metadata = {'render.modes': ['human', 'rgb_array'],'video.frames_per_second' : 3}
def __init__(self,rules='chain',lang_group='human',n_board=4,hold_out=0,permute=0,pretrain=0,render_img=0):
self.viewer = None
self.seed()
action_converter=[]
for i in range(n_board):
for j in range(n_board):
action_converter.append((i,j))
self.action_converter=np.asarray(action_converter)
self.n_board=n_board
self.hold_out=hold_out
self.rules=rules
self.pretrain=pretrain
self.render_img=render_img
self.lang_group=lang_group
if lang_group=='auto':
if self.rules=='gsp_4x4' or self.rules=='gsp_4x4_null':
language_boards=np.load('data/500_gsp_samples.npy')
self.lang_dict={}
for i in range(len(language_boards)):
self.lang_dict[tuple(language_boards[i].flatten())]=np.asarray([language_boards[i].flatten().astype('float')])
else:
if self.rules=='gsp_4x4' or self.rules=='gsp_4x4_null':
if 'program' in lang_group:
program_file=np.load('data/500_gsp_samples-recognition_activations-structurePenalty2.npz')
if 'nolibrary' in lang_group:
key='activations_pen_noConsolidation'
else:
key='activations_pen_structurePenalty1.5'
language_embedings=program_file[key]
language_embedings=np.reshape(language_embedings,(language_embedings.shape[0],-1,language_embedings.shape[1]))
language_boards=np.load('data/500_gsp_samples-recognition_activations.npz')['tasks']
else:
if lang_group=='human':
lang_file='data/500_gsp_samples_text_human_encoded.npy'
if lang_group=='th':
lang_file='data/500_gsp_samples_text_synth_encoded.npy'
language_embedings=np.load(lang_file,allow_pickle=True)
language_boards=np.load('data/500_gsp_samples.npy')
self.lang_dict={}
for i in range(language_boards.shape[0]):
self.lang_dict[tuple(language_boards[i].flatten())]=language_embedings[i]
if hold_out==-1:
self.heldout=np.load('data/'+self.rules+'_sample.npy').reshape((-1,4,4))
self.maze_idx=0
self.maze=np.reshape(self.heldout[self.maze_idx],(4,4))
start=np.load('data/'+self.rules+'_sample_starts.npy')[self.maze_idx]
else:
if hold_out>0:
heldout=np.load('data/'+self.rules+'_sample.npy').reshape((-1,16))
self.heldout=set([tuple(x) for x in heldout])
full_distribution_init=np.load('data/'+self.rules+"_full.npy")
full_probs_init=np.load('data/'+self.rules+"_full_probs.npy")
self.full_distribution=np.asarray([b.flatten() for b in full_distribution_init if tuple(b.flatten()) in self.lang_dict.keys()])
self.full_probs=np.asarray([full_probs_init[i] for i in range(full_probs_init.shape[0]) if tuple(full_distribution_init[i].flatten()) in self.lang_dict.keys()])
if hold_out>0:
for board in heldout:
boarda=np.asarray(board)
for i in range(self.full_distribution.shape[0]):
if np.sum(self.full_distribution[i]==boarda)==16:
self.full_probs[i]=0.0
break
elif np.sum(self.full_distribution[i])==0:
self.full_probs[i]=0.0
for i in range(self.full_distribution.shape[0]):
if np.sum(self.full_distribution[i])==0:
self.full_probs[i]=0.0
self.full_probs=self.full_probs/np.sum(self.full_probs)
elif hold_out==0:
self.heldout=set()
if 'null' not in self.rules:
r_idx=np.random.choice(np.arange(self.full_distribution.shape[0]),p=self.full_probs)
grid=self.full_distribution[r_idx].reshape((4,4))
reds=np.vstack(np.where(grid==1)).T
start=reds[np.random.choice(np.arange(reds.shape[0]))]
gen=(grid,start)
else:
raise Exception("No training on control distribution.")
#network.load_state_dict(torch.load("data/"+self.rules+"_fc_generator.pt"))
#network.eval()
#self.size_buffer=1000
#self.gibbs_buffer=batch_gibbs(S=4,numSweeps=20,batch_size=self.size_buffer,network=network).reshape((-1,4,4))
#self.gibbs_idx=0
#grid=self.null_sample()
#reds=np.vstack(np.where(grid==1)).T
#start=reds[np.random.choice(np.arange(reds.shape[0]))]
#gen=(grid,start)
while tuple(gen[0].flatten()) in self.heldout:
if 'null' not in self.rules:
r_idx=np.random.choice(np.arange(self.full_distribution.shape[0]),p=self.full_probs)
grid=self.full_distribution[r_idx].reshape((4,4))
reds=np.vstack(np.where(grid==1)).T
start=reds[np.random.choice(np.arange(reds.shape[0]))]
gen=(grid,start)
else:
raise Exception("No training on control distribution")
#network.load_state_dict(torch.load("data/"+self.rules+"_fc_generator.pt"))
#network.eval()
#self.size_buffer=1000
#self.gibbs_buffer=batch_gibbs(S=4,numSweeps=20,batch_size=self.size_buffer,network=network).reshape((-1,4,4))
#self.gibbs_idx=0
#grid=self.null_sample()
#reds=np.vstack(np.where(grid==1)).T
#start=reds[np.random.choice(np.arange(reds.shape[0]))]
#gen=(grid,start)
self.maze=grid
self.board=np.ones(self.maze.shape)*-1
self.current_position=start
self.board[self.current_position[0],self.current_position[1]]=1
self.num_hits=0
self.self_hits={}
if self.render_img:
if lang_group=='auto':
self.observation_space = Box(low=-1, high=1, shape=(16+n_board*n_board*1+n_board*n_board+1,), dtype=np.float32)
elif 'program' in lang_group:
self.observation_space = Box(low=-1, high=1, shape=(64+n_board*n_board*1+n_board*n_board+1,), dtype=np.float32)
else:
self.observation_space = Box(low=-1, high=1, shape=(768+n_board*n_board*1+n_board*n_board+1,), dtype=np.float32)
else:
if lang_group=='auto':
self.observation_space = Box(low=-1, high=1, shape=(16+n_board*n_board*1+n_board*n_board+1,), dtype=np.float32)
elif 'program' in lang_group:
self.observation_space = Box(low=-1, high=1, shape=(64+n_board*n_board*1+n_board*n_board+1,), dtype=np.float32)
else:
self.observation_space = Box(low=-1, high=1, shape=(768+n_board*n_board*1+n_board*n_board+1,), dtype=np.float32)
self.action_space = Discrete(np.prod(self.maze.shape))
self.nA=n_board*n_board
self.prev_reward=0
self.prev_action=np.zeros((self.nA,))
self.valid_actions=[1 for _ in range(self.nA)]
"""
def null_sample(self):
if self.gibbs_idx>=self.size_buffer:
self.gibbs_buffer=batch_gibbs(S=4,numSweeps=20,batch_size=self.size_buffer,network=network).reshape((-1,4,4))
self.gibbs_idx=0
grid=self.gibbs_buffer[self.gibbs_idx]
self.gibbs_idx=self.gibbs_idx+1
while np.sum(grid)<3:
if self.gibbs_idx>=self.size_buffer:
self.gibbs_buffer=batch_gibbs(S=4,numSweeps=20,batch_size=self.size_buffer,network=network).reshape((-1,4,4))
self.gibbs_idx=0
grid=self.gibbs_buffer[self.gibbs_idx]
self.gibbs_idx=self.gibbs_idx+1
return grid
"""
def seed(self, seed=None):
self.np_random, seed = seeding.np_random(seed)
return [seed]
def step(self, action):
#print(self.tot_solves)
#print(action,self.valid_actions,self.valid_actions[action])
prev_position=self.current_position
self.current_position=self.action_converter[action]
reward=0
if self.board[self.current_position[0],self.current_position[1]]==-1:
if self.maze[self.current_position[0],self.current_position[1]]==1:
self.board[self.current_position[0],self.current_position[1]]=1
self.num_hits+=1
reward=1
else:
self.board[self.current_position[0],self.current_position[1]]=0
reward=-1
else:
reward=-4
if (self.current_position[0],self.current_position[1]) not in self.self_hits.keys():
self.self_hits[(self.current_position[0],self.current_position[1])]=1
else:
self.self_hits[(self.current_position[0],self.current_position[1])]+=1
if self._is_goal():
reward=+5
done = True
if self.hold_out==-1:
self.maze_idx+=1
else:
done = False
p_action=self.prev_action
p_reward=self.prev_reward
self.prev_action=np.zeros((self.nA,))
self.prev_action[action]=1
self.prev_reward=reward
if self.render_img:
obs=self.get_image().flatten()
else:
obs=self.board.flatten()
self.valid_actions=[0 for _ in range(self.nA)]
for i in range(self.nA):
pos=self.action_converter[i]
if self.board[pos[0],pos[1]]==0:
self.valid_actions[i]=1
if tuple(self.maze.flatten()) in self.lang_dict:
choices=self.lang_dict[tuple(self.maze.flatten())]
c_idx=np.random.choice(np.arange(choices.shape[0]))
lang_embedding=choices[c_idx]
else:
sz=self.lang_dict[list(self.lang_dict.keys())[0]][0].shape[0]
lang_embedding=(np.ones((sz,))*-1).astype('float')
obs_array=np.concatenate((lang_embedding,obs,p_action,[p_reward]))
return obs_array,reward,done,{'valid_actions':self.valid_actions}
def get_action_mask(self):
return self.valid_actions
def _is_goal(self):
return np.sum(self.board==1)==np.sum(self.maze==1)
def get_image(self):
"""
img=np.empty((28,28, 1), dtype=np.int8)
#fills=[[0,0,255],[255,0,0],[255,255,255]]
#fills=[0,128,255]
for r in range(self.board.shape[0]):
for c in range(self.board.shape[1]):
#fill=fills[self.board[r,c].astype('int')]
img[4*r:4*r+4,4*c:4*c+4]=self.board[r,c]
return img
"""
return np.reshape(self.board,(4,4,1))
def set_task(self,task):
self.maze = task
self.board=np.zeros(self.maze.shape)
self.current_position=[np.random.choice(range(self.maze.shape[0])),np.random.choice(self.maze.shape[1])]
self.num_hits=0
self.self_hits={}
return self.board.flatten()
def reset(self):
if self.hold_out==-1:
self.maze=np.reshape(self.heldout[self.maze_idx%len(self.heldout)],(4,4))
start=np.load('data/'+self.rules+'_sample_starts.npy')[self.maze_idx%len(self.heldout)]
else:
if 'null' not in self.rules:
r_idx=np.random.choice(np.arange(self.full_distribution.shape[0]),p=self.full_probs)
grid=self.full_distribution[r_idx].reshape((4,4))
reds=np.vstack(np.where(grid==1)).T
start=reds[np.random.choice(np.arange(reds.shape[0]))]
gen=(grid,start)
else:
grid=self.null_sample()
reds=np.vstack(np.where(grid==1)).T
start=reds[np.random.choice(np.arange(reds.shape[0]))]
gen=(grid,start)
while tuple(gen[0].flatten()) in self.heldout:
if 'null' not in self.rules:
r_idx=np.random.choice(np.arange(self.full_distribution.shape[0]),p=self.full_probs)
grid=self.full_distribution[r_idx].reshape((4,4))
reds=np.vstack(np.where(grid==1)).T
start=reds[np.random.choice(np.arange(reds.shape[0]))]
gen=(grid,start)
else:
grid=self.null_sample()
reds=np.vstack(np.where(grid==1)).T
start=reds[np.random.choice(np.arange(reds.shape[0]))]
gen=(grid,start)
self.maze=grid
self.board=np.ones(self.maze.shape)*-1
self.current_position=start
self.board[self.current_position[0],self.current_position[1]]=1
self.num_hits=0
self.self_hits={}
if self.render_img:
obs=self.get_image().flatten()
else:
obs=self.board.flatten()
if tuple(self.maze.flatten()) in self.lang_dict:
choices=self.lang_dict[tuple(self.maze.flatten())]
c_idx=np.random.choice(np.arange(choices.shape[0]))
lang_embedding=choices[c_idx]
else:
sz=self.lang_dict[list(self.lang_dict.keys())[0]][0].shape[0]
lang_embedding=(np.ones((sz,))*-1).astype('float')
obs_array=np.concatenate((lang_embedding,obs,self.prev_action,[self.prev_reward]))
self.valid_actions=[1 for _ in range(self.nA)]
return obs_array
def render(self, mode='human', max_width=500):
img = self.get_image()
img = np.asarray(img).astype(np.uint8)
img_height, img_width = img.shape[:2]
ratio = max_width/img_width
img = PILImage.fromarray(img).resize([int(ratio*img_width), int(ratio*img_height)])
img = np.asarray(img)
if mode == 'rgb_array':
return img
elif mode == 'human':
from gym.envs.classic_control.rendering import SimpleImageViewer
if self.viewer is None:
self.viewer = SimpleImageViewer()
self.viewer.imshow(img)
return self.viewer.isopen
def close(self):
if self.viewer is not None:
self.viewer.close()
self.viewer = None
def register_small_env(env_id,rules,lang_group='human',n_board=4,hold_out=0,permute=0,max_episode_steps=25,pretrain=0,render_img=0):
gym.envs.register(id=env_id, entry_point=BattleshipEnv, max_episode_steps=max_episode_steps,kwargs={'rules':rules,'lang_group':lang_group,'n_board':n_board,'hold_out':hold_out,'permute':permute,'pretrain':pretrain,'render_img':render_img})