-
Notifications
You must be signed in to change notification settings - Fork 0
/
guide.py
362 lines (321 loc) · 16.7 KB
/
guide.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import pyro
import pyro.infer
import pyro.distributions as dist
import pyro.infer.csis.proposal_dists as proposal_dists
import numpy as np
from attention import FourierLocationEmbedder,\
LearnedLocationEmbedder,\
MultiHeadAttention
from generic_nn import Administrator
from graphics import AttentionTracker
from cnn import ViewEmbedder, FullViewEmbedder
class Guide(nn.Module):
def __init__(self,
d_k=64,
d_emb=128,
n_queries=4,
hidden_size=1024,
lstm_layers=1,
smp_emb_dim=32,
n_attention_heads=4,
lstm_dropout=0.1,
use_low_res_view=True,
low_res_view_as_attention_loc=False,
low_res_emb_size=128,
cuda=False,
share_smp_embedder=False,
share_qry_layer=False,
share_prop_layer=False,
keys_use_view=True,
keys_use_loc=True,
vals_use_view=True,
vals_use_loc=True,
learn_loc_embs=False,
wiggle_picture=False,
max_loc_emb_wiggle=0,
add_linear_loc_emb=True,
random_colour=True,
random_bar_width=True,
random_line_colour=True,
random_line_width=True,
scale="fixed",
attention_graphics_path=None,
collect_history=False):
super(Guide, self).__init__()
self.HYPERPARAMS = {"d_k": d_k,
"d_emb": d_emb, # cannot be changed without changing FourierLocationEmbedder
"n_queries": n_queries,
"hidden_size": hidden_size,
"lstm_layers": lstm_layers,
"smp_emb_dim": smp_emb_dim,
"n_attention_heads": n_attention_heads,
"lstm_dropout": lstm_dropout,
"use_low_res_view": use_low_res_view,
"low_res_emb_size": low_res_emb_size,
"low_res_view_as_attention_loc": low_res_view_as_attention_loc,
"CUDA": cuda,
"share_smp_embedder": share_smp_embedder,
"share_qry_layer": share_qry_layer,
"share_prop_layer": share_prop_layer,
"keys_use_view": keys_use_view,
"keys_use_loc": keys_use_loc,
"vals_use_view": vals_use_view,
"vals_use_loc": vals_use_loc,
"max_loc_emb_wiggle": max_loc_emb_wiggle}
self.CUDA = cuda
self.random_colour = random_colour
self.random_bar_width = random_bar_width
self.random_line_colour = random_line_colour
self.random_line_width = random_line_width
self.wiggle_picture = wiggle_picture
self.scale = scale
self.sample_statements = {"num_bars": {"instances": 1,
"dist": dist.categorical,
"output_dim": 5},
"bar_height": {"instances": 5,
"dist": dist.uniform}}
if wiggle_picture:
shifts = {shift: {"instances": 1,
"dist": dist.categorical,
"output_dim": 10}
for shift in ("x_shift", "y_shift")}
self.sample_statements.update(shifts)
if scale == "discrete":
self.sample_statements.update({"max_height": {"instances": 1,
"dist": dist.categorical,
"output_dim": 3}})
elif scale == "continuous":
self.sample_statements.update({"max_height": {"instances": 1,
"dist": dist.uniform}})
else:
assert scale == "fixed", "scale argument given is invalid"
if random_colour:
colour_samples = {col: {"instances": 1,
"dist": dist.uniform}
for col in ("red", "green", "blue")}
self.sample_statements.update(colour_samples)
if random_bar_width:
self.sample_statements.update({"bar_width": {"instances": 1,
"dist": dist.uniform}})
if random_line_colour:
colour_samples = {"line_{}".format(col): {"instances": 1,
"dist": dist.uniform}
for col in ("red", "green", "blue")}
self.sample_statements.update(colour_samples)
if random_line_width:
self.sample_statements.update({"line_width": {"instances": 1,
"dist": dist.uniform}})
self.administrator = Administrator(self.sample_statements,
self.HYPERPARAMS)
self.view_embedder = ViewEmbedder(output_dim=d_emb)
if use_low_res_view:
self.low_res_embedder = FullViewEmbedder(output_dim=low_res_emb_size)
if learn_loc_embs:
self.location_embedder = LearnedLocationEmbedder(d_emb)
else:
self.location_embedder = FourierLocationEmbedder(d_emb, 200, 200, add_linear_loc_emb)
self.mha = MultiHeadAttention(h=n_attention_heads, d_k=d_k, d_v=d_emb, d_model=d_emb)
self.initial_hidden = nn.Parameter(torch.normal(torch.zeros(lstm_layers, 1, hidden_size), 1))
self.initial_cell = nn.Parameter(torch.normal(torch.zeros(lstm_layers, 1, hidden_size), 1))
lstm_input_size = n_queries*d_emb + self.administrator.t_dim
self.lstm = nn.LSTM(input_size=lstm_input_size,
hidden_size=hidden_size,
num_layers=lstm_layers,
dropout=lstm_dropout)
if attention_graphics_path is not None:
self.attention_tracker = AttentionTracker(attention_graphics_path)
else:
self.attention_tracker = None
self.collect_history = collect_history
if collect_history is True:
self.history = []
if cuda:
self.cuda()
def init_lstm(self, keys, values, low_res_emb=None):
"""
run at the start of each trace
intialises LSTM hidden state etc.
"""
self.hidden, self.cell = self.initial_hidden, self.initial_cell
self.instances_dict = {key: -1 for key in self.sample_statements} # maybe messy to start from -1
self.keys = keys
self.values = values
self.low_res_emb = low_res_emb
self.prev_sample_name = None
self.prev_instance = None
def time_step(self, current_sample_name, prev_sample_value):
"""
perform one LSTM time step
returns proposal parameters for `current_sample_name`
"""
self.instances_dict[current_sample_name] += 1
current_instance = self.instances_dict[current_sample_name]
t = self.administrator.t(current_instance,
current_sample_name,
self.prev_instance,
self.prev_sample_name,
prev_sample_value,
self.low_res_emb)
queries = self.administrator.get_query_layer(current_sample_name, current_instance)(t=t,
prev_hidden=self.hidden)
if self.attention_tracker is None:
attention_output = self.mha(queries, self.keys, self.values).view(1, -1)
else:
attention_output = self.mha(queries, self.keys, self.values, self.attention_tracker).view(1, -1)
lstm_input = torch.cat([attention_output, t], 1).view(1, 1, -1)
lstm_output, (hidden, cell) = self.lstm(lstm_input, (self.hidden, self.cell))
del self.hidden
del self.cell
self.hidden, self.cell = hidden, cell
proposal_params = self.administrator.get_proposal_layer(current_sample_name, current_instance)(lstm_output)
self.prev_sample_name = current_sample_name
self.prev_instance = current_instance
if self.CUDA:
try:
proposal_params = proposal_params.cpu()
except AttributeError:
proposal_params = tuple(param.cpu() for param in proposal_params)
if self.collect_history:
self.history[-1]["{}_{}".format(current_sample_name, current_instance)] = proposal_params
return proposal_params
def forward(self, observed_image=None):
x = observed_image.view(1, 3, 210, 210)
if self.CUDA:
x = x.cuda()
if self.collect_history:
self.history.append({})
if self.HYPERPARAMS["use_low_res_view"]:
low_res_img = nn.AvgPool2d(10)(x)
low_res_emb = self.low_res_embedder(low_res_img)
""" this bit should probably be moved """
# find and embed each seperate location
views = [x[:, :, 10*j:10*(j+2), 10*i:10*(i+2)].clone().view(1, 3, 20, 20) for i in range(20) for j in range(20)]
views = torch.cat(views, 0)
view_embeddings = self.view_embedder(views)
# add location embeddings
x_offset = np.random.uniform(0, self.HYPERPARAMS["max_loc_emb_wiggle"]) # will be 0 by default
y_offset = np.random.uniform(0, self.HYPERPARAMS["max_loc_emb_wiggle"]) # will be 0 by default
location_embeddings = [self.location_embedder(i, j, x_offset=x_offset, y_offset=y_offset)
for i in range(0, 200, 10)
for j in range(0, 200, 10)]
if isinstance(x.data, torch.cuda.FloatTensor):
location_embeddings = [emb.cuda() for emb in location_embeddings]
location_embeddings = torch.cat(location_embeddings, 0)
keys = Variable(torch.zeros(view_embeddings.shape))
values = Variable(torch.zeros(view_embeddings.shape))
if self.CUDA:
keys = keys.cuda()
values = values.cuda()
if self.HYPERPARAMS["keys_use_view"]:
keys += view_embeddings
if self.HYPERPARAMS["keys_use_loc"]:
keys += location_embeddings
if self.HYPERPARAMS["vals_use_view"]:
values += view_embeddings
if self.HYPERPARAMS["vals_use_loc"]:
values += location_embeddings
if self.HYPERPARAMS["use_low_res_view"] and self.HYPERPARAMS["low_res_view_as_attention_loc"]:
keys = torch.cat([keys, low_res_emb.view(1, -1)], 0)
values = torch.cat([values, low_res_emb.view(1, -1)], 0)
# low_res_img = nn.AvgPool2d(10)(observed_image.view(1, 3, 200, 200))
# full_pic_embedding = self.big_picture_embedder(low_res_img)
# x = torch.cat([view_embeddings, full_pic_embedding], 0)
""""""
if self.HYPERPARAMS["use_low_res_view"] and not self.HYPERPARAMS["low_res_view_as_attention_loc"]:
self.init_lstm(keys, values, low_res_emb)
else:
self.init_lstm(keys, values, None)
prev_sample_value = None
if self.wiggle_picture:
for shift in ("x_shift", "y_shift"):
ps = self.time_step(shift, prev_sample_value)
prev_sample_value = pyro.sample(shift,
proposal_dists.categorical_proposal,
ps=ps).type(torch.FloatTensor)
max_max_height = 100
if self.scale == "discrete":
ps = self.time_step("max_height", prev_sample_value)
prev_sample_value = pyro.sample("max_height",
proposal_dists.categorical_proposal,
ps=ps).type(torch.FloatTensor)
max_height = 100
elif self.scale == "continuous":
mode, certainty = self.time_step("max_height", prev_sample_value)
prev_sample_value = pyro.sample("max_height",
proposal_dists.uniform_proposal,
Variable(torch.Tensor([0])),
Variable(torch.Tensor([max_max_height])),
mode*max_max_height,
certainty)
max_height = 100
else:
max_height = 10
if self.random_colour:
for colour in ("red", "green", "blue"):
mode, certainty = self.time_step(colour,
prev_sample_value)
# mode, certainty = modes[0], certainties[0]
prev_sample_value = pyro.sample(colour,
proposal_dists.uniform_proposal,
Variable(torch.Tensor([0])),
Variable(torch.Tensor([1])),
mode,
certainty)
if self.random_bar_width:
mode, certainty = self.time_step("bar_width",
prev_sample_value)
# mode, certainty = modes[0], certainties[0]
prev_sample_value = pyro.sample("bar_width",
proposal_dists.uniform_proposal,
Variable(torch.Tensor([0])),
Variable(torch.Tensor([1])),
mode,
certainty)
if self.random_line_colour:
for colour in ("red", "green", "blue"):
mode, certainty = self.time_step("line_{}".format(colour),
prev_sample_value)
# mode, certainty = modes[0], certainties[0]
prev_sample_value = pyro.sample("line_{}".format(colour),
proposal_dists.uniform_proposal,
Variable(torch.Tensor([0])),
Variable(torch.Tensor([1])),
mode,
certainty)
if self.random_line_width:
mode, certainty = self.time_step("line_width",
prev_sample_value)
# mode, certainty = modes[0], certainties[0]
prev_sample_value = pyro.sample("line_width",
proposal_dists.uniform_proposal,
Variable(torch.Tensor([0])),
Variable(torch.Tensor([2.5])),
mode*2.5,
certainty)
ps = self.time_step("num_bars",
prev_sample_value)
num_bars = pyro.sample("num_bars",
proposal_dists.categorical_proposal,
ps=ps)
prev_sample_value = num_bars.type(torch.FloatTensor)
for _ in range(num_bars):
mode, certainty = self.time_step("bar_height",
prev_sample_value)
# mode, certainty = modes[0], certainties[0]
print(mode.data.numpy()[0])
prev_sample_value = pyro.sample("{}_{}".format("bar_height", self.instances_dict["bar_height"]),
proposal_dists.uniform_proposal,
Variable(torch.Tensor([0])),
Variable(torch.Tensor([max_height])),
mode*max_height,
certainty)
if self.attention_tracker is not None:
self.attention_tracker.save_graphics()
def get_history(self):
if not self.collect_history:
raise Exception("collect_history is set to False")
return self.history