-
Notifications
You must be signed in to change notification settings - Fork 17
/
noTraceBaseline.py
424 lines (344 loc) · 14.3 KB
/
noTraceBaseline.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
from DSL import *
from graphicsSearch import serializeLine
from dispatch import dispatch
import time
import random
import numpy as np
import cPickle as pickle
import torch
from torch import nn
from torch.nn import functional as F
from torch.autograd import Variable
import torch.optim as optimization
import torch.cuda as cuda
from torch.nn.utils.rnn import pack_padded_sequence
GPU = cuda.is_available()
def variable(x, volatile=False):
if isinstance(x,list): x = np.array(x)
if isinstance(x,(np.ndarray,np.generic)): x = torch.from_numpy(x)
if GPU: x = x.cuda()
return Variable(x, volatile=volatile)
LEXICON = ["START","END",
"circle",
"rectangle",
"line","arrow = True","arrow = False","solid = True","solid = False",
"for",
"reflect","x","y",
"}",
"if",
"i","j","k","None"] + map(str,range(-5,20))
symbolToIndex = dict(zip(LEXICON,range(len(LEXICON))))
@dispatch(Loop)
def serializeProgram(l):
return serializeLine(l) + ([] if l.boundary == None else ["if"] + serializeProgram(l.boundary) + ["}"]) + \
serializeProgram(l.body) + ["}"]
@dispatch(Reflection)
def serializeProgram(l):
return serializeLine(l) + serializeProgram(l.body) + ["}"]
@dispatch(Primitive)
def serializeProgram(l): return serializeLine(l)
@dispatch(Block)
def serializeProgram(l):
return [ c for x in l.items for c in serializeProgram(x) ]
def parseOutput(l):
def get(l):
n = l[0]
del l[0]
return n
def parseLinear(l):
b = int(get(l))
x = get(l)
m = int(get(l))
if x == 'None': x = None
return LinearExpression(m,x,b)
def parseBody(l):
items = []
while True:
if l == []: return Block(items)
if l[0] == "}":
get(l)
return Block(items)
items.append(parseAtomic(l))
def parseAtomic(l):
k = get(l)
if k == 'circle':
x = parseLinear(l)
y = parseLinear(l)
return Primitive(k,x,y)
if k == 'rectangle':
x1 = parseLinear(l)
y1 = parseLinear(l)
x2 = parseLinear(l)
y2 = parseLinear(l)
return Primitive(k,x1,y1,x2,y2)
if k == 'line':
x1 = parseLinear(l)
y1 = parseLinear(l)
x2 = parseLinear(l)
y2 = parseLinear(l)
a = get(l)
s = get(l)
return Primitive(k,x1,y1,x2,y2,
"arrow = True" == a,
"solid = True" == s)
if k == 'for':
v = get(l)
b = parseLinear(l)
if l[0] == "if":
get(l)
boundary = parseBody(l)
else: boundary = None
body = parseBody(l)
return Loop(v = v, bound = b, boundary = boundary, body = body)
if k == 'reflect':
a = get(l)
c = int(get(l))
body = parseBody(l)
return Reflection(body = body, axis = a, coordinate = c)
raise Exception('parsing line '+k)
return parseBody(l)
class CaptionEncoder(nn.Module):
def __init__(self):
super(CaptionEncoder, self).__init__()
(squareFilters,rectangularFilters,numberOfFilters,kernelSizes,poolSizes,poolStrides) = (20,2,[10],
[9,9],
[8,4],
[4,4])
self.squareFilters = nn.Conv2d(1, squareFilters, kernelSizes[0], padding = kernelSizes[0]/2)
self.verticalFilters = nn.Conv2d(1, rectangularFilters,
(kernelSizes[0]/2 - 1,kernelSizes[0]*2 - 1),
padding = (kernelSizes[0]/4 - 1,kernelSizes[0] - 1))
self.horizontalFilters = nn.Conv2d(1, rectangularFilters,
(kernelSizes[0]*2 - 1,kernelSizes[0]/2 - 1),
padding = (kernelSizes[0] - 1,kernelSizes[0]/4 - 1))
self.laterStages = nn.Sequential(nn.ReLU(),
nn.MaxPool2d(poolSizes[0],poolStrides[0],padding = poolSizes[0]/2 - 1),
nn.Conv2d(squareFilters + 2*rectangularFilters,
numberOfFilters[0],
kernelSizes[1],
padding = kernelSizes[1]/2),
nn.ReLU(),
nn.MaxPool2d(poolSizes[1],poolStrides[1],padding = poolSizes[1]/2 - 1))
def forward(self,x):
c1 = self.squareFilters(x)
c2 = self.verticalFilters(x)
c3 = self.horizontalFilters(x)
c0 = torch.cat((c1,c2,c3),dim = 1)
output = self.laterStages(c0)
return output
#return output.view(output.size(0),-1)
class CaptionDecoder(nn.Module):
def __init__(self):
super(CaptionDecoder, self).__init__()
IMAGEFEATURESIZE = 2560
EMBEDDINGSIZE = 64
INPUTSIZE = IMAGEFEATURESIZE + EMBEDDINGSIZE
HIDDEN = 1024
LAYERS = 2
# self.embedding : list of N indices (BxW) -> (B,W,EMBEDDINGSIZE)
self.embedding = nn.Embedding(len(LEXICON),EMBEDDINGSIZE)
# The embedding is combined with the image features at each time step
self.rnn = nn.LSTM(INPUTSIZE, HIDDEN, LAYERS, batch_first = True)
self.tokenPrediction = nn.Linear(HIDDEN,len(LEXICON))
def forward(self, features, captions, lengths):
# flatten the convolution output
features = features.view(features.size(0),-1)
e = self.embedding(captions) # e: BxLx embeddingSize
#print "e = ",e.size()
#expandedFeatures: BxTx2560
expandedFeatures = features.unsqueeze(1).expand(features.size(0),e.size(1),features.size(1))
#recurrentInputs: bxtxINPUTSIZE
recurrentInputs = torch.cat((expandedFeatures,e),2)
#print "recurrentInputs = ",recurrentInputs.size()
packed = pack_padded_sequence(recurrentInputs, lengths, batch_first = True)
hidden,_ = self.rnn(packed)
outputs = self.tokenPrediction(hidden[0])
#print "outputs = ",outputs.size()
return outputs
def sample(self, features):
result = ["START"]
# (1,1,F)
features = features.view(-1).unsqueeze(0).unsqueeze(0)
#features: 1x1x2560
states = None
while True:
e = self.embedding(variable([symbolToIndex[result[-1]]]).view((1,-1)))
recurrentInput = torch.cat((features,e),2)
output, states = self.rnn(recurrentInput,states)
distribution = self.tokenPrediction(output).view(-1)
distribution = F.log_softmax(distribution).data.exp()
draw = torch.multinomial(distribution,1)[0]
c = LEXICON[draw]
if len(result) > 20 or c == "END":
return result[1:]
else:
result.append(c)
def buildCaptions(self,tokens):
'''returns inputs, sizes, targets'''
#tokens = [ [self.symbolToIndex["START"]] + [ self.symbolToIndex[s] for s in serializeProgram(p) ] + [self.symbolToIndex["END"]]
# for p in programs ]
# The full token sequences are START, ..., END
# Training input sequences are START, ...
# Target output sequences are ..., END
# the sizes are actually one smaller therefore
# Make sure that the token sequences are decreasing in size
previousLength = None
for t in tokens:
assert previousLength == None or len(t) <= previousLength
previousLength = len(t)
sizes = map(lambda t: len(t) - 1,tokens)
maximumSize = max(sizes)
tokens = [ np.concatenate((p, np.zeros(maximumSize + 1 - len(p),dtype = np.int)))
for p in tokens ]
tokens = np.array(tokens)
return variable(tokens[:,:-1]),sizes,variable(tokens[:,1:])
class NoTrace(nn.Module):
def __init__(self):
super(NoTrace, self).__init__()
self.encoder = CaptionEncoder()
self.decoder = CaptionDecoder()
def sampleMany(self, sequence, duration):
image = variable(np.array([ sequence.draw() ], dtype = np.float32), volatile = True).unsqueeze(1)
startTime = time()
imageFeatures = self.encoder(image)
#imageFeatures: 1x10x16x16
programs = []
while time() < startTime + duration:
nextSequence = self.decoder.sample(imageFeatures)
try:
p = parseOutput(nextSequence)
print "Sampled",p
programs.append({"time": time() - startTime,
"program": p,
"spec": p.convertToSequence()})
except: continue
return programs
def loss(self,examples):
# IMPORTANT: Sort the examples by their size. recurrent network stuff needs this
examples.sort(key = lambda e: len(e.tokens), reverse = True)
x = variable(np.array([ e.sequence.draw() for e in examples], dtype = np.float32))
x = x.unsqueeze(1) # insert the channel
imageFeatures = self.encoder(x)
inputs, sizes, T = self.decoder.buildCaptions([ e.tokens for e in examples ])
outputDistributions = self.decoder(imageFeatures, inputs, sizes)
T = pack_padded_sequence(T, sizes, batch_first = True)[0]
return F.cross_entropy(outputDistributions, T)
def load(self,path):
if os.path.isfile(path):
if not GPU: stuff = torch.load(path,map_location = lambda s,l: s)
else: stuff = torch.load(path)
self.load_state_dict(stuff)
print "Loaded checkpoint",path
else:
print "Could not find checkpoint",path
def dump(self,path):
torch.save(self.state_dict(),path)
print "Dumped checkpoint",path
class TrainingExample():
def __init__(self,p):
try:
self.tokens = np.array([symbolToIndex["START"]] + [ symbolToIndex[s] for s in serializeProgram(p) ] + [symbolToIndex["END"]])
except KeyError:
print "Key error in tokenization",serializeProgram(p)
assert False
self.sequence = p.convertToSequence()
#self.program = p
if str(parseOutput(serializeProgram(p))) != str(p):
print "Serialization failure for program",p
print serializeProgram(p)
print parseOutput(serializeProgram(p))
assert False
def loadTrainingData(n):
print "About to load the examples"
alternatives = ['/scratch/ellisk/randomlyGeneratedPrograms.p',
'randomlyGeneratedPrograms.p']
for alternative in alternatives:
if os.path.exists(alternative):
trainingDataPath = alternative
print "Loading training data from",trainingDataPath
break
with open(trainingDataPath,'rb') as handle:
X = pickle.load(handle)
print "Keeping %d/%d examples"%(n,len(X))
pruned = []
for x in X:
x = pickle.loads(x)
if x.items != []:
pruned.append(TrainingExample(x))
if len(pruned) >= n:
break
print "Pruned down to %d examples"%(len(pruned))
return pruned
if __name__ == "__main__":
import sys
model = NoTrace()
if GPU:
print "Using the GPU"
model = model.float().cuda()
else:
print "Using the CPU"
model = model.float()
model.load("checkpoints/noTrace.torch")
if 'test' in sys.argv:
from groundTruthParses import *
import os
target = getGroundTruthParse('drawings/expert-%s.png'%(sys.argv[2]))
results = model.sampleMany(target, 60*60)
results.sort(key = lambda z: (z["spec"] - target, z["program"].totalCost(), z["time"]))
if len(results) > 0:
z = results[0]
print "Best program:"
print z["program"].pretty()
#showImage(np.concatenate((1 - target.draw(),s.draw()),axis = 1))
saveMatrixAsImage(255*np.concatenate((1 - target.draw(),z["spec"].draw()),axis = 0),
"noTraceOutputs/%s.png"%(sys.argv[2]))
with open("noTraceOutputs/%s.p"%(sys.argv[2]),'wb') as handle:
pickle.dump(z,handle)
os.exit(0)
if 'statistics' in sys.argv:
from groundTruthParses import *
import os
results = []
gt = []
for n in xrange(100):
gt = gt + [getGroundTruthParse('drawings/expert-%s.png'%(n))]
f = "noTraceOutputs/%d.p"%n
try:
with open(f,"rb") as handle:
result = pickle.load(handle)
results.append(result)
except:
results.append(None)
times = [ r["time"] if r else float('inf')
for r in results ]
medianTime = sorted(times)[len(times)/2]
print "Median time",medianTime
successful = sum(r is not None and (r["spec"] - g) == 0
for r,g in zip(results,gt) )
print "# times that we got a program which was consistent with the data",successful
os.exit(0)
#print "# Learnable parameters:",sum([ parameter.view(-1).shape[0] for parameter in model.parameters() ])
N = 1*(10**7)
B = 64
X = loadTrainingData(N)
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
E = 0
while True:
E += 1
print "epic",E
# scrambled the data
X = list(np.random.permutation(X))
start = 0
batchesPerLoop = N/B
batchIndex = 0
while start < N:
batch = X[start:start + B]
model.zero_grad()
L = model.loss(batch)
if batchIndex%50 == 0:
print "Batch [%d/%d], LOSS = %s"%(batchIndex,batchesPerLoop,L.data[0])
model.dump("checkpoints/noTrace.torch")
L.backward()
optimizer.step()
start += B
batchIndex += 1