-
Notifications
You must be signed in to change notification settings - Fork 0
/
fromTranscripts_bow.py
332 lines (294 loc) · 12.3 KB
/
fromTranscripts_bow.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
import numpy as np
import torch.nn as nn
import torch.optim as optim
import torch
import torch.nn.functional as F
import sys
import matplotlib
import matplotlib.pyplot as plt
plt.switch_backend('agg')
import pdb
import random
from random import shuffle
import argparse
import copy
parser = argparse.ArgumentParser(description='DA tagging')
parser.add_argument('--nEpoch', type=int, default=10,
help='number of epochs')
parser.add_argument('--rSeed', type=int, default=100,
help='random seed')
parser.add_argument('--nClasses', type=int, default=43,
help='number of classes')
parser.add_argument('--bSize', type=int, default=30,
help='batch size')
parser.add_argument('--nLayers', type=int, default=1,
help='number of layers')
parser.add_argument('--shuffle', type=bool, default=True,
help='shuffling of data before each epoch')
parser.add_argument('--init', type=bool, default=True,
help='modified initialization or the default one')
parser.add_argument('--lr', type=float, default=0.0015,
help='initial learning rate')
parser.add_argument('--optim', type=str, default='Adam',
help='type of optimizer (SGD, Adagrad, Adam, RMSprop)')
parser.add_argument('--initM', type=str, default='xavier_normal_',
help='type of optimizer (xavier_uniform_,xavier_normal_,uniform_, \
normal_,constant_,ones_,zeros_)')
parser.add_argument('--mode', type=str, default='train',
help='mode of operation (dataproc or train)')
parser.add_argument('--nl', type=str, default='relu',
help='non linearity (relu, sigmoid, tanh')
parser.add_argument('--hiddenSize', type=int, default=256,
help='number of hidden units')
parser.add_argument('--lrs', type=float, default=1,
help='learning rate scheduling')
# for debugging purpose
parser.add_argument('--overfit', type=bool, default=False,
help='running an overfitting experiment')
parser.add_argument('--debug', type=bool, default=False,
help='printing outputs')
args = parser.parse_args()
rSeed = args.rSeed
random.seed(rSeed)
torch.manual_seed(rSeed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(rSeed)
if args.mode=='dataproc':
def getData(dataSplit='train',vocab=None,tag2label=None,tag2freq=None,thresh=2000):
dataSize = 0
dataList = open('dataSplit/fromPaper/dataset_'+dataSplit+'.txt').readlines()
if(not vocab):
vocabDict = {}
for line in dataList:
index = [] # bag of words feature for this utterance
words = line.split('\t')[1].split(' ') # list of words
for word in words:
if(word not in vocabDict): vocabDict[word] = 0
vocabDict[word] += 1
# retaining the top 'thresh' frequent words in the vocabulary
# sort the dict according to values
vocabDictSorted = sorted(vocabDict,key=vocabDict.get,reverse=True)
# retain top n keys
vocab = vocabDictSorted[:thresh]
vocab.insert(0,'UNK')
outfile = open('vocab.txt','w')
for word in vocab:
outfile.write(word+'\n')
outfile.close()
if(not tag2label): tag2label = {}
labelId = 0
indices = [] # bag of words features
labels = []
for line in dataList:
index = [] # bag of words feature for this utterance
words = line.split('\t')[1].split(' ') # list of words
tag = line.split('\t')[2]
if(tag=='statement' or tag=='backchannel'):
# if(tag=='statement'): pass
# else:
if(dataSplit=='train' and tag not in tag2label):
tag2label[tag] = labelId
labelId += 1
# labelWeights.append(tag2freq[tag])
labels.append(tag2label[tag])
for word in words:
try: index.append(vocab.index(word))
except: index.append(0) # for 'UNK'
indices.append(index)
dataSize += 1
data = np.zeros([dataSize,len(vocab)])
dataTensor = [None]*dataSize
for i in range(dataSize):
data[i][indices[i]] = 1
dataTensor[i] = torch.tensor(data[i],dtype=torch.float)
if(dataSplit=='train'):
# outfile = open('labelWeights.txt','w')
# for weight in labelWeights:
# outfile.write(weight+'\n')
# outfile.close()
return dataTensor, labels, vocab, tag2label
else:
return dataTensor, labels
def getFreqDict():
freqFile = open('dataStats.csv','r')
tag2freq = {}
for line in freqFile:
tag = line.split(',')[0]
freq = line.split(',')[1]
tag2freq[tag] = freq
return tag2freq
print('Reading data')
tag2freq = getFreqDict()
trainDataTensor, trainLabels, vocab, tag2label = getData('train',tag2freq=tag2freq)
devDataTensor, devLabels = getData('dev',vocab,tag2label)
testDataTensor, testLabels = getData('test',vocab,tag2label)
print(tag2label)
print('Saving data')
torch.save(trainDataTensor, 'trainData.pt')
torch.save(trainLabels, 'trainLabels.pt')
torch.save(devDataTensor, 'devData.pt')
torch.save(devLabels, 'devLabels.pt')
torch.save(testDataTensor, 'testData.pt')
torch.save(testLabels, 'testLabels.pt')
elif args.mode=='train':
trainData = torch.load('trainData.pt')
# pdb.set_trace()
trainData = torch.stack(trainData)
trainLabels = torch.load('trainLabels.pt')
batchSize = args.bSize
if(args.overfit):
# debugging experimentation: overfits?
trainData = trainData[:100]
trainLabels = trainLabels[:100]
batchSize = 2
devData = torch.stack(torch.load('devData.pt'))
devLabels = torch.load('devLabels.pt')
testData = torch.stack(torch.load('testData.pt'))
testLabels = torch.load('testLabels.pt')
inSize = len(open('vocab.txt','rb').readlines())
nClasses = args.nClasses
nHidden = args.hiddenSize
nl = args.nl
initialization = args.init
# weights = torch.zeros(nClasses)
# weightFile = open('labelWeights.txt','r').readlines()
# for i in range(len(weightFile)):
# weights[i] = (1-float(weightFile[i].strip('\n')))/(nClasses-1)
class Net(nn.Module):
def __init__(self,nClasses,inputSize,hiddenSize,nLayers=1,nl='relu',init=False,initMethod='xavier_uniform_'):
super(Net, self).__init__()
# self.embed = nn.Embedding(inputSize,hiddenSize1)
self.nl = nl
self.nLayers = nLayers
self.affineIn = nn.Linear(inputSize,hiddenSize)
self.affineOut = nn.Linear(hiddenSize,nClasses)
self.hiddenLayers = nn.Sequential()
self.dropout = torch.nn.Dropout(p=0.25)
for i in range(nLayers):
self.hiddenLayers.add_module('layer'+str(i), nn.Linear(hiddenSize, hiddenSize))
if(init):
for m in self.modules():
if isinstance(m, nn.Linear):
# nn.init.xavier_normal_(m.weight, 1)
getattr(nn.init,initMethod)(m.weight,1)
nn.init.constant_(m.bias, 0)
def forward(self,x):
out = getattr(F, self.nl)(self.dropout(self.affineIn(x)))
out = getattr(F, self.nl)(self.hiddenLayers.forward(out))
out = getattr(F, self.nl)(self.affineOut(out))
return out
lrs = args.lrs
lr = args.lr
ff = Net(nClasses,inSize,nHidden,nLayers=args.nLayers,nl=nl,init=initialization,initMethod=args.initM).cuda()
# optimizer = getattr(optim, args.optim)(ff.parameters(),lr=lr)
# optimizer = optim.SGD(ff.parameters(),lr=0.1)
# optimizer = optim.Adagrad(ff.parameters(),lr=0.01)
criterion = torch.nn.CrossEntropyLoss(size_average=False).cuda()
nEpoch = args.nEpoch
iterNo = 0
shuffle = args.shuffle
trainLoss = []
devLoss = []
trainAcc = []
devAcc = []
weightNormDiff = []
def validationStep(ff,devAcc,devLoss):
totalDev = len(devData)
correctDev = 0
lossTotDev = 0
for i in range(totalDev):
input = devData[i].cuda()
output = ff(input)
lossTotDev += criterion(output.view(1,nClasses), torch.tensor([devLabels[i]]).cuda()).detach()
pred = np.argmax(output.cpu().detach().numpy())
if(pred==devLabels[i]): correctDev += 1
print("Val acc: ", correctDev/float(totalDev))
devAcc.append(correctDev/float(totalDev))
devLoss.append(lossTotDev/float(totalDev))
return devAcc, devLoss
ff.train()
# print(ff.affineIn.weight)
while(iterNo<nEpoch):
print("Epoch no: ", iterNo)
# print(ff.training)
# total = len(trainData)
# print(trainData[0:batchSize])
optimizer = getattr(optim, args.optim)(ff.parameters(),lr=lr)
lr = lr*lrs
if(shuffle):
temp = list(zip(trainData,trainLabels))
random.shuffle(temp)
[trainData1,trainLabels1] = zip(*temp)
trainData1 = torch.stack(trainData1)
else:
trainData1 = copy.deepcopy(trainData)
trainLabels1 = copy.deepcopy(trainLabels)
# print(trainData1[0:batchSize])
total = 0
correct = 0
lossTot = 0
nBatches = len(trainData1)//batchSize
startIdx = 0
for batchIdx in range(nBatches):
optimizer.zero_grad()
total += batchSize
start, end = batchIdx*batchSize, (batchIdx+1)*batchSize
data, target = trainData1[start:end].cuda(), torch.tensor(trainLabels1[start:end]).cuda()
# input = trainData[i]
output = ff(data)
output = output.view(batchSize,nClasses)
# loss = criterion(output.view(1,37), torch.tensor([trainLabels[i]]))
loss = criterion(output, target)
loss.backward()
optimizer.step()
if(batchIdx==0): prevStepWeight = ff.affineIn.weight.detach().cpu().numpy()
lossTot += loss.detach()
pred = torch.max(output, 1)[1]
correct += torch.sum(pred==target).cpu().numpy()
# pred = np.argmax(output.detach().cpu().numpy())
# if(pred==trainLabels[i]): correct += 1
if(((batchIdx+1)%(nBatches//4))==0):
# print(np.count_nonzero(list(ff.parameters())[0].grad.detach().cpu().numpy()))
print("Train acc: ", correct/float(total))
# print(ff.affineIn.weights.grad)
if(args.debug):
print(target)
print(pred)
weightNormDiff.append(np.sqrt(np.sum((ff.affineIn.weight.detach().cpu().numpy() - prevStepWeight)**2)))
prevStepWeight = ff.affineIn.weight.detach().cpu().numpy()
trainAcc.append(correct/float(total))
trainLoss.append(lossTot/float(total))
devAcc, devLoss = validationStep(ff,devAcc,devLoss)
iterNo += 1
ff.eval()
total = len(testData)
correct = 0
for i in range(total):
input = testData[i].cuda()
output = ff(input)
pred = np.argmax(output.cpu().detach().numpy())
if(pred==testLabels[i]): correct += 1
print("Test acc: ", correct/float(total))
xaxis = np.arange(1,len(trainLoss)+1)
fig, ax1 = plt.subplots()
ax1.plot(xaxis,trainLoss,'g',label='train loss')
ax1.plot(xaxis,devLoss,'r',label='dev loss')
ax1.set_xlabel('Number of quarter epochs')
ax1.set_ylabel('Negative log-likelihood', color='g')
plt.legend(loc=6)
ax2 = ax1.twinx()
ax2.plot(xaxis,trainAcc,'b',label='train Acc')
ax2.plot(xaxis,devAcc,'k',label='dev Acc')
ax2.set_ylabel('Accuracy', color='b')
plt.legend(loc=7)
fig.tight_layout()
plt.savefig('LossCompare.png')
plt.figure()
plt.plot(weightNormDiff)
plt.xlabel("number of quarter epochs")
plt.ylabel("norm difference in learned weights")
plt.savefig('weightNorm.png')
# plt.show()
# plt.close()
else:
print("Enter a valid argument")