else:
print('Full softmax')
print()
embedding = Embedding(sourceEmbedDim, targetEmbedDim, corpus.sourceVoc.size(),
corpus.targetVoc.size())
encdec = EncDec(sourceEmbedDim,
targetEmbedDim,
hiddenDim,
corpus.targetVoc.size(),
useSmallSoftmax=useSmallSoftmax,
dropoutRate=dropoutRate,
numLayers=numLayers)
if useSmallSoftmax:
vocGen = VocGenerator(vocGenHiddenDim, corpus.targetVoc.size(),
corpus.sourceVoc.size())
vocGen.load_state_dict(torch.load(vocGenFile))
vocGen.cuda()
vocGen.eval()
encdec.softmaxLayer.weight.weight = embedding.targetEmbedding.weight
embedding.cuda()
encdec.cuda()
batchListTrain = utils.buildBatchList(len(corpus.trainData), batchSize)
batchListDev = utils.buildBatchList(len(corpus.devData), batchSize)
withoutWeightDecay = []
withWeightDecay = []
for name, param in list(embedding.named_parameters()) + list(
torch.cuda.manual_seed(seed)
corpus = Corpus(sourceTrainFile, sourceOrigTrainFile, targetTrainFile,
sourceDevFile, sourceOrigDevFile, targetDevFile,
minFreqSource, minFreqTarget, maxLen)
print('Source vocabulary size: '+str(corpus.sourceVoc.size()))
print('Target vocabulary size: '+str(corpus.targetVoc.size()))
print()
print('# of training sentences: '+str(len(corpus.trainData)))
print('# of develop sentnces: '+str(len(corpus.devData)))
print('Random seed: ', str(seed))
print('K = ', K)
print()
vocGen = VocGenerator(vocGenHiddenDim, corpus.targetVoc.size(), corpus.sourceVoc.size(), dropoutRate)
vocGen.to(device)
batchListTrain = utils.buildBatchList(len(corpus.trainData), batchSize)
batchListDev = utils.buildBatchList(len(corpus.devData), batchSize)
withoutWeightDecay = []
withWeightDecay = []
for name, param in vocGen.named_parameters():
if 'bias' in name or 'softmax' in name or 'Embedding' in name:
withoutWeightDecay += [param]
else:
withWeightDecay += [param]
optParams = [{'params': withWeightDecay, 'weight_decay': weightDecay},
{'params': withoutWeightDecay, 'weight_decay': 0.0}]
out_features = 12772
voc_features = 226658
seed = 1
gpuId = 0
torch.set_num_threads(1)
torch.manual_seed(seed)
random.seed(seed)
torch.cuda.set_device(gpuId)
torch.cuda.manual_seed(seed)
device = torch.device('cuda:' + str(gpuId))
cpu = torch.device('cpu')
vocGen = VocGenerator(vocGenHiddenDim, out_features, voc_features)
vocGen.load_state_dict(torch.load(vocGenFile))
vocGen.cuda()
vocGen.eval()
sourceTrainFile = '../../WikiCatSum/film_tok_min5_L7.5k/train.bow.src'
sourceOrigTrainFile = sourceTrainFile
targetTrainFile = '../../WikiCatSum/film_tok_min5_L7.5k/train.bow.tgt'
sourceTestFile = '../../WikiCatSum/film_tok_min5_L7.5k/test.bow.src'
sourceOrigTestFile = sourceTestFile
targetTestFile = '../../WikiCatSum/film_tok_min5_L7.5k/test.bow.tgt'
minFreqSource = 2
minFreqTarget = 2
maxLen = 10000
trainPickle = 'trainData.pt'
devPickle = 'devData.pt'
torch.manual_seed(seed)
random.seed(seed)
torch.cuda.set_device(gpuId)
torch.cuda.manual_seed(seed)
device = torch.device('cuda:'+str(gpuId))
cpu = torch.device('cpu')
corpus = Corpus(sourceTrainFile, sourceOrigTrainFile, targetTrainFile,
sourceTestFile, sourceOrigTestFile, targetTestFile,
minFreqSource, minFreqTarget, maxLen, trainPickle, devPickle)
batchSize = len(corpus.devData)
vocGen = VocGenerator(vocGenHiddenDim, 12772, 226658)
vocGen.load_state_dict(torch.load(vocGenFile))
vocGen.cuda()
vocGen.eval()
batchListDev = utils.buildBatchList(len(corpus.devData), batchSize)
for batch in batchListDev:
batchSize = batch[1]-batch[0]+1
batchInputSource, lengthsSource, batchInputTarget, batchTarget, lengthsTarget, tokenCount, batchData, maxTargetLen = corpus.processBatchInfoNMT(batch, train = False, volatile = True, device = device)
targetVocGen, inputVocGen = corpus.processBatchInfoVocGen(batchData, smoothing = False, volatile = True, device = device)
outputVocGen = vocGen(inputVocGen)
<<<<<<< HEAD
#tmp = F.sigmoid(outputVocGen.data).data + targetVocGen.data