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
'''
# how to load NMT model parameters
all_state_dict = torch.load(nmtFile)
embedding_state_dict = {}
encdec_state_dict = {}
for s in all_state_dict:
if 'Embedding' in s:
embedding_state_dict[s] = all_state_dict[s]
else:
encdec_state_dict[s] = all_state_dict[s]
embedding.load_state_dict(embedding_state_dict)
encdec.load_state_dict(encdec_state_dict)
'''
embedding.cuda()
encdec.cuda()
stateDict[elem] = stateDict[elem].to(cpu)
torch.save(stateDict, './params/embedding.bin')
stateDict = encdec.state_dict()
for elem in stateDict:
stateDict[elem] = stateDict[elem].to(cpu)
torch.save(stateDict, './params/encdec.bin')
prevDevGleu = devGleu
if train:
exit(0)
torch.set_grad_enabled(False)
embedding.load_state_dict(torch.load('./params/embedding.bin'))
encdec.load_state_dict(torch.load('./params/encdec.bin'))
embedding.to(device)
encdec.to(device)
embedding.eval()
encdec.eval()
f_trans = open('./trans.txt', 'w')
f_gold = open('./gold.txt', 'w')
devPerp = 0.0
totalTokenCount = 0.0
for batch in batchListDev:
if os.path.exists(wordParamsFile):
embedding.wordEmbedding.load_state_dict(torch.load(wordParamsFile))
else:
utils.loadEmbeddings(embedding.wordEmbedding, corpus.voc,
wordEmbeddingFile)
torch.save(embedding.wordEmbedding.state_dict(), wordParamsFile)
if os.path.exists(charParamsFile):
embedding.charEmbedding.load_state_dict(torch.load(charParamsFile))
else:
utils.loadEmbeddings(embedding.charEmbedding, corpus.charVoc,
charEmbeddingFile)
torch.save(embedding.charEmbedding.state_dict(), charParamsFile)
if test:
tagger.load_state_dict(torch.load(taggerParamsFile))
embedding.load_state_dict(torch.load(embeddingParamsFile))
if useGpu:
if torch.cuda.is_available():
torch.cuda.set_device(args.gpuId)
torch.cuda.manual_seed(seed)
embedding.cuda()
tagger.cuda()
print('**** Running with GPU-' + str(args.gpuId) + ' ****\n')
else:
useGpu = False
print('**** Warning: GPU is not available ****\n')
criterionTagger = nn.CrossEntropyLoss(size_average=False, ignore_index=-1)
batchListTrain = utils.buildBatchList(len(corpus.trainData), batchSize)
vocGen.cuda()
vocGen.eval()
encdec.softmaxLayer.weight.weight = embedding.targetEmbedding.weight
# how to load NMT model parameters
if os.path.exists(nmtFile):
all_state_dict = torch.load(nmtFile)
embedding_state_dict = {}
encdec_state_dict = {}
for s in all_state_dict:
if 'Embedding' in s:
embedding_state_dict[s] = all_state_dict[s]
else:
encdec_state_dict[s] = all_state_dict[s]
embedding.load_state_dict(embedding_state_dict)
encdec.load_state_dict(encdec_state_dict)
else:
print('****** No pre-trained model found --> training from scratch ******')
embedding.cuda()
encdec.cuda()
bse.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(
encdec.named_parameters()):
}))
lstm = LSTM(
Config({
'input_size': word_embed.output_size + char_cnn.output_size,
'hidden_size': train_args['lstm_hidden_size'],
'forget_bias': 1.0,
'batch_first': True,
'bidirectional': True
}))
crf = CRF(Config({'label_vocab': label_vocab}))
output_linear = Linear(
Config({
'in_features': lstm.output_size,
'out_features': len(label_vocab)
}))
word_embed.load_state_dict(state['model']['word_embed'])
char_cnn.load_state_dict(state['model']['char_cnn'])
char_highway.load_state_dict(state['model']['char_highway'])
lstm.load_state_dict(state['model']['lstm'])
crf.load_state_dict(state['model']['crf'])
output_linear.load_state_dict(state['model']['output_linear'])
lstm_crf = LstmCrf(token_vocab=token_vocab,
label_vocab=label_vocab,
char_vocab=char_vocab,
word_embedding=word_embed,
char_embedding=char_cnn,
crf=crf,
lstm=lstm,
univ_fc_layer=output_linear,
embed_dropout_prob=train_args['embed_dropout'],
lstm_dropout_prob=train_args['lstm_dropout'],