def getBleuOnVal(self, params, reverse_vocab, val_feed, sess, model_name): val_encoder_inputs, val_decoder_inputs, val_decoder_outputs, val_decoder_outputs_matching_inputs = val_feed decoder_outputs_inference, decoder_ground_truth_outputs = self.solveAll(params, val_encoder_inputs, val_decoder_outputs, reverse_vocab, sess=sess, print_progress=False) validOutFile_name = "./tmp/tmp_" + model_name +".valid.output" original_data_path = data_dir + "valid.original.nltktok" BLEUOutputFile_path = "./tmp/tmp_" + model_name + ".valid.BLEU" utilities.getBlue(validOutFile_name, original_data_path, BLEUOutputFile_path, decoder_outputs_inference, decoder_ground_truth_outputs, params['preprocessing']) return open(BLEUOutputFile_path,"r").read()
def main():
# params
params = {}
params['embeddings_dim'] = config.embeddings_dim
params['lstm_cell_size'] = config.lstm_cell_size
params['max_input_seq_length'] = config.max_input_seq_length
# inputs are all but last element,
# outputs are all but first element
params['max_output_seq_length'] = config.max_output_seq_length - 1
params['batch_size'] = config.batch_size
params['pretrained_embeddings'] = config.use_pretrained_embeddings
params[
'share_encoder_decoder_embeddings'] = config.share_encoder_decoder_embeddings
params['use_pointer'] = config.use_pointer
params['pretrained_embeddings_path'] = config.pretrained_embeddings_path
params[
'pretrained_embeddings_are_trainable'] = config.pretrained_embeddings_are_trainable
params[
'use_additional_info_from_pretrained_embeddings'] = config.use_additional_info_from_pretrained_embeddings
params['max_vocab_size'] = config.max_vocab_size
params['do_vocab_pruning'] = config.do_vocab_pruning
params['use_reverse_encoder'] = config.use_reverse_encoder
params['use_sentinel_loss'] = config.use_sentinel_loss
params['lambd'] = config.lambd
params['use_context_for_out'] = config.use_context_for_out
print "PARAMS:"
for key, value in params.items():
print " -- ", key, " = ", value
buckets = {
0: {
'max_input_seq_length': params['max_input_seq_length'],
'max_output_seq_length': params['max_output_seq_length']
}
}
# Preprocess data
preprocessing, val_encoder_inputs, val_decoder_outputs = preprocess(
"style_transfer_test.in")
params['vocab_size'] = preprocessing.vocab_size
# Pretrained embeddibngs
if params['pretrained_embeddings']:
pretrained_embeddings = pickle.load(
open(params['pretrained_embeddings_path'], "r"))
word_to_idx = preprocessing.word_to_idx
encoder_embedding_matrix = np.random.rand(params['vocab_size'],
params['embeddings_dim'])
decoder_embedding_matrix = np.random.rand(params['vocab_size'],
params['embeddings_dim'])
not_found_count = 0
for token, idx in word_to_idx.items():
if token in pretrained_embeddings:
encoder_embedding_matrix[idx] = pretrained_embeddings[token]
decoder_embedding_matrix[idx] = pretrained_embeddings[token]
else:
if not_found_count < 10:
print "No pretrained embedding for (only first 10 such cases will be printed. other prints are suppressed) ", token
not_found_count += 1
print "not found count = ", not_found_count
params['encoder_embeddings_matrix'] = encoder_embedding_matrix
params['decoder_embeddings_matrix'] = decoder_embedding_matrix
if params['use_additional_info_from_pretrained_embeddings']:
additional_count = 0
tmp = []
for token in pretrained_embeddings:
if token not in preprocessing.word_to_idx:
preprocessing.word_to_idx[
token] = preprocessing.word_to_idx_ctr
preprocessing.idx_to_word[
preprocessing.word_to_idx_ctr] = token
preprocessing.word_to_idx_ctr += 1
tmp.append(pretrained_embeddings[token])
additional_count += 1
print "additional_count = ", additional_count
params['vocab_size'] = preprocessing.word_to_idx_ctr
tmp = np.array(tmp)
encoder_embedding_matrix = np.vstack(
[encoder_embedding_matrix, tmp])
decoder_embedding_matrix = np.vstack(
[decoder_embedding_matrix, tmp])
print "decoder_embedding_matrix.shape ", decoder_embedding_matrix.shape
print "New vocab size = ", params['vocab_size']
# Make inference
saved_model_path = sys.argv[2]
print "saved_model_path = ", saved_model_path
inference_type = "greedy"
print "inference_type = ", inference_type
params['saved_model_path'] = saved_model_path
rnn_model = solver.Solver(params, buckets=None, mode='inference')
_ = rnn_model.getModel(params, mode='inference', reuse=False, buckets=None)
print "----Running inference-----"
# val
print "val_encoder_inputs = ", val_encoder_inputs
if len(val_decoder_outputs.shape) == 3:
val_decoder_outputs = np.reshape(
val_decoder_outputs,
(val_decoder_outputs.shape[0], val_decoder_outputs.shape[1]))
decoder_outputs_inference, decoder_ground_truth_outputs = rnn_model.solveAll(
params,
val_encoder_inputs,
val_decoder_outputs,
preprocessing.idx_to_word,
inference_type=inference_type)
validOutFile_name = saved_model_path + ".style_transfer_test.output"
original_data_path = data_src + "style_transfer_test.in"
BLEUOutputFile_path = saved_model_path + ".style_transfer_test.BLEU"
utilities.getBlue(validOutFile_name, original_data_path,
BLEUOutputFile_path, decoder_outputs_inference,
decoder_ground_truth_outputs, preprocessing)
def main():
# params
params = {}
params['embeddings_dim'] = config.embeddings_dim
params['lstm_cell_size'] = config.lstm_cell_size
params['max_input_seq_length'] = config.max_input_seq_length
params['max_output_seq_length'] = config.max_output_seq_length-1 #inputs are all but last element, outputs are al but first element
params['batch_size'] = config.batch_size
params['pretrained_embeddings'] = config.use_pretrained_embeddings
params['share_encoder_decoder_embeddings'] = config.share_encoder_decoder_embeddings
params['use_pointer'] = config.use_pointer
params['pretrained_embeddings_path'] = config.pretrained_embeddings_path
params['pretrained_embeddings_are_trainable'] = config.pretrained_embeddings_are_trainable
params['use_additional_info_from_pretrained_embeddings'] = config.use_additional_info_from_pretrained_embeddings
params['max_vocab_size'] = config.max_vocab_size
params['do_vocab_pruning'] = config.do_vocab_pruning
params['use_reverse_encoder'] = config.use_reverse_encoder
params['use_sentinel_loss'] =config.use_sentinel_loss
params['lambd'] = config.lambd
params['use_context_for_out'] = config.use_context_for_out
print("PARAMS:")
for key,value in params.items():
print(" -- ",key," = ",value)
buckets = { 0:{'max_input_seq_length':params['max_input_seq_length'], 'max_output_seq_length':params['max_output_seq_length']} }
#print "buckets = ",buckets
# train
mode=sys.argv[1]
print("mode = ",mode)
########### PREPROCESSING
if mode=="preprocessing":
# preprocessing
print("------------------------------------------------------------------------")
preprocessing = PreProcessing()
splits =["train","valid","test"]
#for split in splits: preprocessing.loadVocab(split)
preprocessing.loadVocab('train')
if params['do_vocab_pruning']:
preprocessing.pruneVocab(max_vocab_size=params['max_vocab_size'])
data_seq = {split:preprocessing.loadData(split=split) for split in splits}
data = { split:preprocessing.prepareMTData(cur_data) for split,cur_data in data_seq.items() }
for split,split_data in data.items():
#print "Split: ",split
inp,dinp,dout,dout_inp_matches = split_data
#print inp.shape, dinp.shape, dout.shape, dout_inp_matches.shape
#print "------------------------------------------------------------------------"
#print ""
pickle.dump(data,open(data_src + "data.obj",mode="wb"))
pickle.dump(preprocessing, open(data_src + "preprocessing.obj",mode="wb") )
return
else:
data = pickle.load(open(data_src + "data.obj","rb") )
preprocessing = pickle.load(open(data_src + "preprocessing.obj","rb") )
params['vocab_size'] = preprocessing.vocab_size
params['preprocessing'] = preprocessing
train = data['train']
val = data['valid']
test = data['test']
# DEBUG
if mode=="debug":
lim = 64
else:
lim= int(params['batch_size'] * ( len(train[0])/params['batch_size']))
if lim!=-1:
train_encoder_inputs, train_decoder_inputs, train_decoder_outputs, train_decoder_outputs_matching_inputs = train
train_encoder_inputs = train_encoder_inputs[:lim]
train_decoder_inputs = train_decoder_inputs[:lim]
train_decoder_outputs = train_decoder_outputs[:lim]
train_decoder_outputs_matching_inputs = train_decoder_outputs_matching_inputs[:lim]
train = train_encoder_inputs, train_decoder_inputs, train_decoder_outputs, train_decoder_outputs_matching_inputs
#Pretrained embeddibngs
if params['pretrained_embeddings']:
pretrained_embeddings = pickle.load(open(params['pretrained_embeddings_path'],mode="rb"),encoding='bytes')
word_to_idx = preprocessing.word_to_idx
encoder_embedding_matrix = np.random.rand( params['vocab_size'], params['embeddings_dim'] )
decoder_embedding_matrix = np.random.rand( params['vocab_size'], params['embeddings_dim'] )
not_found_count = 0
for token,idx in word_to_idx.items():
if token in pretrained_embeddings:
encoder_embedding_matrix[idx]=pretrained_embeddings[token]
decoder_embedding_matrix[idx]=pretrained_embeddings[token]
else:
if not_found_count<10:
print("No pretrained embedding for (only first 10 such cases will be printed. other prints are suppressed) ",token)
not_found_count+=1
#print "not found count = ", not_found_count
params['encoder_embeddings_matrix'] = encoder_embedding_matrix
params['decoder_embeddings_matrix'] = decoder_embedding_matrix
if params['use_additional_info_from_pretrained_embeddings']:
additional_count=0
tmp=[]
for token in pretrained_embeddings:
if token not in preprocessing.word_to_idx:
preprocessing.word_to_idx[token] = preprocessing.word_to_idx_ctr
preprocessing.idx_to_word[preprocessing.word_to_idx_ctr] = token
preprocessing.word_to_idx_ctr+=1
tmp.append(pretrained_embeddings[token])
additional_count+=1
#print "additional_count = ",additional_count
params['vocab_size'] = preprocessing.word_to_idx_ctr
tmp = np.array(tmp)
encoder_embedding_matrix = np.vstack([encoder_embedding_matrix,tmp])
decoder_embedding_matrix = np.vstack([decoder_embedding_matrix,tmp])
#print "decoder_embedding_matrix.shape ",decoder_embedding_matrix.shape
#print "New vocab size = ",params['vocab_size']
# TRAIN/DEBUG
if mode=='train' or mode=="debug":
if mode=="train":
training_iters = int(sys.argv[2])
model_name = sys.argv[3]
else:
training_iters = 5
model_name = "test"
params['training_iters'] = training_iters
params['model_name'] = model_name
train_buckets = {}
for bucket,_ in enumerate(buckets):
train_buckets[bucket] = train
rnn_model = solver.Solver(params,buckets)
_ = rnn_model.getModel(params, mode='train',reuse=False, buckets=buckets)
rnn_model.trainModel(config=params, train_feed_dict=train_buckets, val_feed_dct=val, reverse_vocab=preprocessing.idx_to_word, do_init=True)
# INFERENCE
elif mode=="inference":
saved_model_path = sys.argv[2]
print("saved_model_path = ",saved_model_path)
inference_type = sys.argv[3] # greedy / beam
print("inference_type = ",inference_type)
params['saved_model_path'] = saved_model_path
rnn_model = solver.Solver(params, buckets=None, mode='inference')
_ = rnn_model.getModel(params, mode='inference', reuse=False, buckets=None)
print("----Running inference-----")
#val
val_encoder_inputs, val_decoder_inputs, val_decoder_outputs, val_decoder_outputs_matching_inputs = val
#print "val_encoder_inputs = ",val_encoder_inputs
if len(val_decoder_outputs.shape)==3:
val_decoder_outputs=np.reshape(val_decoder_outputs, (val_decoder_outputs.shape[0], val_decoder_outputs.shape[1]))
decoder_outputs_inference, decoder_ground_truth_outputs = rnn_model.solveAll(params, val_encoder_inputs, val_decoder_outputs, preprocessing.idx_to_word, inference_type=inference_type)
validOutFile_name = saved_model_path+".valid.output"
original_data_path = data_src + "valid.original.nltktok"
BLEUOutputFile_path = saved_model_path + ".valid.BLEU"
utilities.getBlue(validOutFile_name, original_data_path, BLEUOutputFile_path, decoder_outputs_inference, decoder_ground_truth_outputs, preprocessing)
print("VALIDATION: ",open(BLEUOutputFile_path,"r").read())
#test
test_encoder_inputs, test_decoder_inputs, test_decoder_outputs, test_decoder_outputs_matching_inputs = test
if len(test_decoder_outputs.shape)==3:
test_decoder_outputs=np.reshape(test_decoder_outputs, (test_decoder_outputs.shape[0], test_decoder_outputs.shape[1]))
decoder_outputs_inference, decoder_ground_truth_outputs = rnn_model.solveAll(params, test_encoder_inputs, test_decoder_outputs, preprocessing.idx_to_word, inference_type=inference_type)
validOutFile_name = saved_model_path+".test.output"
original_data_path = data_src + "test.original.nltktok"
BLEUOutputFile_path = saved_model_path + ".test.BLEU"
utilities.getBlue(validOutFile_name, original_data_path, BLEUOutputFile_path, decoder_outputs_inference, decoder_ground_truth_outputs, preprocessing)
print("TEST: ",open(BLEUOutputFile_path,"r").read())
else:
print("Please see usage")