def predict(self, batch_index=0):
_, i, *__ = self.feed.nth_batch(batch_index)
self.runner.model.eval()
output = self.runner.run(i)
results = ListTable()
results.extend(self.repr_function(output, self.feed, batch_index))
output_ = output.data
del output
del _, i, __
return output_, results
def predict(self, batch_index=0):
log.debug('batch_index: {}'.format(batch_index))
_, i, *__ = self.feed.nth_batch(batch_index)
log.debug('input_shape: {}'.format([j.shape for j in i]))
self.runner.model.eval()
output = self.runner.run(i)
results = ListTable()
results.extend(self.repr_function(output, self.feed, batch_index))
output_ = output
return output_, results
def experiment(eons=1000, epochs=10, checkpoint=5):
try:
try:
model = BiLSTMDecoderModel(Config(), len(INPUT_VOCAB), len(CHAR_VOCAB), len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
model.load_state_dict(torch.load('attn_model.pth'))
log.info('loaded the old image for the model')
except:
log.exception('failed to load the model')
model = BiLSTMDecoderModel(Config(), len(INPUT_VOCAB), len(CHAR_VOCAB), len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
print('**** the model', model)
train_feed, test_feed, predictor_feed = {}, {}, {}
trainer, predictor = {}, {}
max_size = max( sorted( [len(i[0]) for i in classified_datapoints.values()] )[:-1] )
#max_size = max( sorted( [len(i[0]) for i in classified_datapoints.values()] ) )
for label in classified_datapoints.keys():
if len(classified_datapoints[label][0]) < 1: continue
label_desc = '-'.join([OUTPUT_VOCAB[l] for l in [i for i, x in enumerate(label) if x == 1]] )
print('label: {} and size: {}'.format(label, len(classified_datapoints[label][0])))
train_feed[label] = DataFeed(label_desc, classified_datapoints[label][0], batchop=batchop, batch_size=max(128, int(len(classified_datapoints[label][0])/600)) )
test_feed[label] = DataFeed(label_desc, classified_datapoints[label][1], batchop=batchop, batch_size=32)
predictor_feed[label] = DataFeed(label_desc, classified_datapoints[label][1], batchop=batchop, batch_size=12)
turns = int(max_size/train_feed[label].size) + 1
trainer[label] = Trainer(name=label_desc,
model=model,
loss_function=partial(loss, scale=1), accuracy_function=accuracy, f1score_function=f1score_function,
checkpoint=checkpoint, epochs=epochs,
feeder = Feeder(train_feed[label], test_feed[label]))
predictor[label] = Predictor(model=model, feed=predictor_feed[label], repr_function=repr_function)
test_predictor_feed = DataFeed('test', test_datapoints, batchop=test_batchop, batch_size=128)
test_predictor = Predictor(model=model, feed=test_predictor_feed, repr_function=test_repr_function)
all_class_train_feed = MultiplexedDataFeed('atrain', train_feed.values(), batchop=batchop, batch_size=256)
all_class_test_feed = MultiplexedDataFeed('atest', test_feed.values(), batchop=batchop, batch_size=256)
all_class_predictor_feed = MultiplexedDataFeed('apredict',predictor_feed.values(), batchop=batchop, batch_size=256)
all_class_trainer = Trainer(name='all_class_trainer',
model=model,
loss_function=partial(loss, scale=1), accuracy_function=accuracy, f1score_function=f1score_function,
checkpoint=checkpoint, epochs=epochs,
feeder = Feeder(all_class_train_feed, all_class_test_feed))
all_class_predictor = Predictor(model=model, feed=all_class_predictor_feed, repr_function=repr_function)
label_trainer_triples = sorted( [(l, t, train_feed[l].size) for l, t in trainer.items()], key=lambda x: x[2] )
log.info('trainers built {}'.format(pformat(label_trainer_triples)))
dump = open('results/experiment_attn.csv', 'w').close()
for e in range(eons):
dump = open('results/experiment_attn.csv', 'a')
dump.write('#========================after eon: {}\n'.format(e))
dump.close()
log.info('on {}th eon'.format(e))
if e and not e % 1:
test_results = ListTable()
test_dump = open('results/experiment_attn_over_test_{}.csv'.format(e), 'w')
test_dump.write('|'.join(['id', 'toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']) + '\n')
log.info('running over test')
for i in tqdm(range(test_predictor_feed.num_batch)):
log.debug('i: {}'.format(i))
output, results = test_predictor.predict(i)
test_results += results
test_dump.write(repr(test_results))
test_dump.close()
with open('results/experiment_attn.csv', 'a') as dump:
output, results = all_class_predictor.predict(random.choice(range(all_class_predictor_feed.num_batch)))
dump.write(repr(results))
del output, results
all_class_trainer.train()
"""
for label, _, _ in reversed(label_trainer_triples):
if not sum(label) and e and not e % 10: #Avoid neutral classes in every epoch
continue
label_desc = '-'.join([OUTPUT_VOCAB[l] for l in [i for i, x in enumerate(label) if x == 1]] )
log.info('=================================== training for {} datapoints ========================================'.format(label_desc))
with open('results/experiment_attn.csv', 'a') as dump:
output, results = predictor[label].predict(random.choice(range(predictor_feed[label].num_batch)))
dump.write(repr(results))
del output, results
turns = int(max_size/train_feed[label].size/6) + 1
log.info('======================== size: {} and turns: {}==========================================='.format(train_feed[label].size, turns))
for turn in range(turns):
log.info('================================== label: {} and turn: {}/{}====================================='.format(label_desc, turn, turns))
trainer[label].train()
"""
except:
log.exception('####################')
torch.save(model.state_dict(), open('attn_model.pth', 'wb'))
return locals()
def experiment(VOCAB,
LABELS,
WORD2INDEX,
LABEL2INDEX,
raw_samples,
loss_weight,
datapoints=[[], []],
eons=1000,
epochs=10,
checkpoint=5):
try:
try:
model = BiLSTMDecoderModel(Config(), len(VOCAB), len(LABELS))
if Config().cuda: model = model.cuda()
log.info('loaded the old image for the model, from {}'.format(
'{}.{}'.format(SELF_NAME, 'pth')))
model.load_state_dict(torch.load('{}.{}'.format(SELF_NAME, 'pth')))
log.info('loaded the old image for the model')
except:
log.exception('failed to load the model')
model = BiLSTMDecoderModel(Config(), len(VOCAB), len(LABELS))
if Config().cuda: model = model.cuda()
print('**** the model', model)
name = SELF_NAME
_batchop = partial(batchop, LABEL2INDEX=LABEL2INDEX)
train_feed = DataFeed(name,
datapoints[0],
batchop=_batchop,
vocab=WORD2INDEX,
batch_size=1024)
test_feed = DataFeed(name,
datapoints[1],
batchop=_batchop,
vocab=WORD2INDEX,
batch_size=256)
predictor_feed = DataFeed(name,
datapoints[1],
batchop=_batchop,
vocab=WORD2INDEX,
batch_size=128)
loss_weight = Variable(torch.Tensor(loss_weight))
if Config.cuda: loss_weight = loss_weight.cuda()
_loss = partial(loss, loss_function=nn.NLLLoss(loss_weight))
trainer = Trainer(name=name,
model=model,
loss_function=_loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(train_feed, test_feed))
predictor = Predictor(model=model,
feed=predictor_feed,
repr_function=partial(repr_function,
VOCAB=VOCAB,
LABELS=LABELS,
raw_samples=raw_samples))
for e in range(eons):
dump = open('results/experiment_attn.csv', 'a')
dump.write('#========================after eon: {}\n'.format(e))
dump.close()
log.info('on {}th eon'.format(e))
with open('results/experiment_attn.csv', 'a') as dump:
results = ListTable()
for ri in range(predictor_feed.num_batch):
output, _results = predictor.predict(ri)
results.extend(_results)
dump.write(repr(results))
if not trainer.train():
raise Exception
except:
log.exception('####################')
trainer.save_best_model()
return locals()
def experiment(eons=1000, epochs=1, checkpoint=1):
try:
model = BiLSTMDecoderModel(Config(), len(INPUT_VOCAB), len(CHAR_VOCAB),
len(OUTPUT_VOCAB))
if Config().cuda: model = model.cuda()
classified_train_feed = DataFeed(classified_datapoints[0],
batchop=batchop,
batch_size=128)
classified_test_feed = DataFeed(classified_datapoints[1],
batchop=batchop,
batch_size=128)
classified_trainer = Trainer(model=model,
loss_function=loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(classified_train_feed,
classified_test_feed))
classified_predictor_feed = DataFeed(classified_datapoints[1],
batchop=batchop,
batch_size=12)
classified_predictor = Predictor(model=model,
feed=classified_predictor_feed,
repr_function=repr_function)
non_classified_train_feed = DataFeed(non_classified_datapoints[0],
batchop=batchop,
batch_size=128)
non_classified_test_feed = DataFeed(non_classified_datapoints[1],
batchop=batchop,
batch_size=128)
non_classified_trainer = Trainer(model=model,
loss_function=loss,
accuracy_function=accuracy,
checkpoint=checkpoint,
epochs=epochs,
feeder=Feeder(
non_classified_train_feed,
non_classified_test_feed))
non_classified_predictor_feed = DataFeed(non_classified_datapoints[1],
batchop=batchop,
batch_size=12)
non_classified_predictor = Predictor(
model=model,
feed=non_classified_predictor_feed,
repr_function=repr_function)
test_predictor_feed = DataFeed(test_datapoints,
batchop=test_batchop,
batch_size=128)
test_predictor = Predictor(model=model,
feed=test_predictor_feed,
repr_function=test_repr_function)
dump = open('results/experiment_attn.csv', 'w')
for e in range(eons):
dump.write('#========================after eon: {}\n'.format(e))
log.info('on {}th eon'.format(e))
output, results = classified_predictor.predict(
random.choice(range(classified_predictor_feed.num_batch)))
dump.write(repr(results))
del output, results
output, results = non_classified_predictor.predict(
random.choice(range(non_classified_predictor_feed.num_batch)))
dump.write(repr(results))
del output, results
#non_classified_trainer.train()
#for i in range(int(non_classified_train_feed.size/classified_train_feed.size)):
classified_trainer.train()
if e and not e % 10:
test_results = ListTable()
test_dump = open(
'results/experiment_attn_over_test_{}.csv'.format(e), 'w')
test_dump.write('|'.join([
'id', 'toxic', 'severe_toxic', 'obscene', 'threat',
'insult', 'identity_hate'
]) + '\n')
log.info('running over test')
for i in tqdm(range(test_predictor_feed.num_batch)):
log.debug('i: {}'.format(i))
output, results = test_predictor.predict(i)
test_results += results
test_dump.write(repr(test_results))
test_dump.close()
except KeyboardInterrupt:
torch.save(model.state_dict(), open('attn_model.pth', 'wb'))
dump.close()
return locals()