def experiment_fn(run_config, params):
data = Data(FLAGS)
data.initialize_word_vectors()
model = Seq2seq(data.vocab_size, FLAGS, data.embeddings_mat)
estimator = tf.estimator.Estimator(
model_fn=model.make_graph,
# model_dir=FLAGS.model_dir,
config=run_config,
params=FLAGS)
train_input_fn, train_feed_fn = data.make_input_fn('train')
eval_input_fn, eval_feed_fn = data.make_input_fn('test')
print_vars = [
'source', 'predict'
# 'decoder_output',
# 'actual'
]
print_inputs = tf.train.LoggingTensorHook(print_vars,
every_n_iter=FLAGS.print_every,
formatter=data.get_formatter(
['source', 'predict']))
experiment = tf.contrib.learn.Experiment(
estimator=estimator,
train_input_fn=train_input_fn,
eval_input_fn=eval_input_fn,
train_steps=FLAGS.iterations,
min_eval_frequency=FLAGS.print_every,
train_monitors=[tf.train.FeedFnHook(train_feed_fn), print_inputs],
eval_hooks=[tf.train.FeedFnHook(eval_feed_fn)],
eval_steps=10)
return experiment
def main(args):
# tf.logging._logger.setLevel(logging.INFO)
tf.logging.set_verbosity(logging.INFO)
data = Data(FLAGS)
model = Seq2seq(data.vocab_size, FLAGS)
input_fn, feed_fn = data.make_input_fn()
print_inputs = tf.train.LoggingTensorHook(
['source', 'target', 'predict'],
every_n_iter=FLAGS.print_every,
formatter=data.get_formatter(['source', 'target', 'predict']))
estimator = tf.estimator.Estimator(
model_fn=model.make_graph, model_dir=FLAGS.model_dir) #, params=FLAGS)
estimator.train(input_fn=input_fn,
hooks=[tf.train.FeedFnHook(feed_fn), print_inputs],
steps=FLAGS.iterations)
def trainWithPreviousKnowledge(datasetPath,
datasetSize,
transferMethod=None,
transferVocabularyPath=None,
sourceCheckpointPath=None,
suffix=""):
percentages = [0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
checkpoint_filename = "checkpointsAdaptiveWithPrev" + suffix
size = datasetSize
epoch = 5
test_source = datasetPath + "test_source.txt"
test_target = datasetPath + "test_target.txt"
resultPath = datasetPath + "AdaptiveWithPrev"
if (transferMethod is None):
vocabulary = datasetPath + "v.txt"
else:
vocabulary = transferVocabularyPath
data = Data(FLAGS, "", "", "", "", vocabulary)
model = Seq2seq(data.vocab_size, FLAGS, transferMethod,
sourceCheckpointPath)
for i in percentages:
source_filename = datasetPath + "without-pool/" + format(
i, '.1f') + "_source" + ".txt"
target_filename = datasetPath + "without-pool/" + format(
i, '.1f') + "_target" + ".txt"
data = Data(FLAGS, source_filename, target_filename, test_source,
test_target, vocabulary)
iterations = int(round(size * i * epoch / FLAGS.batch_size))
# iterations = 1
input_fn, feed_fn = data.make_input_fn()
test_fn = data.make_test_fn()
print_inputs = tf.train.LoggingTensorHook(
['source', 'target', 'predict'],
every_n_iter=FLAGS.print_every,
formatter=data.get_formatter(['source', 'target', 'predict']))
estimator = tf.estimator.Estimator(model_fn=model.make_graph,
model_dir=checkpoint_filename,
params=FLAGS)
print("Training with " + format(i, '.2f') + " percent of the dataset.")
estimator.train(input_fn=input_fn,
hooks=[tf.train.FeedFnHook(feed_fn), print_inputs],
steps=iterations)
model.setLoadParameters(False)
# test_paraphrases = list(estimator.predict(test_fn))
a = estimator.predict(test_fn)
test_paraphrases = []
# a = list(a)
# for i in a:
# print(i.shape)
# print(len(a))
for j in a:
# j = i[:, 0]
test_paraphrases.append(j)
data.builtTranslationCorpus(test_paraphrases)
scr = evaluate(data.reference_corpus, data.translation_corpus)
print(i, scr)
saveResult(i, scr, resultPath)
def supervisedLearning(datasetPath,
datasetSize,
transferMethod=None,
transferVocabularyPath=None,
sourceCheckpointPath=None,
suffix=""):
test_source = datasetPath + "test_source.txt"
test_target = datasetPath + "test_target.txt"
train_source = datasetPath + "train_source.txt"
train_target = datasetPath + "train_target.txt"
resultPath = datasetPath + "SL"
if (transferMethod is None):
vocabulary = datasetPath + "quora_msr_vocabulary.txt"
else:
vocabulary = transferVocabularyPath
data = Data(FLAGS, train_source, train_target, test_source, test_target,
vocabulary)
model = Seq2seq(data.vocab_size, FLAGS, transferMethod,
sourceCheckpointPath)
size = datasetSize
epoch = 5
# iterations = int(round(size * epoch / FLAGS.batch_size))
iterations = 4
# var_list = checkpoint_utils.list_variables('checkpoints')
# for v in var_list: print(v)
input_fn, feed_fn = data.make_input_fn()
print_inputs = tf.train.LoggingTensorHook(
['source', 'target', 'predict'],
every_n_iter=FLAGS.print_every,
formatter=data.get_formatter(['source', 'target', 'predict']))
# estimator = tf.estimator.Estimator(model_fn = model.make_graph, model_dir="checkpointsQuoraMSR")
# estimator.train(input_fn=input_fn, hooks=[tf.train.FeedFnHook(feed_fn), print_inputs], steps=iterations)
# modelInfer = Seq2seq(data.vocab_size, FLAGS, transferMethod, sourceCheckpointPath, False, inferGraph = 1)
estimator = tf.estimator.Estimator(
model_fn=model.make_graph,
model_dir="data/cps/checkpointsQuoraMSRinit")
model.setLoadParameters(False)
test_fn = data.make_test_fn()
a = estimator.predict(test_fn)
test_paraphrases = []
# a = list(a)
# for i in a:
# print(i.shape)recent # print(len(a))
for j in a:
# j = i[:, 0]
test_paraphrases.append(j)
data.builtTranslationCorpus(test_paraphrases)
scr = evaluate(data.reference_corpus, data.translation_corpus)
print(data.translation_corpus)
print(scr)
saveResult(100, scr, resultPath)