def write_to_tfrecords(filename, destination_dir, responses, prompts, q_ids, grades, speakers, targets, predictions,
debug=False):
# Check that all the input lists are of equal lengths
assert len({len(responses), len(prompts), len(q_ids), len(grades), len(speakers), len(targets), len(predictions)}) == 1
# Create the training TF Record file
print('Writing: ', filename)
writer = tf.python_io.TFRecordWriter(os.path.join(destination_dir, filename))
for response, prompt, q_id, grd, spkr, tgt, example_pred, idx in zip(responses, prompts, q_ids, grades, speakers, targets, predictions, range(len(q_ids))):
example = tf.train.SequenceExample(
context=tf.train.Features(feature={
'targets': tfrecord_utils.float_feature([tgt]),
'grade': tfrecord_utils.float_feature([float(grd)]),
'teacher_pred': tfrecord_utils.float_feature(list(example_pred)),
'spkr': tfrecord_utils.bytes_feature([spkr]),
'q_id': tfrecord_utils.int64_feature([q_id]),
'example_idx': tfrecord_utils.int64_feature([idx]) # Stores the example number for easy back-reference to txt files even when examples get shuffled (0 indexed)
}),
feature_lists=tf.train.FeatureLists(feature_list={
'response': tfrecord_utils.int64_feature_list(response),
'prompt': tfrecord_utils.int64_feature_list(prompt)}))
if debug:
# Print out the data that is going to be saved:
print("-----------------\n", "EXAMPLE: \n", "Response: {}\nPrompt: {}\nQ_id: {}\n\ntarget: {}\ngrade: {}\n,teacher_pred: {}\nexample_num: {}\n\n".format(response, prompt, q_id, tgt, grd, example_pred, idx))
writer.write(example.SerializeToString())
writer.close()
return
def main(argv=None):
"""Converts a dataset to tfrecords."""
args = commandLineParser.parse_args()
if os.path.isdir(args.destination_dir):
print 'destination directory exists. Exiting...'
else:
os.makedirs(args.destination_dir)
if not os.path.isdir('CMDs'):
os.makedirs('CMDs')
with open('CMDs/step_preprocess_data.cmd', 'a') as f:
f.write(' '.join(sys.argv) + '\n')
f.write('--------------------------------\n')
# Load responses and prompts as sequences of word ids
responses, _ = load_text(args.input_data_path, args.input_wlist_path)
prompts, _ = load_text(args.input_prompt_path, args.input_wlist_path)
# Load up the prompts as sequences of words
with open(args.input_prompt_path, 'r') as file:
topics = [line.replace('\n', '') for line in file.readlines()]
# Get unique set of topics and topic counts (and sort tem)
unique_topics, topic_counts = np.unique(topics, return_counts=True)
topics = unique_topics[np.flip(np.argsort(topic_counts), 0)]
topic_counts = np.flip(np.sort(topic_counts), 0)
# Create dictionary for topics mapping sentence to topic id
# Also create file of sorted topics and unigrams file
# Unigram file later used for training
topic_dict = {}
with open(os.path.join(args.destination_dir, 'unigrams.txt'),
'w') as ufile:
with open(os.path.join(args.destination_dir, 'sorted_topics.txt'),
'w') as tfile:
for i, topic, count in zip(xrange(topics.shape[0]), topics,
topic_counts):
topic_dict[topic] = i
ufile.write(str(i) + ',' + str(int(count)) + '\n')
tfile.write(topic + '\n')
# Load up the speakers and speakers
grades = np.loadtxt(args.input_grade_path)
with open(args.input_spkr_path, 'r') as file:
speakers = np.asarray(
[line.replace('\n', '') for line in file.readlines()])
# Create a list of topic IDs for every response
with open(args.input_prompt_path, 'r') as file:
q_ids = np.asarray(
[topic_dict[line.replace('\n', '')] for line in file.readlines()])
### Split data into train and validation data sets
n = len(responses)
train_size = int(n * (1.0 - args.valid_fraction))
valid_size = n - train_size
print 'Total dataset size', n, 'Train dataset size', train_size, 'Valid dataset size', valid_size
np.random.seed(1000)
permutation = np.random.choice(np.arange(n), n, replace=False)
index_train = permutation[:train_size]
inded_valid = permutation[train_size:]
trn_responses = responses[index_train]
trn_prompts = prompts[index_train]
trn_q_ids = q_ids[index_train]
trn_speakers = speakers[index_train]
trn_grades = grades[index_train]
valid_responses = responses[inded_valid]
valid_prompts = prompts[inded_valid]
valid_q_ids = q_ids[inded_valid]
valid_speakers = speakers[inded_valid]
valid_grades = grades[inded_valid]
# Create the training TF Record file
filename = 'relevance.train.tfrecords'
print 'Writing', filename
writer = tf.python_io.TFRecordWriter(
os.path.join(args.destination_dir, filename))
for response, prompt, q_id, grd, spkr in zip(trn_responses, trn_prompts,
trn_q_ids, trn_grades,
trn_speakers):
example = tf.train.SequenceExample(
context=tf.train.Features(
feature={
'targets': tfrecord_utils.float_feature([1.0]),
'grade': tfrecord_utils.float_feature([grd]),
'spkr': tfrecord_utils.bytes_feature([spkr]),
'q_id': tfrecord_utils.int64_feature([q_id])
}),
feature_lists=tf.train.FeatureLists(
feature_list={
'response': tfrecord_utils.int64_feature_list(response),
'prompt': tfrecord_utils.int64_feature_list(prompt)
}))
writer.write(example.SerializeToString())
writer.close()
# Create the validation TF Record file
filename = 'relevance.valid.tfrecords'
print 'Writing', filename
writer = tf.python_io.TFRecordWriter(
os.path.join(args.destination_dir, filename))
for response, prompt, q_id, grd, spkr in zip(valid_responses,
valid_prompts, valid_q_ids,
valid_grades, valid_speakers):
example = tf.train.SequenceExample(
context=tf.train.Features(
feature={
'targets': tfrecord_utils.float_feature([1.0]),
'grade': tfrecord_utils.float_feature([grd]),
'spkr': tfrecord_utils.bytes_feature([spkr]),
'q_id': tfrecord_utils.int64_feature([q_id])
}),
feature_lists=tf.train.FeatureLists(
feature_list={
'response': tfrecord_utils.int64_feature_list(response),
'prompt': tfrecord_utils.int64_feature_list(prompt)
}))
writer.write(example.SerializeToString())
writer.close()
def main(argv=None):
"""Converts a dataset to tfrecords."""
args = commandLineParser.parse_args()
if not os.path.isdir(args.destination_dir):
os.makedirs(args.destination_dir)
if not os.path.isdir('CMDs'):
os.makedirs('CMDs')
with open('CMDs/step_preprocess_test_data.cmd', 'a') as f:
f.write(' '.join(sys.argv) + '\n')
f.write('--------------------------------\n')
# Load responses and prompts as sequences of word ids
responses, _ = load_text(args.input_data_path, args.input_wlist_path)
prompts, _ = load_text(args.input_prompt_path, args.input_wlist_path)
# Load up the grades, targets and speakers
grades = np.loadtxt(args.input_grade_path)
targets = np.loadtxt(args.input_tgt_path, dtype=np.float32)
with open(args.input_spkr_path, 'r') as file:
speakers = np.asarray(
[line.replace('\n', '') for line in file.readlines()])
# Load up sorted topics and (re)construct the topic dict so that I map each prompt word sequence to its q_id
topic_dict = {}
i = 0
with open(os.path.join(args.sorted_topics_path), 'r') as tfile:
for topic in tfile.readlines():
topic_dict[topic.replace('\n', '')] = i
i += 1
# Load up the prompts as sequences of words and convert to q_id
try:
with open(args.input_prompt_path, 'r') as file:
q_ids = np.asarray([
topic_dict[line.replace('\n', '')]
for line in file.readlines()
])
except:
with open(args.input_prompt_path, 'r') as file:
q_ids = np.asarray([-1 for line in file.readlines()])
# Create the training TF Record file
filename = args.name + '.tfrecords'
print 'Writing', filename
writer = tf.python_io.TFRecordWriter(
os.path.join(args.destination_dir, filename))
for response, prompt, q_id, grd, spkr, tgt in zip(responses, prompts,
q_ids, grades, speakers,
targets):
example = tf.train.SequenceExample(
context=tf.train.Features(
feature={
'targets': tfrecord_utils.float_feature([tgt]),
'grade': tfrecord_utils.float_feature([grd]),
'spkr': tfrecord_utils.bytes_feature([spkr]),
'q_id': tfrecord_utils.int64_feature([q_id])
}),
feature_lists=tf.train.FeatureLists(
feature_list={
'response': tfrecord_utils.int64_feature_list(response),
'prompt': tfrecord_utils.int64_feature_list(prompt)
}))
writer.write(example.SerializeToString())
writer.close()
def write_to_tfrecords(filename,
destination_dir,
responses,
prompts,
q_ids,
grades,
speakers,
targets=1.0,
debug=False):
# Check that all the input lists are of equal lengths
# TEMP
print(len(responses))
print(len(prompts))
print(len(q_ids))
print(len(grades))
print(len(speakers))
assert len(
{len(responses),
len(prompts),
len(q_ids),
len(grades),
len(speakers)}) == 1
if type(targets) is float or type(targets) is int:
# If targets is an integer make each target this value
targets = [float(targets)] * len(responses)
else:
assert type(targets) is list
assert len(targets) == len(responses)
# Create the training TF Record file
print('Writing: ', filename)
writer = tf.python_io.TFRecordWriter(
os.path.join(destination_dir, filename))
for response, prompt, q_id, grd, spkr, tgt in zip(responses, prompts,
q_ids, grades, speakers,
targets):
if debug:
# Print out the data that is going to be saved:
print(
"-----------------\n", "EXAMPLE: \n",
"Response: {}\nPrompt: {}\nQ_id: {}\n\ntarget: {}\ngrade: {}\n\n"
.format(response, prompt, q_id, tgt, grd))
example = tf.train.SequenceExample(
context=tf.train.Features(
feature={
'targets': tfrecord_utils.float_feature([tgt]),
'grade': tfrecord_utils.float_feature([float(grd)]),
'spkr': tfrecord_utils.bytes_feature([spkr]),
'q_id': tfrecord_utils.int64_feature([q_id])
}),
feature_lists=tf.train.FeatureLists(
feature_list={
'response': tfrecord_utils.int64_feature_list(response),
'prompt': tfrecord_utils.int64_feature_list(prompt)
}))
writer.write(example.SerializeToString())
writer.close()
return