removed tfrecord tensor code and remnants
Malar Kannan
parent
e9b18921ee
commit
1190312def
205
speech_data.py
205
speech_data.py
|
|
@ -110,67 +110,6 @@ def padd_zeros(spgr, max_samples):
|
||||||
return np.lib.pad(spgr, [(0, max_samples - spgr.shape[0]), (0, 0)],
|
return np.lib.pad(spgr, [(0, max_samples - spgr.shape[0]), (0, 0)],
|
||||||
'constant')
|
'constant')
|
||||||
|
|
||||||
def find_max_n(trf):
|
|
||||||
max_n,n_records = 0,0
|
|
||||||
max_n_it = tf.python_io.tf_record_iterator(path=trf)
|
|
||||||
for string_record in max_n_it:
|
|
||||||
example = tf.train.Example()
|
|
||||||
example.ParseFromString(string_record)
|
|
||||||
spec_n1 = example.features.feature['spec_n1'].int64_list.value[0]
|
|
||||||
spec_n2 = example.features.feature['spec_n2'].int64_list.value[0]
|
|
||||||
max_n = max([max_n,spec_n1,spec_n2])
|
|
||||||
n_records+=1
|
|
||||||
return (max_n,n_records)
|
|
||||||
|
|
||||||
def padd_zeros_siamese_tfrecords(audio_group='audio'):
|
|
||||||
records_file = os.path.join('./outputs',audio_group+'.tfrecords')
|
|
||||||
record_iterator = tf.python_io.tf_record_iterator(path=records_file)
|
|
||||||
print('finding max_n...')
|
|
||||||
max_n,n_records = find_max_n(records_file)
|
|
||||||
p_spec1 = None
|
|
||||||
print('reading tfrecords...')
|
|
||||||
writer = tf.python_io.TFRecordWriter('./outputs/' + audio_group + '_padded.tfrecords')
|
|
||||||
for string_record in tqdm(record_iterator,desc='padding siamese record',total=n_records):
|
|
||||||
example = tf.train.Example()
|
|
||||||
example.ParseFromString(string_record)
|
|
||||||
spec_n1 = example.features.feature['spec_n1'].int64_list.value[0]
|
|
||||||
spec_n2 = example.features.feature['spec_n2'].int64_list.value[0]
|
|
||||||
spec_w1 = example.features.feature['spec_w1'].int64_list.value[0]
|
|
||||||
spec_w2 = example.features.feature['spec_w2'].int64_list.value[0]
|
|
||||||
spec1 = np.array(example.features.feature['spec1'].float_list.value).reshape(spec_n1,spec_w1)
|
|
||||||
spec2 = np.array(example.features.feature['spec2'].float_list.value).reshape(spec_n2,spec_w2)
|
|
||||||
p_spec1,p_spec2 = padd_zeros(spec1,max_n),padd_zeros(spec2,max_n)
|
|
||||||
output = example.features.feature['output'].int64_list.value
|
|
||||||
w_example = tf.train.Example(features=tf.train.Features(
|
|
||||||
feature={
|
|
||||||
'spec1':_float_feature(p_spec1.reshape(-1)),
|
|
||||||
'spec2':_float_feature(p_spec2.reshape(-1)),
|
|
||||||
'output':_int64_feature(output)
|
|
||||||
}
|
|
||||||
))
|
|
||||||
writer.write(w_example.SerializeToString())
|
|
||||||
const_file = os.path.join('./outputs',audio_group+'.constants')
|
|
||||||
pickle.dump((max_n,p_spec1.shape[1],n_records),open(const_file,'wb'))
|
|
||||||
writer.close()
|
|
||||||
|
|
||||||
def pickle_constants(audio_group='audio'):
|
|
||||||
records_file = os.path.join('./outputs',audio_group+'_padded.tfrecords')
|
|
||||||
record_iterator = tf.python_io.tf_record_iterator(path=records_file)
|
|
||||||
print('finding max_n...')
|
|
||||||
max_n,n_records = find_max_n(records_file)
|
|
||||||
spec1 = 0
|
|
||||||
print('finding spec_w1...')
|
|
||||||
record_iterator = tf.python_io.tf_record_iterator(path=records_file)
|
|
||||||
for string_record in record_iterator:
|
|
||||||
example = tf.train.Example()
|
|
||||||
example.ParseFromString(string_record)
|
|
||||||
spec1 = len(example.features.feature['spec1'].float_list.value)//max_n
|
|
||||||
print('found spec_w1...')
|
|
||||||
break
|
|
||||||
const_file = os.path.join('./outputs',audio_group+'.constants')
|
|
||||||
print(max_n,spec1,n_records)
|
|
||||||
pickle.dump((max_n,spec1,n_records),open(const_file,'wb'))
|
|
||||||
|
|
||||||
def reservoir_sample(iterable, k):
|
def reservoir_sample(iterable, k):
|
||||||
it = iter(iterable)
|
it = iter(iterable)
|
||||||
if not (k > 0):
|
if not (k > 0):
|
||||||
|
|
@ -185,36 +124,6 @@ def reservoir_sample(iterable, k):
|
||||||
sample[j] = item # replace item with gradually decreasing probability
|
sample[j] = item # replace item with gradually decreasing probability
|
||||||
return sample
|
return sample
|
||||||
|
|
||||||
def read_siamese_tfrecords_oneshot(audio_group='audio',sample_size=3000):
|
|
||||||
records_file = os.path.join('./outputs',audio_group+'_padded.tfrecords')
|
|
||||||
record_iterator = tf.python_io.tf_record_iterator(path=records_file)
|
|
||||||
input_pairs = []
|
|
||||||
output_class = []
|
|
||||||
const_file = os.path.join('./outputs',audio_group+'.constants')
|
|
||||||
(n_spec,n_features,n_records) = pickle.load(open(const_file,'rb'))
|
|
||||||
print('reading tfrecords({})...'.format(audio_group))
|
|
||||||
samples = min([sample_size,n_records])
|
|
||||||
input_data = np.zeros((samples,2,n_spec,n_features))
|
|
||||||
output_data = np.zeros((samples,2))
|
|
||||||
random_samples = enumerate(reservoir_sample(record_iterator,samples))
|
|
||||||
for (i,string_record) in tqdm(random_samples,total=samples):
|
|
||||||
example = tf.train.Example()
|
|
||||||
example.ParseFromString(string_record)
|
|
||||||
spec_n1 = example.features.feature['spec_n1'].int64_list.value[0]
|
|
||||||
spec_n2 = example.features.feature['spec_n2'].int64_list.value[0]
|
|
||||||
spec_w1 = example.features.feature['spec_w1'].int64_list.value[0]
|
|
||||||
spec_w2 = example.features.feature['spec_w2'].int64_list.value[0]
|
|
||||||
spec1 = np.array(example.features.feature['spec1'].float_list.value).reshape(spec_n1,spec_w1)
|
|
||||||
spec2 = np.array(example.features.feature['spec2'].float_list.value).reshape(spec_n2,spec_w2)
|
|
||||||
p_spec1,p_spec2 = padd_zeros(spec1,n_spec),padd_zeros(spec2,n_spec)
|
|
||||||
input_data[i] = np.asarray([spec1,spec2])
|
|
||||||
output = example.features.feature['output'].int64_list.value
|
|
||||||
output_data[i] = np.asarray(output)
|
|
||||||
# print('converting to nparray...')
|
|
||||||
# tr_pairs,te_pairs,tr_y,te_y = train_test_split(input_data,output_data,test_size=0.1)
|
|
||||||
# result = (tr_pairs,te_pairs,tr_y,te_y,n_spec,n_features)
|
|
||||||
return input_data,output_data
|
|
||||||
|
|
||||||
def read_siamese_tfrecords_generator(audio_group='audio',batch_size=32,sample_size=100):
|
def read_siamese_tfrecords_generator(audio_group='audio',batch_size=32,sample_size=100):
|
||||||
records_file = os.path.join('./outputs',audio_group+'.train.tfrecords')
|
records_file = os.path.join('./outputs',audio_group+'.train.tfrecords')
|
||||||
input_pairs = []
|
input_pairs = []
|
||||||
|
|
@ -273,120 +182,8 @@ def read_siamese_tfrecords_generator(audio_group='audio',batch_size=32,sample_si
|
||||||
|
|
||||||
return record_generator,input_data,output_data,n_spec,n_features,n_records
|
return record_generator,input_data,output_data,n_spec,n_features,n_records
|
||||||
|
|
||||||
def read_siamese_tfrecords_generator_old(audio_group='audio',batch_size=32):
|
|
||||||
records_file = os.path.join('./outputs',audio_group+'_padded.tfrecords')
|
|
||||||
input_pairs = []
|
|
||||||
output_class = []
|
|
||||||
const_file = os.path.join('./outputs',audio_group+'.constants')
|
|
||||||
(n_spec,n_features,n_records) = pickle.load(open(const_file,'rb'))
|
|
||||||
print('reading tfrecords({})...'.format(audio_group))
|
|
||||||
def record_generator():
|
|
||||||
input_data = []
|
|
||||||
output_data = []
|
|
||||||
while True:
|
|
||||||
record_iterator = tf.python_io.tf_record_iterator(path=records_file)
|
|
||||||
for (i,string_record) in tqdm(enumerate(record_iterator),total=n_records):
|
|
||||||
example = tf.train.Example()
|
|
||||||
example.ParseFromString(string_record)
|
|
||||||
spec1 = np.array(example.features.feature['spec1'].float_list.value).reshape(n_spec,n_features)
|
|
||||||
spec2 = np.array(example.features.feature['spec2'].float_list.value).reshape(n_spec,n_features)
|
|
||||||
input_data.append(np.asarray([spec1,spec2]))
|
|
||||||
output = example.features.feature['output'].int64_list.value
|
|
||||||
output_data.append(np.asarray(output))
|
|
||||||
if len(input_data) == batch_size:
|
|
||||||
input_arr = np.asarray(input_data)
|
|
||||||
output_arr = np.asarray(output_data)
|
|
||||||
yield ([input_arr[:, 0], input_arr[:, 1]],output_arr)
|
|
||||||
input_data = []
|
|
||||||
output_data = []
|
|
||||||
return record_generator,n_spec,n_features,n_records
|
|
||||||
|
|
||||||
def read_siamese_tfrecords(audio_group='audio'):
|
|
||||||
audio_group='story_words_test'
|
|
||||||
|
|
||||||
record_file = os.path.join('./outputs',audio_group+'_padded.tfrecords')
|
|
||||||
const_file = os.path.join('./outputs',audio_group+'.constants')
|
|
||||||
(n_spec,n_features) = pickle.load(open(const_file,'rb'))
|
|
||||||
|
|
||||||
filename_queue = tf.train.string_input_producer([record_file])
|
|
||||||
reader = tf.TFRecordReader()
|
|
||||||
_, serialized_example = reader.read(filename_queue)
|
|
||||||
features = tf.parse_single_example(serialized_example,
|
|
||||||
features={
|
|
||||||
'spec1': tf.FixedLenFeature([1,n_spec,n_features], tf.float32),
|
|
||||||
'spec2': tf.FixedLenFeature([1,n_spec,n_features], tf.float32),
|
|
||||||
'output':tf.FixedLenFeature([2], tf.int64)
|
|
||||||
})
|
|
||||||
spec1 = features['spec1']
|
|
||||||
spec1 = tf.cast(spec1, tf.float32) * (1. / 255)
|
|
||||||
spec2 = features['spec2']
|
|
||||||
spec2 = tf.cast(spec2, tf.float32) * (1. / 255)
|
|
||||||
output = tf.cast(features['output'], tf.int32)
|
|
||||||
return spec1,spec2, output,n_spec,n_features
|
|
||||||
|
|
||||||
def read_siamese_tfrecords_batch(audio_group='audio', batch_size=32):
|
|
||||||
audio_group='story_words_test'
|
|
||||||
record_file = os.path.join('./outputs',audio_group+'_padded.tfrecords')
|
|
||||||
""" Return tensor to read from TFRecord """
|
|
||||||
print('Creating graph for loading {} ...'.format(record_file))
|
|
||||||
const_file = os.path.join('./outputs',audio_group+'.constants')
|
|
||||||
(n_spec,n_features) = pickle.load(open(const_file,'rb'))
|
|
||||||
records_file = os.path.join('./outputs',audio_group+'.tfrecords')
|
|
||||||
record_iterator = tf.python_io.tf_record_iterator(path=records_file)
|
|
||||||
n_records = len([i for i in record_iterator])
|
|
||||||
batch_shape=[batch_size, n_spec, n_features]
|
|
||||||
with tf.variable_scope("SiameseTFRecords"):
|
|
||||||
record_input = data_flow_ops.RecordInput(record_file, batch_size=batch_size)
|
|
||||||
records_op = record_input.get_yield_op()
|
|
||||||
records_op = tf.split(records_op, batch_shape[0], 0)
|
|
||||||
records_op = [tf.reshape(record, []) for record in records_op]
|
|
||||||
specs1, specs2 = [],[]
|
|
||||||
outputs = []
|
|
||||||
for i, serialized_example in tqdm(enumerate(records_op)):
|
|
||||||
with tf.variable_scope("parse_siamese_pairs", reuse=True):
|
|
||||||
features = tf.parse_single_example(
|
|
||||||
serialized_example,
|
|
||||||
features={
|
|
||||||
'spec1': tf.FixedLenFeature([n_spec,n_features], tf.float32),
|
|
||||||
'spec2': tf.FixedLenFeature([n_spec,n_features], tf.float32),
|
|
||||||
'output':tf.FixedLenFeature([2], tf.int64)
|
|
||||||
})
|
|
||||||
spec1 = features['spec1']
|
|
||||||
spec1 = tf.cast(spec1, tf.float32) * (1. / 255)
|
|
||||||
spec2 = features['spec2']
|
|
||||||
output = tf.cast(spec2, tf.float32) * (1. / 255)
|
|
||||||
output = tf.cast(features['output'], tf.float32)
|
|
||||||
specs1.append(spec1)
|
|
||||||
specs2.append(spec2)
|
|
||||||
outputs.append(output)
|
|
||||||
|
|
||||||
specs1 = tf.parallel_stack(specs1, 0)
|
|
||||||
specs2 = tf.parallel_stack(specs2, 0)
|
|
||||||
outputs = tf.parallel_stack(outputs, 0)
|
|
||||||
specs1 = tf.cast(specs1, tf.float32)
|
|
||||||
specs2 = tf.cast(specs2, tf.float32)
|
|
||||||
|
|
||||||
specs1 = tf.reshape(specs1, shape=batch_shape)
|
|
||||||
specs2 = tf.reshape(specs1, shape=batch_shape)
|
|
||||||
specs1_shape = specs1.get_shape()
|
|
||||||
specs2_shape = specs2.get_shape()
|
|
||||||
outputs_shape = outputs.get_shape()
|
|
||||||
copy_stage = data_flow_ops.StagingArea(
|
|
||||||
[tf.float32, tf.float32, tf.float32],
|
|
||||||
shapes=[specs1_shape, specs2_shape, outputs_shape])
|
|
||||||
copy_stage_op = copy_stage.put(
|
|
||||||
[specs1, specs2, outputs])
|
|
||||||
staged_specs1, staged_specs2, staged_outputs = copy_stage.get()
|
|
||||||
return specs1, spec2, outputs,n_spec,n_features,n_records
|
|
||||||
|
|
||||||
def audio_samples_word_count(audio_group='audio'):
|
def audio_samples_word_count(audio_group='audio'):
|
||||||
audio_group = 'story_all'
|
audio_samples = pd.read_csv( './outputs/' + audio_group + '.csv')
|
||||||
audio_samples = pd.read_csv( './outputs/' + audio_group + '.csv'
|
|
||||||
, names=['word','phonemes', 'voice', 'language', 'rate', 'variant', 'file']
|
|
||||||
, quoting=csv.QUOTE_NONE)
|
|
||||||
audio_samples['file_path'] = audio_samples.loc[:, 'file'].apply(lambda x: 'outputs/' + audio_group + '/' + x)
|
|
||||||
audio_samples['file_exists'] = apply_by_multiprocessing(audio_samples['file_path'], os.path.exists)
|
|
||||||
audio_samples = audio_samples[audio_samples['file_exists'] == True].reset_index()
|
|
||||||
return len(audio_samples.groupby(audio_samples['word']))
|
return len(audio_samples.groupby(audio_samples['word']))
|
||||||
|
|
||||||
def fix_csv(audio_group='audio'):
|
def fix_csv(audio_group='audio'):
|
||||||
|
|
|
||||||
Loading…
Reference in New Issue