tfrecords wip

speech_data.py

@@ -6,11 +6,11 @@ import numpy as np
 from spectro_gen import generate_aiff_spectrogram
 from sklearn.model_selection import train_test_split
 import itertools
+import os
 import random
 import csv
 import gc
 
-
 def get_siamese_pairs(groupF1, groupF2):
     group1 = [r for (i, r) in groupF1.iterrows()]
     group2 = [r for (i, r) in groupF2.iterrows()]
@@ -83,12 +83,46 @@ def create_spectrogram_data(audio_group='audio'):
         , quoting=csv.QUOTE_NONE)
     # audio_samples = audio_samples.loc[audio_samples['word'] ==
     #                                   'sunflowers'].reset_index(drop=True)
-    file_names = audio_samples.loc[:, 'file'].apply(lambda x: 'outputs/' + audio_group + '/' + x)
-    audio_samples['spectrogram'] = apply_by_multiprocessing(file_names,generate_aiff_spectrogram)#.apply(
-            #generate_aiff_spectrogram)
+    audio_samples['file_paths'] = audio_samples.loc[:, 'file'].apply(lambda x: 'outputs/' + audio_group + '/' + x)
+    audio_samples['file_exists'] = apply_by_multiprocessing(audio_samples['file_paths'], os.path.exists)
+    audio_samples = audio_samples[audio_samples['file_exists'] == True].reset_index()
+    audio_samples['spectrogram'] = apply_by_multiprocessing(audio_samples['file_paths'],generate_aiff_spectrogram)#.apply(
     audio_samples['window_count'] = audio_samples.loc[:,'spectrogram'].apply(lambda x: x.shape[0])
     audio_samples.to_pickle('outputs/{}-spectrogram.pkl'.format(audio_group))
 
+def create_spectrogram_tfrecords(audio_group='audio'):
+    audio_samples = pd.read_csv( './outputs/' + audio_group + '.csv'
+        , names=['word','phonemes', 'voice', 'language', 'rate', 'variant', 'file']
+        , quoting=csv.QUOTE_NONE)
+    # audio_samples = audio_samples.loc[audio_samples['word'] ==
+    #                                   'sunflowers'].reset_index(drop=True)
+    audio_samples['file_paths'] = audio_samples.loc[:, 'file'].apply(lambda x: 'outputs/' + audio_group + '/' + x)
+    audio_samples['file_exists'] = apply_by_multiprocessing(audio_samples['file_paths'], os.path.exists)
+    audio_samples = audio_samples[audio_samples['file_exists'] == True].reset_index()
+    # audio_samples['spectrogram'] = apply_by_multiprocessing(audio_samples['file_paths'],generate_aiff_spectrogram)#.apply(
+    # audio_samples['window_count'] = audio_samples.loc[:,'spectrogram'].apply(lambda x: x.shape[0])
+    # audio_samples.to_pickle('outputs/{}-spectrogram.pkl'.format(audio_group))
+
+    def _float_feature(value):
+      return tf.train.Feature(float_list=tf.train.FloatList(value=value))
+
+    def _int64_feature(value):
+      return tf.train.Feature(int64_list=tf.train.Int64List(value=value))
+
+    def _bytes_feature(value):
+      return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))
+    writer = tf.python_io.TFRecordWriter(output_path)
+
+    for sample in audio_samples:
+        example = tf.train.Example(features=tf.train.Features(
+          feature={
+            'label': _int64_feature([label]),
+            'path': _bytes_feature([image_path]),
+            'instance' : _bytes_feature([instance_id])
+          }
+        ))
+    writer.write(example.SerializeToString())
+    writer.close()
 
 def create_tagged_data(audio_samples):
     same_data, diff_data = [], []
@@ -105,28 +139,8 @@ def create_tagged_data(audio_samples):
     Y = to_onehot(Y_f.astype(np.int8))
     print('casting as array speech pairs')
     X = np.asarray(same_data + diff_data)
-    # del same_data
-    # del diff_data
-    # gc.collect()
     return X,Y
 
-# def create_padded_spectrogram(audio_group='audio'):
-#     audio_samples = pd.read_pickle('outputs/{}-spectrogram.pkl'.format(audio_group))
-#     daf_audio_samples = ddf.from_pandas(audio_samples,npartitions=4)
-#     def spec_size(s):
-#         return s['spectrogram'].shape[0]
-#     max_samples = daf_audio_samples.apply(spec_size,axis=1, meta=('x', 'i8')).max().compute()
-#     print('max sample count is ',max_samples)
-#     def padd_zeros_fixed(sp):
-#         x = sp['spectrogram']
-#         bounds = [(0, max_samples - x.shape[0]), (0, 0)]
-#         sp['spectrogram'] = np.lib.pad(x,bounds,'constant')
-#         return sp
-#     padded_audio_samples = daf_audio_samples.apply(padd_zeros_fixed,axis=1, meta=audio_samples).compute()#,max_samples=max_samples)
-#     import pdb; pdb.set_trace()
-#     padded_spectrogram = np.asarray(padded_audio_samples['spectrogram'].tolist())
-#     np.save('outputs/{}-padded_spectrogram.npy'.format(audio_group),padded_spectrogram)
-
 def create_speech_pairs_data(audio_group='audio'):
     audio_samples = pd.read_pickle('outputs/{}-spectrogram.pkl'.format(audio_group))
     # sample_size = audio_samples['spectrogram'][0].shape[1]
@@ -139,26 +153,11 @@ def create_speech_pairs_data(audio_group='audio'):
         np.random.shuffle(X)
         np.random.set_state(rng_state)
         np.random.shuffle(Y)
-        # p = np.random.permutation(len(X))
-        # X = X_i[p]
-        # Y = Y_i[p]
         print('pickling X/Y')
         np.save('outputs/{}-train-X.npy'.format(audio_group), X)
         np.save('outputs/{}-train-Y.npy'.format(audio_group), Y)
     save_samples_for('train',tr_audio_samples)
     save_samples_for('test',te_audio_samples)
-    # print('generating test siamese speech pairs ')
-    # X,Y = create_tagged_data(te_audio_samples,max_samples)
-    # print('pickling X/Y')
-    # np.save('outputs/{}-test-X.npy'.format(audio_group), X)
-    # np.save('outputs/{}-test-Y.npy'.format(audio_group), Y)
-    # # print('train/test splitting speech pairs')
-    # # tr_pairs, te_pairs, tr_y, te_y = train_test_split(X, Y, test_size=0.1)
-    # # print('pickling train/test')
-    # # np.save('outputs/{}-tr_pairs.npy'.format(audio_group), tr_pairs)
-    # # np.save('outputs/{}-te_pairs.npy'.format(audio_group), te_pairs)
-    # # np.save('outputs/{}-tr_y.npy'.format(audio_group), tr_y)
-    # # np.save('outputs/{}-te_y.npy'.format(audio_group), te_y)
 
 def speech_data(audio_group='audio'):
     X = np.load('outputs/{}-X.npy'.format(audio_group)) / 255.0