implemented siamese pair tfrecord writer

speech_data.py

@@ -1,7 +1,8 @@
 import pandas as pd
 from pandas_parallel import apply_by_multiprocessing
-import dask as dd
-import dask.dataframe as ddf
+# import dask as dd
+# import dask.dataframe as ddf
+import tensorflow as tf
 import numpy as np
 from spectro_gen import generate_aiff_spectrogram
 from sklearn.model_selection import train_test_split
@@ -22,6 +23,15 @@ def get_siamese_pairs(groupF1, groupF2):
     # return (random.sample(same,10), random.sample(diff,10))
     return same[:10],diff[:10]
 
+def siamese_pairs(rightGroup, wrongGroup):
+    group1 = [r for (i, r) in rightGroup.iterrows()]
+    group2 = [r for (i, r) in wrongGroup.iterrows()]
+    rightWrongPairs = [(g1, g2) for g2 in group2 for g1 in group1]
+    rightRightPairs = [i for i in itertools.combinations(group1, 2)]
+    random.shuffle(rightWrongPairs)
+    random.shuffle(rightRightPairs)
+    # return (random.sample(same,10), random.sample(diff,10))
+    return rightRightPairs[:10],rightWrongPairs[:10]
 
 def append_zeros(spgr, max_samples):
     return np.lib.pad(spgr, [(0, max_samples - spgr.shape[0]), (0, 0)],
@@ -96,12 +106,9 @@ def create_spectrogram_tfrecords(audio_group='audio'):
         , 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['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()
-    # 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))
@@ -111,17 +118,46 @@ def create_spectrogram_tfrecords(audio_group='audio'):
 
     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 = tf.python_io.TFRecordWriter('./outputs/' + audio_group + '.tfrecords')
+    # audio_samples = audio_samples[:100]
+    for (w, word_group) in audio_samples.groupby(audio_samples['word']):
+        g = word_group.reset_index()
+        g['spectrogram'] = apply_by_multiprocessing(g['file_path'],generate_aiff_spectrogram)
+        sample_right = g.loc[audio_samples['variant'] == 'low']
+        sample_wrong = g.loc[audio_samples['variant'] == 'medium']
+        same, diff = siamese_pairs(sample_right, sample_wrong)
+        groups = [([0,1],same),([1,0],diff)]
+        for (output,group) in groups:
+            for sample1,sample2 in group:
+                spectro1,spectro2 = sample1['spectrogram'],sample2['spectrogram']
+                spec_n1,spec_n2 = spectro1.shape[0],spectro2.shape[0]
+                spec_w1,spec_w2 = spectro1.shape[1],spectro2.shape[1]
+                spec1,spec2 = spectro1.reshape(-1),spectro2.reshape(-1)
+                example = tf.train.Example(features=tf.train.Features(
+                  feature={
+                    'word': _bytes_feature([w.encode('utf-8')]),
+                    'phoneme1': _bytes_feature([sample1['phonemes'].encode('utf-8')]),
+                    'phoneme2': _bytes_feature([sample2['phonemes'].encode('utf-8')]),
+                    'voice1': _bytes_feature([sample1['voice'].encode('utf-8')]),
+                    'voice2': _bytes_feature([sample2['voice'].encode('utf-8')]),
+                    'language': _bytes_feature([sample1['language'].encode('utf-8')]),
+                    'rate1':_int64_feature([sample1['rate']]),
+                    'rate2':_int64_feature([sample2['rate']]),
+                    'variant1': _bytes_feature([sample1['variant'].encode('utf-8')]),
+                    'variant2': _bytes_feature([sample2['variant'].encode('utf-8')]),
+                    'file1': _bytes_feature([sample1['file'].encode('utf-8')]),
+                    'file2': _bytes_feature([sample2['file'].encode('utf-8')]),
+                    'spec1':_float_feature(spec1),
+                    'spec2':_float_feature(spec2),
+                    'spec_n1':_int64_feature([spec_n1]),
+                    'spec_w1':_int64_feature([spec_w1]),
+                    'spec_n2':_int64_feature([spec_n2]),
+                    'spec_w2':_int64_feature([spec_w2]),
+                    'output':_int64_feature(output)
+                  }
+                ))
+                writer.write(example.SerializeToString())
     writer.close()
 
 def create_tagged_data(audio_samples):
@@ -177,7 +213,8 @@ def speech_model_data():
 if __name__ == '__main__':
     # sunflower_pairs_data()
     # create_spectrogram_data()
-    create_spectrogram_data('story_words')
+    # create_spectrogram_data('story_words')
+    create_spectrogram_tfrecords('story_words')
     # create_padded_spectrogram()
     # create_speech_pairs_data()
     # print(speech_model_data())