pickling intermediate data to save memory usage

.gitignore

@@ -140,3 +140,4 @@ Temporary Items
 outputs/*
 inputs/mnist
 inputs/audio*
+*.pkl

speech_data.py

@@ -2,7 +2,7 @@ import pandas as pd
 import numpy as np
 from spectro_gen import generate_aiff_spectrogram
 from sklearn.model_selection import train_test_split
-import tensorflow as tf
+import pickle,gc
 
 def sunflower_data():
     audio_samples = pd.read_csv('./outputs/audio.csv',names=['word','voice','rate','variant','file'])
@@ -62,11 +62,12 @@ def create_spectrogram_data(audio_group='audio'):
     audio_samples.loc[:,'spectrogram'] = audio_samples.loc[:,'file'].apply(lambda x:'outputs/'+audio_group+'/'+x).apply(generate_aiff_spectrogram)
     audio_samples.to_pickle('outputs/spectrogram.pkl')
 
-def speech_pairs_data(audio_group='audio'):
+def create_speech_pairs_data(audio_group='audio'):
     audio_samples = pd.read_pickle('outputs/spectrogram.pkl')
     y_data = audio_samples['variant'].apply(lambda x:x=='normal').values
     max_samples = audio_samples['spectrogram'].apply(lambda x:x.shape[0]).max()
     sample_size = audio_samples['spectrogram'][0].shape[1]
+    pickle.dump((max_samples,sample_size),open('./spectrogram_vars.pkl','wb'))
     audio_samples_pos = audio_samples[audio_samples['variant'] == 'normal'].reset_index(drop=True)
     audio_samples_neg = audio_samples[audio_samples['variant'] == 'phoneme'].reset_index(drop=True)
     def append_zeros(spgr):
@@ -74,17 +75,47 @@ def speech_pairs_data(audio_group='audio'):
     def create_data(sf):
         sample_count = sf['spectrogram'].shape[0]
         pad_sun = sf['spectrogram'].apply(append_zeros).values
+        print('appended zeros')
         x_data = np.vstack(pad_sun).reshape((sample_count,max_samples,sample_size))
+        print('reshaped')
         return x_data
+    print('creating speech pair data')
     x_data_pos = create_data(audio_samples_pos)
     x_data_neg = create_data(audio_samples_neg)
+    np.save('outputs/x_data_pos.npy',x_data_pos)
+    np.save('outputs/x_data_neg.npy',x_data_neg)
+    print('pickled speech pairs')
+
+def create_speech_model_data():
+    (max_samples,sample_size) = pickle.load(open('./spectrogram_vars.pkl','rb'))
+    x_data_pos = np.load('outputs/x_data_pos.npy')
+    x_data_neg = np.load('outputs/x_data_neg.npy')
     x_pos_train, x_pos_test, x_neg_train, x_neg_test =train_test_split(x_data_pos,x_data_neg,test_size=0.33)
+    del x_data_pos
+    del x_data_neg
+    gc.collect()
+    print('split train and test')
     tr_y = np.array(x_pos_train.shape[0]*[[1,0]])
     te_y = np.array(x_pos_test.shape[0]*[[1,0]])
     tr_pairs = np.array([x_pos_train,x_neg_train]).reshape(x_pos_train.shape[0],2,max_samples,sample_size)
     te_pairs = np.array([x_pos_test,x_neg_test]).reshape(x_pos_test.shape[0],2,max_samples,sample_size)
+    print('reshaped to input dim')
+    np.save('outputs/tr_pairs.npy',tr_pairs)
+    np.save('outputs/te_pairs.npy',te_pairs)
+    np.save('outputs/tr_y.npy',tr_y)
+    np.save('outputs/te_y.npy',te_y)
+    print('pickled speech model data')
+    # return tr_pairs,te_pairs,tr_y,te_y
+
+def speech_model_data():
+    tr_pairs = np.load('outputs/tr_pairs.npy')
+    te_pairs = np.load('outputs/te_pairs.npy')
+    tr_y = np.load('outputs/tr_y.npy')
+    te_y = np.load('outputs/te_y.npy')
     return tr_pairs,te_pairs,tr_y,te_y
 
 if __name__ == '__main__':
-    create_spectrogram_data()
-    print(speech_pairs_data())
+    #create_spectrogram_data()
+    # create_speech_pairs_data()
+    # create_speech_model_data()
+    print(speech_model_data())

speech_siamese.py

@@ -16,7 +16,7 @@ import numpy as np
 
 import random
 # from keras.datasets import mnist
-from speech_data import sunflower_pairs_data
+from speech_data import speech_model_data
 from keras.models import Model
 from keras.layers import Dense, Dropout, Input, Lambda, LSTM, SimpleRNN
 from keras.optimizers import RMSprop
@@ -66,7 +66,7 @@ def accuracy(y_true, y_pred):
 
 
 # the data, shuffled and split between train and test sets
-tr_pairs,te_pairs,tr_y,te_y = sunflower_pairs_data()
+tr_pairs,te_pairs,tr_y,te_y = speech_model_data()
  # y_train.shape,y_test.shape
 # x_train.shape,x_test.shape
 # x_train = x_train.reshape(60000, 784)