Merge branch 'master' of ssh://invnuc/~/Public/Repos/speech_scoring

2. fixed last batch carry-over

speech_data.py

@@ -161,7 +161,7 @@ def read_siamese_tfrecords_generator(audio_group='audio',batch_size=32,test_size
                 input_data.append(np.asarray([p_spec1,p_spec2]))
                 output = example.features.feature['output'].int64_list.value
                 output_data.append(np.asarray(output))
-                if len(input_data) == batch_size:
+                if len(input_data) == batch_size or i == n_records-1:
                     input_arr = np.asarray(input_data)
                     output_arr = np.asarray(output_data)
                     yield ([input_arr[:, 0], input_arr[:, 1]],output_arr)

speech_model.py

@@ -12,7 +12,7 @@ from keras.utils import to_categorical
 from keras.optimizers import RMSprop
 from keras.callbacks import TensorBoard, ModelCheckpoint
 from keras import backend as K
-from speech_tools import create_dir
+from speech_tools import create_dir,step_count
 
 # def euclidean_distance(vects):
 #     x, y = vects
@@ -96,7 +96,7 @@ def load_model_arch(mod_file):
 def train_siamese(audio_group = 'audio'):
     # the data, shuffled and split between train and test sets
     # tr_pairs, te_pairs, tr_y_e, te_y_e = speech_model_data()
-    batch_size = 128
+    batch_size = 256
     model_dir = './models/'+audio_group
     create_dir(model_dir)
     log_dir = './logs/'+audio_group
@@ -141,12 +141,14 @@ def train_siamese(audio_group = 'audio'):
     #     epochs=100,
     #     validation_data=([te_pairs[:, 0], te_pairs[:, 1]], te_y),
     #     callbacks=[tb_cb, cp_cb])
+    epoch_n_steps = step_count(n_records,batch_size)
     model.fit_generator(tr_gen
-                        ,epochs=1000
+                        , epochs=1000
-                        ,steps_per_epoch=n_records//batch_size
+                        , steps_per_epoch=epoch_n_steps
-                        ,validation_data=([te_pairs[:, 0], te_pairs[:, 1]], te_y)
+                        , validation_data=([te_pairs[:, 0], te_pairs[:, 1]], te_y)
-                        ,use_multiprocessing=True, workers=1
+                        # ,use_multiprocessing=True, workers=1
-                        ,callbacks=[tb_cb, cp_cb])
+                        , max_queue_size=32
+                        , callbacks=[tb_cb, cp_cb])
     model.save(model_dir+'/siamese_speech_model-final.h5')
     # compute final accuracy on training and test sets
     # y_pred = model.predict([tr_pairs[:, 0], tr_pairs[:, 1]])
@@ -160,5 +162,5 @@ def train_siamese(audio_group = 'audio'):
 
 
 if __name__ == '__main__':
-    train_siamese('story_words')
+    train_siamese('story_words_test')
     # train_siamese('audio')

speech_samplegen.py

@@ -5,16 +5,16 @@ from Foundation import NSURL
 import json
 import csv
 import random
+import string
 import os
 import re
 import subprocess
 import time
-import progressbar
+from tqdm import tqdm
 
 from generate_similar import similar_phoneme_phrase,similar_phrase
-from speech_tools import format_filename
 
-OUTPUT_NAME = 'story_words_test'
+OUTPUT_NAME = 'story_phrases'
 dest_dir = os.path.abspath('.') + '/outputs/' + OUTPUT_NAME + '/'
 dest_file = './outputs/' + OUTPUT_NAME + '.csv'
 
@@ -24,21 +24,25 @@ def hms_string(sec_elapsed):
     s = sec_elapsed % 60.
     return "{}:{:>02}:{:>05.2f}".format(h, m, s)
 
-def prog_bar(title):
-    widgets = [title, progressbar.Counter(), 'th entry - ', progressbar.FormatLabel(
-        ''), ' [', progressbar.Bar(), '] - ', progressbar.ETA()]
-    prog = progressbar.ProgressBar(widgets=widgets)
-
-    def update_prog(current):
-        widgets[3] = progressbar.FormatLabel(current)
-        prog.update()
-    return (update_prog, prog)
-
-
 def create_dir(direc):
     if not os.path.exists(direc):
         os.makedirs(direc)
 
+def format_filename(s):
+    """
+    Take a string and return a valid filename constructed from the string.
+    Uses a whitelist approach: any characters not present in valid_chars are
+    removed. Also spaces are replaced with underscores.
+
+    Note: this method may produce invalid filenames such as ``, `.` or `..`
+    When I use this method I prepend a date string like '2009_01_15_19_46_32_'
+    and append a file extension like '.txt', so I avoid the potential of using
+    an invalid filename.
+    """
+    valid_chars = "-_.() %s%s" % (string.ascii_letters, string.digits)
+    filename = ''.join(c for c in s if c in valid_chars)
+    filename = filename.replace(' ','_') # I don't like spaces in filenames.
+    return filename
 
 def dest_filename(w, v, r, t):
     rand_no = str(random.randint(0, 10000))
@@ -55,7 +59,7 @@ def dest_path(v, r, n):
 def cli_gen_audio(speech_cmd, rate, voice, out_path):
     subprocess.call(
         ['say', '-v', voice, '-r',
-         str(rate), '-o', out_path, speech_cmd])
+         str(rate), '-o', out_path, "'"+speech_cmd+"'"])
 
 
 class SynthFile(object):
@@ -173,7 +177,7 @@ class SynthVariant(object):
 
 def synth_generator():
     us_voices_ids = SynthVariant.voices_for_lang('en-US')
-    voice_rates = [150, 180, 210, 250]
+    voice_rates = [150, 180, 210]#, 250]
     voice_synths = []
     create_dir(dest_dir)
     for vp in us_voices_ids:
@@ -191,9 +195,10 @@ def synth_generator():
         for s in voice_synths:
             s.create_synth_dirs()
             for v in ['low', 'medium', 'high']:
-                (update, prog) = prog_bar(prog_title)
+                prog = tqdm(words)
-                for w in prog(words):
+                prog.set_postfix(variant=v,voice=s.name,rate=s.rate)
-                    update('"{}" with {} variant ({})'.format(w, s, v))
+                for w in tqdm(words):
+                    prog.set_postfix(word=w)
                     synthed = s.generate_audio(w, v)
                     writer(synthed)
         end_time = time.time()
@@ -201,7 +206,7 @@ def synth_generator():
         print("It took {} to synthsize all variants.".format(time_str))
     return synth_for_words
 
-def synth_logger(fname, csv=False):
+def synth_logger(fname, csv_mode=False):
     f = open(fname, 'w')
     s_csv_w = csv.writer(f, quoting=csv.QUOTE_MINIMAL)
     def csv_writer(s):
@@ -212,18 +217,18 @@ def synth_logger(fname, csv=False):
         synth_list.append(s)
 
     def close_file():
-        if csv:
+        if csv_mode:
             f.close()
         else:
             json.dump([s.get_json() for s in synth_list], f)
             f.close()
-    if csv:
+    if csv_mode:
         return csv_writer, close_file
     else:
         return json_writer, close_file
 
 def generate_audio_for_text_list(text_list):
-    (writer, closer) = synth_logger(dest_file, csv=True)
+    (writer, closer) = synth_logger(dest_file, csv_mode=True)
     synth_for_texts = synth_generator()
     try:
         synth_for_texts(text_list, writer)
@@ -239,7 +244,7 @@ def generate_audio_for_stories():
     story_file = './inputs/all_stories.json'
     stories_data = json.load(open(story_file))
     # text_list = [t[0] for i in stories_data.values() for t in i]
-    text_list = [i.replace('-','') for g in stories_data.values() for i in g]
+    text_list = [i for g in stories_data.values() for i in g]
     generate_audio_for_text_list(text_list)
 
 def generate_test_audio_for_stories():
@@ -257,6 +262,6 @@ def generate_test_audio_for_stories():
 
 
 if __name__ == '__main__':
-    generate_test_audio_for_stories()
+    # generate_test_audio_for_stories()
     # generate_audio_for_text_list(['I want to go home','education'])
-    # generate_audio_for_stories()
+    generate_audio_for_stories()

speech_tools.py

@@ -1,4 +1,5 @@
 import os
+import math
 import threading
 import multiprocessing
 import pandas as pd
@@ -11,6 +12,9 @@ from speech_spectrum import plot_stft, generate_spec_frec
 SAMPLE_RATE = 22050
 N_CHANNELS = 2
 
+def step_count(n_records,batch_size):
+    return int(math.ceil(n_records*1.0/batch_size))
+
 def file_player():
     p_oup = pyaudio.PyAudio()
     def play_file(audiopath,plot=False):
@@ -132,21 +136,3 @@ def threadsafe_generator(f):
     def g(*a, **kw):
         return threadsafe_iter(f(*a, **kw))
     return g
-
-
-
-def format_filename(s):
-    """
-    Take a string and return a valid filename constructed from the string.
-    Uses a whitelist approach: any characters not present in valid_chars are
-    removed. Also spaces are replaced with underscores.
-
-    Note: this method may produce invalid filenames such as ``, `.` or `..`
-    When I use this method I prepend a date string like '2009_01_15_19_46_32_'
-    and append a file extension like '.txt', so I avoid the potential of using
-    an invalid filename.
-    """
-    valid_chars = "-_.() %s%s" % (string.ascii_letters, string.digits)
-    filename = ''.join(c for c in s if c in valid_chars)
-    filename = filename.replace(' ','_') # I don't like spaces in filenames.
-    return filename