Added support for audio databases

A cleaner version of code for audio support

bob/learn/tensorflow/script/train_voicepa.py

-#!/usr/bin/env python
-# vim: set fileencoding=utf-8 :
-# @author: Pavel Korshunov <pavel.korshunov@idiap.ch>
-# @date: Wed 19 Oct 23:43:22 2016
-
-
-"""
-Simple script that trains voicePA using Tensor flow
-
-Usage:
-  train_voicepa.py [--batch-size=<arg> --validation-batch-size=<arg> --iterations=<arg> --validation-interval=<arg> --intermediate_model=<arg> --temp-dir=<arg> --snapshot=<arg> --use-gpu]
-  train_voicepa.py -h | --help
-Options:
-  -h --help     Show this screen.
-  --batch-size=<arg>  [default: 1]
-  --validation-batch-size=<arg>   [default: 128]
-  --iterations=<arg>  [default: 30000]
-  --validation-interval=<arg>  [default: 100]
-  --intermediate_model=<arg> [default: ""]
-  --temp-dir=<arg>  [default: "cnn"]
-  --snapshot=<arg>  [default: 1000]
-"""
-
-from docopt import docopt
-
-SEED = 1980
-
-from bob.learn.tensorflow.datashuffler import DiskAudio
-from bob.learn.tensorflow.network import SimpleAudio
-from bob.learn.tensorflow.loss import NegLogLoss
-from bob.learn.tensorflow.trainers import Trainer, constant
-
-import tensorflow as tf
-
-tf.logging.set_verbosity(tf.logging.ERROR)
-
-import os
-
-os.environ["CUDA_VISIBLE_DEVICES"] = "1,0,3,2"
-
-def main():
-    args = docopt(__doc__, version='voicePa training with TensorFlow')
-
-    BATCH_SIZE = int(args['--batch-size'])
-    VALIDATION_BATCH_SIZE = int(args['--validation-batch-size'])
-    ITERATIONS = int(args['--iterations'])
-    VALIDATION_INTERVAL = int(args['--validation-interval'])
-    USE_GPU = args['--use-gpu']
-    INTERMEDIATE_MODEL = str(args['--intermediate_model'])
-    TEMP_DIR = str(args['--temp-dir'])
-    TRAIN_SNAPSHOT = int(args['--snapshot'])
-
-    import bob.db.voicepa
-    db_voicepa = bob.db.voicepa.Database()
-    directory = "/idiap/temp/pkorshunov/data/voicePA/"
-
-    # Preparing train set
-    train_objects = db_voicepa.objects(protocol="avspoofPA", groups="train", cls=('real', 'attack'))
-
-    train_file_names = [o.make_path(directory=directory, extension=".wav") for o in train_objects]
-    train_labels = [1 if o.is_attack() else 0 for o in train_objects]
-    n_classes = len(set(train_labels))
-
-    train_data_shuffler = DiskAudio(train_file_names, train_labels,
-                                    batch_size=BATCH_SIZE,
-                                    seed=SEED,
-                                    out_file=os.path.join(TEMP_DIR, "train_batches.txt")
-                                    )
-
-    # Preparing validation set
-    validation_objects = db_voicepa.objects(protocol="avspoofPA", groups="dev", cls=('real', 'attack'))
-
-    validation_file_names = [o.make_path(directory=directory, extension=".wav") for o in validation_objects]
-    validation_labels = [1 if o.is_attack() else 0 for o in validation_objects]
-
-    validation_data_shuffler = DiskAudio(validation_file_names, validation_labels,
-                                         batch_size=VALIDATION_BATCH_SIZE,
-                                         seed=SEED,
-                                         out_file=os.path.join(TEMP_DIR, "validation_batches.txt")
-                                         )
-
-    conv1_kernel_size = 300
-    conv1_output = 20
-    conv1_stride = 100
-    fc1_output = 40
-#    fc1_output = 1
-    # Preparing the architecture
-
-    dnn = SimpleAudio(conv1_kernel_size=conv1_kernel_size, conv1_output=conv1_output, conv1_stride=conv1_stride,
-                      fc1_output=fc1_output,
-                      default_feature_layer='fc2',
-                      seed=SEED, use_gpu=USE_GPU)
-
-    loss = NegLogLoss(tf.reduce_mean)
-    learning_rate = 0.0001
-    trainer = Trainer(
-        architecture=dnn, loss=loss, iterations=ITERATIONS,
-        optimizer=tf.train.GradientDescentOptimizer(learning_rate),
-        learning_rate=constant(learning_rate, name="lr"),
-        temp_dir=TEMP_DIR,
-        snapshot=TRAIN_SNAPSHOT,
-        validation_snapshot=VALIDATION_INTERVAL,
-        prefetch=False,
-        verbosity_level=1,
-#        model_from_file=INTERMEDIATE_MODEL,
-    )
-    trainer.train(train_data_shuffler, validation_data_shuffler)