def model_fn(features, labels, mode, params):
x = tf.reshape(features, [-1, 99, 161, 1], name='input_incep4')
x_norm = tf.layers.batch_normalization(x, training=mode == tf.estimator.ModeKeys.TRAIN, name='x_norm')
if params['verbose_summary']:
tf.summary.image('input', x)
conv1 = tf.layers.conv2d(x_norm, filters=16, kernel_size=3, padding='same', activation=tf.nn.relu, name='conv1')
conv1b = tf.layers.conv2d(conv1, filters=16, kernel_size=3, activation=tf.nn.relu, name='conv1b')
pool1 = tf.layers.max_pooling2d(conv1b, pool_size=[2, 2], strides=2, name='pool1')
if params['verbose_summary']:
log_conv_kernel('conv1')
log_conv_kernel('conv1b')
tf.summary.image('pool1', pool1[:, :, :, 0:1])
incep2 = inception_block(conv1, name='incep2')
incep3 = inception_block(incep2, t1x1=4, t3x3=4, t5x5=4, tmp=4, name='incep3')
incep4 = inception_block(incep3, t1x1=8, t3x3=8, t5x5=8, tmp=8, name='incep4')
incep5 = inception_block(incep4, t1x1=16, t3x3=16, t5x5=16, tmp=16, name='incep5')
incep6 = inception_block(incep5, t1x1=20, t3x3=20, t5x5=20, tmp=20, name='incep6')
flat = flatten(incep6)
dropout4 = tf.layers.dropout(flat, rate=params['dropout_rate'], training=mode == tf.estimator.ModeKeys.TRAIN, name='dropout4')
dense4 = tf.layers.dense(dropout4, units=2048, activation=tf.nn.relu, name='dense4')
logits = tf.layers.dense(dense4, units=params['num_classes'], name='logits')
predictions = {
'classes': tf.argmax(logits, axis=1, name='prediction_classes'),
'probabilities': tf.nn.softmax(logits, name='prediction_softmax')
}
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions={'predictions': predictions['probabilities']})
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=params['num_classes'], name='onehot_labels')
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels, logits=logits)
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=params['learning_rate'])
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss=loss, global_step=tf.train.get_global_step())
eval_metric_ops = {
'accuracy': tf.metrics.accuracy(labels=labels, predictions=predictions['classes'])
}
tf.summary.scalar('accuracy', eval_metric_ops['accuracy'][1])
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops
)
def model_fn(features, labels, mode, params):
x = tf.reshape(features, [-1, 125, 161, 2], name='input_incep100_lg')
x_norm = tf.layers.batch_normalization(x, training=mode == tf.estimator.ModeKeys.TRAIN, name='x_norm')
if params['verbose_summary']:
with tf.variable_scope('input'):
for sample in range(3):
tf.summary.image('input_spec_{}'.format(sample), tf.reshape(x[sample, :, :, 0], [-1, 125, 161, 1]))
tf.summary.image('input_freq_{}'.format(sample), tf.reshape(x[sample, :, :, 1], [-1, 125, 161, 1]))
conv1 = tf.layers.conv2d(x_norm, filters=16, kernel_size=3, padding='same', activation=tf.nn.relu, name='conv1')
conv1b = tf.layers.conv2d(conv1, filters=16, kernel_size=3, activation=tf.nn.relu, name='conv1b')
pool1 = tf.layers.max_pooling2d(conv1b, pool_size=[2, 2], strides=2, name='pool1')
if params['verbose_summary']:
log_conv_kernel('conv1')
log_conv_kernel('conv1b')
tf.summary.image('pool1', pool1[:, :, :, 0:1])
incep2 = inception_block(pool1, t1x1=8, t3x3=8, t5x5=8, tmp=8, name='incep2')
conv3 = tf.layers.conv2d(incep2, filters=32, kernel_size=3, padding='same', activation=tf.nn.relu, name='conv3')
conv3b = tf.layers.conv2d(conv3, filters=32, kernel_size=3, activation=tf.nn.relu, name='conv3b')
pool3 = tf.layers.max_pooling2d(conv3b, pool_size=[2, 2], strides=2, name='pool3')
if params['verbose_summary']:
log_conv_kernel('conv3')
log_conv_kernel('conv3b')
tf.summary.image('pool3', pool3[:, :, :, 0:1])
conv5 = tf.layers.conv2d(pool3, filters=64, kernel_size=3, padding='same', activation=tf.nn.relu, name='conv5')
conv5b = tf.layers.conv2d(conv5, filters=64, kernel_size=3, activation=tf.nn.relu, name='conv5b')
pool5 = tf.layers.max_pooling2d(conv5b, pool_size=[2, 2], strides=2, name='pool5')
if params['verbose_summary']:
log_conv_kernel('conv5')
log_conv_kernel('conv5b')
tf.summary.image('pool5', pool5[:, :, :, 0:1])
incep6 = inception_block(pool5, t1x1=32, t3x3=32, t5x5=32, tmp=32, name='incep6')
conv7 = tf.layers.conv2d(incep6, filters=128, kernel_size=3, padding='same', activation=tf.nn.relu, name='conv7')
conv7b = tf.layers.conv2d(conv7, filters=128, kernel_size=3, activation=tf.nn.relu, name='conv7b')
pool7 = tf.layers.max_pooling2d(conv7b, pool_size=[2, 2], strides=2, name='pool7')
if params['verbose_summary']:
log_conv_kernel('conv7')
log_conv_kernel('conv7b')
tf.summary.image('pool7', pool7[:, :, :, 0:1])
conv8 = tf.layers.conv2d(pool7, filters=256, kernel_size=3, padding='same', activation=tf.nn.relu, name='conv8')
conv8b = tf.layers.conv2d(conv8, filters=256, kernel_size=3, activation=tf.nn.relu, name='conv8b')
pool8 = tf.layers.max_pooling2d(conv8b, pool_size=[2, 2], strides=2, name='pool8')
if params['verbose_summary']:
log_conv_kernel('conv8')
log_conv_kernel('conv8b')
tf.summary.image('pool8', pool8[:, :, :, 0:1])
flat = flatten(pool8)
dense4 = tf.layers.dense(flat, units=2048, activation=tf.nn.relu, name='dense4')
dropout4 = tf.layers.dropout(dense4, rate=params['dropout_rate'], training=mode == tf.estimator.ModeKeys.TRAIN, name='dropout4')
logits = tf.layers.dense(dropout4, units=params['num_classes'], name='logits')
predictions = {
'classes': tf.argmax(logits, axis=1, name='prediction_classes'),
'probabilities': tf.nn.softmax(logits, name='prediction_softmax')
}
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions={'predictions': predictions['probabilities']})
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=params['num_classes'], name='onehot_labels')
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels, logits=logits)
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=params['learning_rate'])
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss=loss, global_step=tf.train.get_global_step())
eval_metric_ops = {
'accuracy': tf.metrics.accuracy(labels=labels, predictions=predictions['classes'])
}
tf.summary.scalar('accuracy', eval_metric_ops['accuracy'][1])
return tf.estimator.EstimatorSpec(
mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops
)
def model_fn(features, labels, mode, params):
with tf.variable_scope('x_prep'):
x = tf.reshape(features, [-1, 125, 161, 2], name='deception1')
x_norm = tf.layers.batch_normalization(
x, training=mode == tf.estimator.ModeKeys.TRAIN, name='x_norm')
x = tf.reshape(x_norm[:, :, :, 0], [-1, 125, 161, 1],
name='reshape_spec')
if params['verbose_summary']:
tf.summary.image('input', x)
conv = tf.layers.conv2d(x,
filters=32,
kernel_size=11,
activation=tf.nn.relu,
name='conv1')
conv = tf.layers.conv2d(conv,
filters=64,
kernel_size=3,
activation=tf.nn.relu,
name='conv3')
conv = tf.layers.conv2d(conv,
filters=128,
kernel_size=3,
activation=tf.nn.relu,
name='conv4')
conv = tf.layers.conv2d(conv,
filters=256,
kernel_size=3,
activation=tf.nn.relu,
name='conv5')
incep = inception_block(conv,
t1x1=32,
t3x3=32,
t5x5=32,
tmp=32,
name='incep6',
norm=True,
mode=mode)
incep = inception_block(incep,
t1x1=64,
t3x3=64,
t5x5=64,
tmp=64,
name='incep7',
norm=True,
mode=mode)
pool = tf.layers.max_pooling2d(incep,
pool_size=[2, 2],
strides=2,
name='pool7')
incep = inception_block(pool,
t1x1=64,
t3x3=64,
t5x5=64,
tmp=64,
name='incep8',
norm=True,
mode=mode)
incep = inception_block(incep,
t1x1=64,
t3x3=64,
t5x5=64,
tmp=64,
name='incep9',
norm=True,
mode=mode)
pool = tf.layers.max_pooling2d(incep,
pool_size=[2, 2],
strides=2,
name='pool9')
incep = inception_block(pool,
t1x1=128,
t3x3=128,
t5x5=128,
tmp=128,
name='incep10',
norm=True,
mode=mode)
pool = tf.layers.max_pooling2d(incep,
pool_size=[2, 2],
strides=2,
name='pool11')
flat = flatten(pool, name='flatten')
dropout = tf.layers.dropout(flat,
rate=params['dropout_rate'],
training=mode == tf.estimator.ModeKeys.TRAIN,
name='dropout')
dense = tf.layers.dense(dropout,
units=2048,
activation=tf.nn.relu,
name='dense')
logits = tf.layers.dense(dense, units=params['num_classes'], name='logits')
predictions = {
'classes': tf.argmax(logits, axis=1, name='prediction_classes'),
'probabilities': tf.nn.softmax(logits, name='prediction_softmax')
}
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions={'predictions': predictions['probabilities']})
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32),
depth=params['num_classes'],
name='onehot_labels')
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=params['learning_rate'])
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
eval_metric_ops = {
'accuracy':
tf.metrics.accuracy(labels=labels, predictions=predictions['classes'])
}
tf.summary.scalar('accuracy', eval_metric_ops['accuracy'][1])
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def branch_incep(x, name, mode, params):
with tf.variable_scope(name):
conv1 = tf.layers.conv2d(x,
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv1')
conv1b = tf.layers.conv2d(conv1,
filters=16,
kernel_size=3,
activation=tf.nn.relu,
name='conv1b')
pool1 = tf.layers.max_pooling2d(conv1b,
pool_size=[2, 2],
strides=2,
name='pool1')
if params['verbose_summary']:
log_conv_kernel('{}/conv1'.format(name))
log_conv_kernel('{}/conv1b'.format(name))
tf.summary.image('pool1', pool1[:, :, :, 0:1])
incep2 = inception_block(pool1,
t1x1=8,
t3x3=8,
t5x5=8,
tmp=8,
name='incep2')
conv3 = tf.layers.conv2d(incep2,
filters=32,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv3')
conv3b = tf.layers.conv2d(conv3,
filters=32,
kernel_size=3,
activation=tf.nn.relu,
name='conv3b')
pool3 = tf.layers.max_pooling2d(conv3b,
pool_size=[2, 2],
strides=2,
name='pool3')
if params['verbose_summary']:
log_conv_kernel('{}/conv3'.format(name))
log_conv_kernel('{}/conv3b'.format(name))
tf.summary.image('pool3', pool3[:, :, :, 0:1])
conv5 = tf.layers.conv2d(pool3,
filters=64,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv5')
conv5b = tf.layers.conv2d(conv5,
filters=64,
kernel_size=3,
activation=tf.nn.relu,
name='conv5b')
pool5 = tf.layers.max_pooling2d(conv5b,
pool_size=[2, 2],
strides=2,
name='pool5')
if params['verbose_summary']:
log_conv_kernel('{}/conv5'.format(name))
log_conv_kernel('{}/conv5b'.format(name))
tf.summary.image('pool5', pool5[:, :, :, 0:1])
incep6 = inception_block(pool5,
t1x1=32,
t3x3=32,
t5x5=32,
tmp=32,
name='incep6')
conv7 = tf.layers.conv2d(incep6,
filters=128,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv7')
conv7b = tf.layers.conv2d(conv7,
filters=128,
kernel_size=3,
activation=tf.nn.relu,
name='conv7b')
pool7 = tf.layers.max_pooling2d(conv7b,
pool_size=[2, 2],
strides=2,
name='pool7')
if params['verbose_summary']:
log_conv_kernel('{}/conv7'.format(name))
log_conv_kernel('{}/conv7b'.format(name))
tf.summary.image('pool7', pool7[:, :, :, 0:1])
conv8 = tf.layers.conv2d(pool7,
filters=256,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv8')
conv8b = tf.layers.conv2d(conv8,
filters=256,
kernel_size=3,
activation=tf.nn.relu,
name='conv8b')
pool8 = tf.layers.max_pooling2d(conv8b,
pool_size=[2, 2],
strides=2,
name='pool8')
if params['verbose_summary']:
log_conv_kernel('{}/conv8'.format(name))
log_conv_kernel('{}/conv8b'.format(name))
tf.summary.image('pool8', pool8[:, :, :, 0:1])
return pool8
def model_fn(features, labels, mode, params):
training = mode == tf.estimator.ModeKeys.TRAIN
# x = tf.layers.batch_normalization(features, training=training, name='x_norm__simple_inception')
tf.summary.image('0_input', features)
x = tf.layers.conv2d(features,
filters=16,
kernel_size=7,
activation=None,
name='l1')
x = tf.layers.batch_normalization(x, training=training, name='batch_norm1')
x = tf.nn.relu(x, name='act1')
tf.summary.image('1_1l1', x[:, :, :, 0:3], max_outputs=4)
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, name='maxpool1')
tf.summary.image('1_2maxpool', x[:, :, :, 0:3], max_outputs=4)
x = tf.layers.conv2d(x,
filters=32,
kernel_size=3,
activation=None,
name='l2')
x = tf.layers.batch_normalization(x, training=training, name='batch_norm2')
x = tf.nn.relu(x, name='act2')
tf.summary.image('2_1l2', x[:, :, :, 0:3], max_outputs=4)
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, name='maxpool2')
tf.summary.image('2_2maxpool', x[:, :, :, 0:3], max_outputs=4)
x = util.inception_block(x,
t1x1=32,
t3x3=42,
t5x5=16,
tmp=16,
training=training,
name='3_incep_1')
x = util.inception_block(x,
t1x1=64,
t3x3=128,
t5x5=52,
tmp=43,
training=training,
name='4_incep_2')
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, name='maxpool3')
tf.summary.image('3_1maxpool', x[:, :, :, 0:3], max_outputs=4)
x = util.inception_block(x,
t1x1=96,
t3x3=128,
t5x5=16,
tmp=16,
training=training,
name='6_incep_3')
x = util.inception_block(x,
t1x1=96,
t3x3=128,
t5x5=16,
tmp=16,
training=training,
name='7_incep_4')
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, name='maxpool4')
tf.summary.image('4_1maxpool', x[:, :, :, 0:3], max_outputs=4)
# x = util.inception_block(x, t1x1=96, t3x3=128, t5x5=16, tmp=16, training=training, name='8_incep_5')
# x = util.inception_block(x, t1x1=96, t3x3=128, t5x5=16, tmp=16, training=training, name='9_incep_6')
# x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, name='maxpool5')
# tf.summary.image('9_maxpool', x[:, :, :, 0:3], max_outputs=4)
x = util.inception_block(x,
t1x1=128,
t3x3=256,
t5x5=32,
tmp=16,
training=training,
name='10_incep_7')
x = util.inception_block(x,
t1x1=128,
t3x3=256,
t5x5=32,
tmp=16,
training=training,
name='11_incep_8')
x = tf.layers.max_pooling2d(x, pool_size=3, strides=2, name='maxpool6')
tf.summary.image('5_1maxpool', x[:, :, :, 0:3], max_outputs=4)
flat = tf.contrib.layers.flatten(x, scope='flatten')
dropout = tf.layers.dropout(flat,
rate=params['dropout_rate'],
training=training,
name='dropout')
fc = tf.layers.dense(dropout, units=1028, name='fc')
logits = tf.layers.dense(fc, units=params['num_classes'], name='logits')
predictions = {
'argmax': tf.argmax(logits, axis=1, name='prediction_classes'),
'softmax': tf.nn.softmax(logits, name='prediction_softmax')
}
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)
onehot_labels = tf.one_hot(indices=labels,
depth=params['num_classes'],
name='onehot_labels')
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=params['learning_rate'])
extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
eval_metric_ops = {
'accuracy_val':
tf.metrics.accuracy(labels=labels, predictions=predictions['argmax'])
}
tf.summary.histogram('labels', labels)
tf.summary.scalar('accuracy_train', eval_metric_ops['accuracy_val'][1])
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def model_fn(features, labels, mode, params):
x = tf.reshape(features, [-1, 125, 161, 2], name='stricCnn5')
x_norm = tf.layers.batch_normalization(
x, training=mode == tf.estimator.ModeKeys.TRAIN, name='x_norm')
x = tf.reshape(x_norm[:, :, :, 0], [-1, 125, 161, 1], name='reshape_spec')
conv = x
conv = stric_block(conv, 32, mode, 'conv_1')
conv = stric_block(conv, 64, mode, 'conv_2')
conv = stric_block(conv, 128, mode, 'conv_3')
conv = stric_block(conv, 256, mode, 'conv_4')
conv = stric_block(conv, 512, mode, 'conv_5', only_same=True)
incep = inception_block(conv,
t1x1=128,
t3x3=128,
t5x5=128,
tmp=128,
name='incep6')
incep = inception_block(incep,
t1x1=128,
t3x3=128,
t5x5=128,
tmp=128,
name='incep7')
incep = inception_block(incep,
t1x1=128,
t3x3=128,
t5x5=128,
tmp=128,
name='incep8')
flat = flatten(incep)
dropout1 = tf.layers.dropout(flat,
rate=params['dropout_rate'],
training=mode == tf.estimator.ModeKeys.TRAIN,
name='dropout1')
dense1 = tf.layers.dense(dropout1,
units=2048,
activation=tf.nn.relu,
name='dense1')
dropout2 = tf.layers.dropout(dense1,
rate=params['dropout_rate'],
training=mode == tf.estimator.ModeKeys.TRAIN,
name='dropout2')
dense2 = tf.layers.dense(dropout2,
units=2048,
activation=tf.nn.relu,
name='dense2')
dropout3 = tf.layers.dropout(dense2,
rate=params['dropout_rate'],
training=mode == tf.estimator.ModeKeys.TRAIN,
name='dropout3')
logits = tf.layers.dense(dropout3,
units=params['num_classes'],
name='logits')
predictions = {
'classes': tf.argmax(logits, axis=1, name='prediction_classes'),
'probabilities': tf.nn.softmax(logits, name='prediction_softmax')
}
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions={'predictions': predictions['probabilities']})
tf.summary.image(
'confusion_matrix',
conf_mat(labels, predictions['classes'], params['num_classes']))
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32),
depth=params['num_classes'],
name='onehot_labels')
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=params['learning_rate'])
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
eval_metric_ops = {
'accuracy':
tf.metrics.accuracy(labels=labels, predictions=predictions['classes'])
}
tf.summary.scalar('accuracy', eval_metric_ops['accuracy'][1])
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
def branch(x, name, mode, params):
with tf.variable_scope(name):
conv1 = tf.layers.conv2d(x,
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv1')
conv1b = tf.layers.conv2d(conv1,
filters=16,
kernel_size=3,
activation=tf.nn.relu,
name='conv1b')
pool1 = tf.layers.max_pooling2d(conv1b,
pool_size=[2, 2],
strides=2,
name='pool1')
if params['verbose_summary']:
log_conv_kernel('{}/conv1'.format(name))
log_conv_kernel('{}/conv1b'.format(name))
tf.summary.image('pool1', pool1[:, :, :, 0:1])
incep2 = inception_block(pool1,
t1x1=8,
t3x3=8,
t5x5=8,
tmp=8,
name='incep2')
conv3 = tf.layers.conv2d(incep2,
filters=32,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv3')
conv3b = tf.layers.conv2d(conv3,
filters=32,
kernel_size=3,
activation=tf.nn.relu,
name='conv3b')
pool3 = tf.layers.max_pooling2d(conv3b,
pool_size=[2, 2],
strides=2,
name='pool3')
if params['verbose_summary']:
log_conv_kernel('{}/conv3'.format(name))
log_conv_kernel('{}/conv3b'.format(name))
tf.summary.image('pool3', pool3[:, :, :, 0:1])
conv5 = tf.layers.conv2d(pool3,
filters=64,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv5')
conv5b = tf.layers.conv2d(conv5,
filters=64,
kernel_size=3,
activation=tf.nn.relu,
name='conv5b')
pool5 = tf.layers.max_pooling2d(conv5b,
pool_size=[2, 2],
strides=2,
name='pool5')
if params['verbose_summary']:
log_conv_kernel('{}/conv5'.format(name))
log_conv_kernel('{}/conv5b'.format(name))
tf.summary.image('pool5', pool5[:, :, :, 0:1])
incep6 = inception_block(pool5,
t1x1=32,
t3x3=32,
t5x5=32,
tmp=32,
name='incep6')
conv7 = tf.layers.conv2d(incep6,
filters=128,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv7')
conv7b = tf.layers.conv2d(conv7,
filters=128,
kernel_size=3,
activation=tf.nn.relu,
name='conv7b')
pool7 = tf.layers.max_pooling2d(conv7b,
pool_size=[2, 2],
strides=2,
name='pool7')
if params['verbose_summary']:
log_conv_kernel('{}/conv7'.format(name))
log_conv_kernel('{}/conv7b'.format(name))
tf.summary.image('pool7', pool7[:, :, :, 0:1])
conv8 = tf.layers.conv2d(pool7,
filters=256,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv8')
conv8b = tf.layers.conv2d(conv8,
filters=256,
kernel_size=3,
activation=tf.nn.relu,
name='conv8b')
pool8 = tf.layers.max_pooling2d(conv8b,
pool_size=[2, 2],
strides=2,
name='pool8')
if params['verbose_summary']:
log_conv_kernel('{}/conv8'.format(name))
log_conv_kernel('{}/conv8b'.format(name))
tf.summary.image('pool8', pool8[:, :, :, 0:1])
flat = flatten(pool8)
dense = tf.layers.dense(flat,
units=2048,
activation=tf.nn.relu,
name='dense')
dropout = tf.layers.dropout(
dense,
rate=params['dropout_rate'],
training=mode == tf.estimator.ModeKeys.TRAIN,
name='dropout')
return dropout
def model_fn(features, labels, mode, params):
with tf.variable_scope('x_prep'):
x = tf.reshape(features, [-1, 16000], name='input_incep108_lg')
x_norm = tf.layers.batch_normalization(
x, training=mode == tf.estimator.ModeKeys.TRAIN, name='x_norm')
spec_pow = get_spectrogram(x_norm, type='power', name='spec_pow')
spec_mag = get_spectrogram(x_norm, type='magnitude', name='spec_mag')
spec_mel = get_spectrogram(x_norm, type='mel', name='spec_mel')
_, h, w, _ = spec_pow.get_shape()
spec = tf.reshape(tf.stack([spec_pow, spec_mag, spec_mel],
name='stack'), [-1, h, w, 3],
name='stack_reshape')
if params['verbose_summary']:
tf.summary.audio('input', x_norm, 16000, max_outputs=12)
tf.summary.image('mag', spec_mag)
tf.summary.image('pow', spec_pow)
tf.summary.image('mel', spec_mel)
tf.summary.image('spec', spec)
conv1 = tf.layers.conv2d(spec,
filters=16,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv1')
conv1b = tf.layers.conv2d(conv1,
filters=16,
kernel_size=3,
activation=tf.nn.relu,
name='conv1b')
conv1c = tf.layers.conv2d(conv1b,
filters=16,
kernel_size=3,
activation=tf.nn.relu,
name='conv1c')
pool1 = tf.layers.max_pooling2d(conv1c,
pool_size=[2, 2],
strides=2,
name='pool1')
if params['verbose_summary']:
log_conv_kernel('conv1')
log_conv_kernel('conv1b')
log_conv_kernel('conv1c')
tf.summary.image('pool1', pool1[:, :, :, 0:1])
incep2 = inception_block(pool1,
t1x1=8,
t3x3=8,
t5x5=8,
tmp=8,
name='incep2')
conv3 = tf.layers.conv2d(incep2,
filters=32,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv3')
conv3b = tf.layers.conv2d(conv3,
filters=32,
kernel_size=3,
activation=tf.nn.relu,
name='conv3b')
conv3c = tf.layers.conv2d(conv3b,
filters=32,
kernel_size=3,
activation=tf.nn.relu,
name='conv3c')
pool3 = tf.layers.max_pooling2d(conv3c,
pool_size=[2, 2],
strides=2,
name='pool3')
if params['verbose_summary']:
log_conv_kernel('conv3')
log_conv_kernel('conv3b')
log_conv_kernel('conv3c')
tf.summary.image('pool3', pool3[:, :, :, 0:1])
conv5 = tf.layers.conv2d(pool3,
filters=64,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv5')
conv5b = tf.layers.conv2d(conv5,
filters=64,
kernel_size=3,
activation=tf.nn.relu,
name='conv5b')
conv5c = tf.layers.conv2d(conv5b,
filters=64,
kernel_size=3,
activation=tf.nn.relu,
name='conv5c')
pool5 = tf.layers.max_pooling2d(conv5c,
pool_size=[2, 2],
strides=2,
name='pool5')
if params['verbose_summary']:
log_conv_kernel('conv5')
log_conv_kernel('conv5b')
log_conv_kernel('conv5c')
tf.summary.image('pool5', pool5[:, :, :, 0:1])
incep6 = inception_block(pool5,
t1x1=32,
t3x3=32,
t5x5=32,
tmp=32,
name='incep6')
conv7 = tf.layers.conv2d(incep6,
filters=128,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv7')
conv7b = tf.layers.conv2d(conv7,
filters=128,
kernel_size=3,
activation=tf.nn.relu,
name='conv7b')
pool7 = tf.layers.max_pooling2d(conv7b,
pool_size=[2, 2],
strides=2,
name='pool7')
if params['verbose_summary']:
log_conv_kernel('conv7')
log_conv_kernel('conv7b')
tf.summary.image('pool7', pool7[:, :, :, 0:1])
conv8 = tf.layers.conv2d(pool7,
filters=256,
kernel_size=3,
padding='same',
activation=tf.nn.relu,
name='conv8')
conv8b = tf.layers.conv2d(conv8,
filters=256,
kernel_size=3,
activation=tf.nn.relu,
name='conv8b')
conv8c = tf.layers.conv2d(conv8b,
filters=256,
kernel_size=3,
activation=tf.nn.relu,
name='conv8c')
if params['verbose_summary']:
log_conv_kernel('conv8')
log_conv_kernel('conv8b')
log_conv_kernel('conv8c')
incep9 = inception_block(conv8c,
t1x1=64,
t3x3=64,
t5x5=64,
tmp=64,
name='incep9')
conv10 = tf.layers.conv2d(incep9,
filters=512,
kernel_size=3,
activation=tf.nn.relu,
name='conv10')
pool10 = tf.layers.max_pooling2d(conv10,
pool_size=[2, 2],
strides=2,
name='pool10')
if params['verbose_summary']:
log_conv_kernel('conv10')
tf.summary.image('pool8', pool10[:, :, :, 0:1])
flat = flatten(pool10)
dense4 = tf.layers.dense(flat,
units=2048,
activation=tf.nn.relu,
name='dense4')
dropout4 = tf.layers.dropout(dense4,
rate=params['dropout_rate'],
training=mode == tf.estimator.ModeKeys.TRAIN,
name='dropout4')
logits = tf.layers.dense(dropout4,
units=params['num_classes'],
name='logits')
predictions = {
'classes': tf.argmax(logits, axis=1, name='prediction_classes'),
'probabilities': tf.nn.softmax(logits, name='prediction_softmax')
}
if mode == tf.estimator.ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(
mode=mode,
predictions={'predictions': predictions['probabilities']})
onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32),
depth=params['num_classes'],
name='onehot_labels')
loss = tf.losses.softmax_cross_entropy(onehot_labels=onehot_labels,
logits=logits)
tf.summary.scalar('loss', loss)
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=params['learning_rate'])
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer.minimize(loss=loss,
global_step=tf.train.get_global_step())
eval_metric_ops = {
'accuracy':
tf.metrics.accuracy(labels=labels, predictions=predictions['classes'])
}
tf.summary.scalar('accuracy', eval_metric_ops['accuracy'][1])
return tf.estimator.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op,
eval_metric_ops=eval_metric_ops)
