def build_model(inputs, mask, units, depth, n_labels, feat_dim, init_lr, direction,
dropout, init_filters, optim, lstm=False, vgg=False):
filters=init_filters
outputs = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(inputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(inputs, init_filters, feat_dim)
outputs = network.network(outputs,units, depth, n_labels, direction, dropout, lstm)
outputs = TimeDistributed(Dense(n_labels+1))(outputs)
outputs = Activation('softmax')(outputs)
model=Model([inputs, mask], outputs)
# we can get accuracy from data along with batch/temporal axes.
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr, clipnorm=50.),
loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
else:
model.compile(keras.optimizers.Adadelta(),
loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
return model
def build_model(inputs,
mask,
units,
depth,
n_labels,
feat_dim,
dropout,
init_filters,
lstm=False,
vgg=False):
outputs = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(outputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(outputs, init_filters, feat_dim)
outputs = Lambda(lambda x: tf.multiply(x[0], x[1]))([outputs, mask])
outputs = Masking(mask_value=0.0)(outputs)
outputs = network.lc_network(outputs, units, depth, n_labels, dropout,
init_filters, lstm)
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
model = Model([inputs, mask], outputs)
return model
def build_model(inputs,
mask,
units,
depth,
n_labels,
feat_dim,
init_lr,
dropout,
init_filters,
optim,
lstm=False,
vgg=False):
masked = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(masked, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(masked, init_filters, feat_dim)
outputs = Lambda(lambda x: tf.multiply(x[0], x[1]))([outputs, mask])
outputs = Masking(mask_value=0.0)(outputs)
outputs1 = network.lc_network(outputs, units, depth, n_labels, dropout,
init_filters, lstm)
depth = 2
init_filters = 16
# 1/2
outputs2 = dilation.VGG2L_Strides(masked, init_filters, feat_dim)
outputs2 = network.lc_network(outputs2, units, depth, n_labels, dropout,
init_filters, lstm)
outputs2 = dilation.VGG2L_Transpose(outputs2, init_filters, units * 2)
# 1/4
outputs3 = dilation.VGG2L_QuadStrides(
masked, init_filters, feat_dim) # output = time/2, feat_dim*filters*2
outputs3 = network.lc_network(outputs3, units, depth, n_labels, dropout,
init_filters,
lstm) # output = time/2, units*2
outputs3 = dilation.VGG2L_QuadTranspose(outputs3, init_filters,
units * 2) #output = time, units*2
#outputs = Add()([outputs, outputs2, outputs3])
outputs = Add()([outputs1, outputs2, outputs3])
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
model = Model([inputs, mask], outputs)
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr),
loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
else:
model.compile(keras.optimizers.Adadelta(lr=init_lr),
loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
return model
def build_model(inputs, mask, units, depth, n_labels, feat_dim, init_lr,
dropout, init_filters, optim, lstm=False, vgg=False):
outputs = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(outputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(outputs, init_filters, feat_dim)
for n in range (depth):
# forward, keep current states
# statefule
if lstm is False:
x=GRU(units, kernel_initializer='glorot_uniform',
return_sequences=True,
stateful=True,
dropout=dropout,
unroll=False)(outputs)
else:
x=LSTM(units, kernel_initializer='glorot_uniform',
return_sequences=True,
unit_forget_bias=True,
stateful=True,
dropout=dropout,
unroll=False)(outputs)
# backward, not keep current states
# do not preserve state values for backward pass
if lstm is False:
y=GRU(units, kernel_initializer='glorot_uniform',
return_sequences=True,
stateful=False,
unroll=False,
dropout=dropout,
go_backwards=True)(outputs)
else:
y=LSTM(units, kernel_initializer='glorot_uniform',
return_sequences=True,
unit_forget_bias=True,
stateful=False,
unroll=False,
dropout=dropout,
go_backwards=True)(outputs)
outputs = Concatenate(axis=-1)([x,y])
outputs=layer_normalization.LayerNormalization()(outputs)
outputs = TimeDistributed(Dense(n_labels+1))(outputs)
outputs = Activation('softmax')(outputs)
#outputs = Lambda(lambda x: tf.multiply(x[0], x[1]))([outputs, mask])
model = Model([inputs, mask], outputs)
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr, clipnorm=50.), loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
else:
model.compile(keras.optimizers.Adadelta(lr=init_lr, clipnorm=50.), loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
return model
def build_model(file,
inputs,
mask,
units,
depth,
n_labels,
feat_dim,
init_lr,
dropout,
init_filters,
optim,
proc_frames,
lstm=False,
vgg=False):
outputs = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(outputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(outputs, init_filters, feat_dim)
vgg_model = Model(inputs, outputs)
vgg_model.trainable = False
inputs = vgg_model.outputs
outputs = Lambda(lambda x: tf.multiply(x[0], x[1]))([inputs, mask])
rnn_inputs = Masking(mask_value=0.0)(outputs)
rnn_model, rnn_trunc_model = network.lc_network(rnn_inputs, units, depth,
n_labels, dropout,
init_filters, proc_frames,
lstm)
outputs = rnn_model.outputs
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
model = Model(inputs=[vgg_model.inputs, mask], outputs)
valid_model = Model(inputs=[vgg_model.inputs, mask],
rnn_trunc_model.outputs)
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr),
loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
else:
model.compile(keras.optimizers.Adadelta(lr=init_lr),
loss=['categorical_crossentropy'],
metrics=['categorical_accuracy'])
return model, valid_model
def build_model(inputs, units, depth, n_labels, feat_dim, init_lr, dropout,
init_filters, optim):
outputs = vgg2l.VGG2L(inputs, init_filters, feat_dim)
outputs = lstm(outputs, units, depth, n_labels, dropout)
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
if optim is not None:
model = CTCModel.CTCModel([inputs], [outputs], greedy=True)
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr))
elif optim == 'sgd':
model.compile(keras.optimizers.SGD(lr=init_lr, momentum=0.9))
else:
model.compile(keras.optimizers.Adadelta(lr=init_lr))
else:
model = Model(inputs, outputs)
return model
def build_model(inputs,
mask,
units,
depth,
n_labels,
feat_dim,
init_lr,
dropout,
init_filters,
optim,
lstm=False,
vgg=False):
outputs = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(outputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(outputs, init_filters, feat_dim)
outputs = Lambda(lambda x: tf.multiply(x[0], x[1]))([outputs, mask])
outputs = Masking(mask_value=0.0)(outputs)
outputs = network.lc_network(outputs, units, depth, n_labels, dropout,
init_filters, lstm)
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
model = Model([inputs, mask], outputs)
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr),
loss=[multi_utils.soft_loss],
metrics=['categorical_accuracy'])
#metrics=[multi_utils.soft_acc])
else:
model.compile(keras.optimizers.Adadelta(lr=init_lr),
loss=[multi_utils.soft_loss],
metrics=['categorical_accuracy'])
#metrics=[multi_utils.soft_acc])
return model
def build_model(inputs, units, depth, n_labels, feat_dim, init_lr, direction,
dropout, init_filters, optim, lstm=False, vgg=False):
if vgg is False:
outputs = vgg2l.VGG2L(inputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(inputs, init_filters, feat_dim)
#outputs = dilation.VGG2L_Strides(inputs, init_filters, feat_dim)
outputs = network.network(outputs,units, depth, n_labels, direction, dropout, lstm)
#outputs = dilation.VGG2L_Transpose(outputs, init_filters, units*2)
outputs = TimeDistributed(Dense(n_labels+1))(outputs)
outputs = Activation('softmax')(outputs)
model=CTCModel.CTCModel([inputs], [outputs], greedy=True)
if optim == 'adam':
model.compile(keras.optimizers.Adam(lr=init_lr))
elif optim == 'sgd':
model.compile(keras.optimizers.SGD(lr=init_lr, momentum=0.9))
else:
model.compile(keras.optimizers.Adadelta(lr=init_lr))
return model
def build_model(inputs,
units,
depth,
n_labels,
feat_dim,
direction,
init_filters,
lstm=False,
vgg=False):
if vgg is False:
outputs = vgg2l.VGG2L(inputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(inputs, init_filters, feat_dim)
outputs = network.network(outputs, units, depth, n_labels, direction, 0.0,
lstm)
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
model = Model(inputs, outputs)
return model
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--data',
type=str,
required=True,
help='training data')
parser.add_argument('--eval', type=str, help='evaluation data')
parser.add_argument('--feat-dim', default=40, type=int, help='feats dim')
parser.add_argument('--n-labels',
default=1024,
type=int,
required=True,
help='number of output labels')
parser.add_argument('--batch-size',
default=1,
type=int,
help='mini-batch size')
parser.add_argument('--snapshot',
type=str,
default='./',
help='snapshot directory')
parser.add_argument('--snapshot-prefix',
type=str,
default='snapshot',
help='snapshot file prefix')
parser.add_argument('--eval-output-prefix', type=str, default='eval_out')
parser.add_argument('--units',
type=int,
default=16,
help='number of LSTM cells')
parser.add_argument('--lstm-depth',
type=int,
default=2,
help='number of LSTM layers')
parser.add_argument('--process-frames',
type=int,
default=10,
help='process frames')
parser.add_argument('--extra-frames1',
type=int,
default=10,
help='1st extra frames')
parser.add_argument('--extra-frames2',
type=int,
default=10,
help='2nd extra frames')
parser.add_argument('--num-extra-frames1',
type=int,
default=1,
help='number of extra frames 1')
parser.add_argument('--filters',
type=int,
default=16,
help='number of filters')
parser.add_argument('--lstm', action='store_true')
parser.add_argument('--vgg', action='store_true')
parser.add_argument('--prior',
type=str,
default=None,
help='prior weights')
parser.add_argument('--prior-scale',
type=float,
default=1.0,
help='prior scaler')
parser.add_argument('--weights',
type=str,
required=True,
help='model weights')
args = parser.parse_args()
eval_in = Input(batch_shape=(args.batch_size, None, args.feat_dim))
eval_mask = Input(batch_shape=(args.batch_size, None,
args.feat_dim * args.filters * 2))
outputs = Masking(mask_value=0.0)(inputs)
if vgg is False:
outputs = vgg2l.VGG2L(outputs, init_filters, feat_dim)
else:
outputs = vgg1l.VGG(outputs, init_filters, feat_dim)
outputs = Lambda(lambda x: tf.multiply(x[0], x[1]))([outputs, mask])
outputs = Masking(mask_value=0.0)(outputs)
outputs = network.lc_network(outputs, units, depth, n_labels, dropout,
lstm)
outputs = TimeDistributed(Dense(n_labels + 1))(outputs)
outputs = Activation('softmax')(outputs)
model = Model([inputs, mask], outputs)
return model
