def res_with_center_loss_model(shape=(32, 1024),
num_classes=500,
n=1,
feature_length=100,
l2_sm=15,
lambda_c=0.2):
origin_model = full_res_net_model(shape, num_classes, n, feature_length,
l2_sm)
origin_input = origin_model.input
origin_feature_output = origin_model.get_layer('feature_layer').output
origin_output = origin_model.get_layer('output_layer').output
input_target = Input(shape=(1, ))
centers = Embedding(num_classes, feature_length,
name='embedding_layer')(input_target)
l2_loss = Lambda(
lambda x: K.sum(K.square(x[0] - x[1][:, 0]), 1, keepdims=True),
name='l2_loss')([origin_feature_output, centers])
model_center_loss = Model(inputs=[origin_input, input_target],
outputs=[origin_output, l2_loss])
model_center_loss.compile(
optimizer=keras.optimizers.Adadelta(),
loss=[l2_softmax(l2_sm), lambda y_true, y_pred: y_pred],
loss_weights=[1, lambda_c],
metrics=['accuracy'])
return model_center_loss
def full_res_net_model_voxceleb(shape=(32, 1024),
num_classes=500,
n=1,
feature_length=100,
l2_sm=15,
**kwargs):
input_array = keras.Input(shape, name='input')
three_d_input = keras.layers.Reshape(target_shape=(*shape, 1))(input_array)
resnet_output = resnet_v2(inputs=three_d_input, n=n)
resnet_output = BatchNormalization()(resnet_output)
mid = keras.layers.Dense(feature_length,
activation='sigmoid',
name="feature_layer")(resnet_output)
mid = BatchNormalization()(mid)
output = keras.layers.Dense(num_classes,
activation='softmax',
name="output_layer")(mid)
model = Model(inputs=input_array, outputs=output)
model.compile(loss=l2_softmax(l2_sm),
optimizer="adam",
metrics=['accuracy'])
model.summary()
return model
def full_res_net_model(shape=(32, 1024),
num_classes=500,
n=1,
feature_length=100,
l2_sm=15,
**kwargs):
input_array = keras.Input(shape, name='input')
# three_d_input = keras.layers.Reshape(target_shape=(*shape, 1))(input_array)
three_d_input = SpatialDropout2D(0.1)(input_array)
resnet_output = resnet_v2(inputs=three_d_input, n=n)
resnet_output = BatchNormalization()(resnet_output)
resnet_output = Dropout(0.1)(resnet_output)
mid = keras.layers.Dense(feature_length,
activation=relu_advanced,
name="feature_layer")(resnet_output)
mid = Dropout(0.05)(mid)
mid = BatchNormalization()(mid)
output = keras.layers.Dense(num_classes,
activation='softmax',
name="output_layer")(mid)
model = Model(inputs=input_array, outputs=output)
model.compile(loss=l2_softmax(l2_sm),
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
model.summary()
return model
def res_plus_attention_model(shape=(32, 1024),
num_classes=500,
n=1,
feature_length=100,
l2_sm=15,
**kwargs):
input_array = keras.Input(shape, name='input')
resnet_output = resnet_1d(inputs=input_array, n=n)
resnet_output = BatchNormalization()(resnet_output)
resnet_output = Dropout(0.3)(resnet_output)
mid = keras.layers.Dense(feature_length,
activation='sigmoid',
name="feature_layer")(resnet_output)
mid = Dropout(0.3)(mid)
mid = BatchNormalization()(mid)
output = keras.layers.Dense(num_classes,
activation='softmax',
name="output_layer")(mid)
model = Model(inputs=input_array, outputs=output)
model.compile(loss=l2_softmax(l2_sm),
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
model.summary()
plot_model(model, "attention.png")
return model
def rnn(shape=(32, 1024), num_classes=500, l2_sm=15, **kwargs):
input_array = keras.Input(shape, name='input')
rnn = GRU(50, return_sequences=False)(input_array)
full = rnn
den = Dense(50, activation="relu", name="feature_layer")(full)
output = Dense(num_classes, activation='softmax', name="output_layer")(den)
model = Model(inputs=input_array, outputs=output)
model.compile(loss=l2_softmax(l2_sm),
optimizer=keras.optimizers.Adadelta(),
metrics=['accuracy'])
model.summary()
return model
def res_plus_transformer_model(shape=(32, 1024), num_classes=500):
input_array = keras.Input(shape)
three_d_input = keras.layers.Reshape(target_shape=(*shape, 1))(input_array)
transformer_output = keras.layers.Flatten()(get_transformer(
transformer_input=input_array, transformer_depth=3))
resnet_output = resnet_v2(inputs=three_d_input, n=1)
mid = keras.layers.concatenate([resnet_output, transformer_output])
output = keras.layers.Dense(num_classes,
activation='relu',
kernel_initializer='he_normal')(mid)
model = Model(inputs=input_array, outputs=output)
model.compile(loss=l2_softmax(5), optimizer="sgd", metrics=['accuracy'])
model.summary()
# plot_model(model, to_file='model.png')
return model
def text_cnn(shape=(500, 64),
num_classes=2,
input_dim=93,
feature_length=100,
l2_sm=15,
**kwargs):
input_array = keras.Input(shape, name='input')
three_d_input = keras.layers.Reshape(target_shape=(*shape, 1))(input_array)
embedding_output = three_d_input
_embed = SpatialDropout2D(0.1)(embedding_output)
warppers = []
num_filters = 64
kernel_size = [3, 5]
conv_action = 'relu'
for _kernel_size in kernel_size:
for dilated_rate in [1, 3]:
conv1d = Conv2D(filters=num_filters,
kernel_size=_kernel_size,
activation=conv_action,
dilation_rate=dilated_rate)(_embed)
warppers.append(GlobalMaxPooling2D()(conv1d))
fc = concatenate(warppers)
fc = BatchNormalization()(fc)
fc = Dropout(0.1)(fc)
fc = Dense(feature_length, activation='sigmoid', name="feature_layer")(fc)
fc = Dropout(0.1)(fc)
preds = Dense(num_classes, activation='softmax', name="output_layer")(fc)
model = Model(inputs=input_array, outputs=preds)
model.summary()
model.compile(loss=l2_softmax(l2_sm),
optimizer='adam',
metrics=['accuracy'])
# plot_model(model, "attention.png")
return model
for layer in origin_model.layers:
layer.trainable = False
input_target = Input(shape=(1, ))
centers = Embedding(class_num, feature_length,
name='embedding_layer')(input_target)
l2_loss = Lambda(
lambda x: K.sum(K.square(x[0] - x[1][:, 0]), 1, keepdims=True),
name='l2_loss')([origin_feature_output, centers])
model_center_loss = Model(inputs=[origin_input, input_target],
outputs=[origin_output, l2_loss])
model_center_loss.compile(
optimizer="adam",
loss=[l2_softmax(l2_lambda), lambda y_true, y_pred: y_pred],
loss_weights=[1, lambda_c],
metrics=['accuracy'])
model = model_center_loss
# create dir
checkpoint = ModelCheckpoint(
filepath='./models/weights.{epoch:02d}-{val_loss:.2f}.hdf5',
monitor='val_acc',
verbose=1,
save_best_only=False)
callbacks = [checkpoint]
print(int(len(x) / batch_size))
if not load_trained_model:
def load_model(model_path, load_type=0) -> keras.Model:
"""
返回训练好的模型
:return:
"""
def temp(a, b):
return b
if load_type == 0:
return keras.models.load_model(model_path,
custom_objects={
'internal': l2_softmax(10),
'<lambda>': temp
})
elif load_type == 2:
model = keras.models.load_model(model_path,
custom_objects={
'internal': l2_softmax(10),
'<lambda>': temp
})
output = model.get_layer('feature_layer').output
new_model = Model(inputs=model.get_layer('input').input,
outputs=output)
return new_model
elif load_type == 1:
model = keras.models.load_model(
model_path, custom_objects={'internal': l2_softmax(10)})
output = model.get_layer('feature_layer').output
new_model = Model(inputs=model.input, outputs=output)
return new_model
elif load_type == 3:
model = keras.models.load_model(model_path,
custom_objects={
'bpr_triplet_loss':
bpr_triplet_loss,
'identity_loss': identity_loss
})
print(len(model.layers))
return Model(inputs=model.get_layer('model_1').get_input_at(0),
output=model.get_layer('model_1').get_output_at(0))
elif load_type == 4:
model = keras.models.load_model(model_path,
custom_objects={
'cos_distance':
cos_distance,
'eucl_dist_output_shape':
eucl_dist_output_shape,
'contrastive_loss':
contrastive_loss,
'accuracy':
accuracy,
'l2_normalize':
l2_normalize,
'euclidean_distance':
euclidean_distance
})
print(len(model.layers))
return Model(inputs=model.get_layer('model_1').get_input_at(0),
output=model.get_layer('model_1').get_output_at(0))
elif load_type == 5:
model = keras.models.load_model(model_path,
custom_objects={
'internal': l2_softmax(10),
'<lambda>': temp
})
return Model(inputs=model.input, output=model.layers[-2].output)