def init_model(self, input_shape, num_classes, **kwargs):
inputs = Input(shape=input_shape)
# bnorm_1 = BatchNormalization(axis=2)(inputs)
lstm_1 = Bidirectional(CuDNNLSTM(64,
name='blstm_1',
return_sequences=True),
merge_mode='concat')(inputs)
activation_1 = Activation('tanh')(lstm_1)
dropout1 = SpatialDropout1D(0.5)(activation_1)
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
pool_1 = GlobalMaxPool1D()(attention_1)
dropout2 = Dropout(rate=0.5)(pool_1)
dense_1 = Dense(units=256, activation='relu')(dropout2)
outputs = Dense(units=num_classes, activation='softmax')(dense_1)
model = TFModel(inputs=inputs, outputs=outputs)
optimizer = optimizers.Adam(
# learning_rate=1e-3,
lr=1e-3,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0002,
amsgrad=True)
model.compile(optimizer=optimizer,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, input_shape, num_classes, **kwargs):
inputs = Input(shape=input_shape)
# bnorm_1 = BatchNormalization(axis=-1)(inputs)
x = Bidirectional(CuDNNLSTM(96, name='blstm1', return_sequences=True),
merge_mode='concat')(inputs)
# activation_1 = Activation('tanh')(lstm_1)
x = SpatialDropout1D(0.1)(x)
x = Attention(8, 16)([x, x, x])
x1 = GlobalMaxPool1D()(x)
x2 = GlobalAvgPool1D()(x)
x = Concatenate(axis=-1)([x1, x2])
x = Dense(units=128, activation='elu')(x)
x = Dense(units=64, activation='elu')(x)
x = Dropout(rate=0.4)(x)
outputs = Dense(units=num_classes, activation='softmax')(x)
model = TFModel(inputs=inputs, outputs=outputs)
optimizer = optimizers.Adam(
# learning_rate=1e-3,
lr=1e-3,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0002,
amsgrad=True)
model.compile(optimizer=optimizer,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, input_shape, num_classes, **kwargs):
inputs = Input(shape=input_shape)
sequence_len = input_shape[0]
lstm_units_array = np.array([32, 64, 128, 256, 512])
lstm_units = lstm_units_array[np.argmin(
np.abs(lstm_units_array - sequence_len))]
lstm_1 = CuDNNLSTM(lstm_units, return_sequences=True)(inputs)
activation_1 = Activation('tanh')(lstm_1)
if num_classes >= 20:
if num_classes < 30:
dropout1 = SpatialDropout1D(0.5)(activation_1)
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
else:
attention_1 = Attention(
8, 16)([activation_1, activation_1, activation_1])
k_num = 10
kmaxpool_l = Lambda(lambda x: tf.reshape(tf.nn.top_k(
tf.transpose(x, [0, 2, 1]), k=k_num, sorted=True)[0],
shape=[-1, k_num, 128]))(
attention_1)
flatten = Flatten()(kmaxpool_l)
dropout2 = Dropout(rate=0.5)(flatten)
else:
dropout1 = SpatialDropout1D(0.5)(activation_1)
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
pool_l = GlobalMaxPool1D()(attention_1)
dropout2 = Dropout(rate=0.5)(pool_l)
dense_1 = Dense(units=256, activation='relu')(dropout2)
# dense_1 = Dense(units=256, activation='softplus',kernel_regularizer=regularizers.l2(0.01),
# activity_regularizer=regularizers.l1(0.01))(dropout2)
#dense_1 = DropConnect(Dense(units=256, activation='softplus'), prob=0.5)(dropout2)
outputs = Dense(units=num_classes, activation='softmax')(dense_1)
loss_fun = CategoricalCrossentropy(label_smoothing=0.2)
model = TFModel(inputs=inputs, outputs=outputs)
optimizer = optimizers.Nadam(lr=0.002,
beta_1=0.9,
beta_2=0.999,
epsilon=None,
schedule_decay=0.004)
model.compile(
optimizer=optimizer,
loss=loss_fun,
#loss="sparse_categorical_crossentropy",
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, config):
input_shape = config['max_len']
num_classes = config['num_classes']
inputs = Input(shape=(input_shape, 96))
x = inputs
cnn1 = Conv1D(50,
kernel_size=1,
strides=1,
padding='same',
kernel_initializer='he_normal')(x)
cnn1 = BatchNormalization(axis=-1)(cnn1)
cnn1 = LeakyReLU()(cnn1)
cnn1 = GlobalMaxPooling1D()(
cnn1) # CNN_Dynamic_MaxPooling(cnn1,50,2,2)
cnn2 = Conv1D(50,
kernel_size=3,
strides=1,
padding='same',
kernel_initializer='he_normal')(x)
cnn2 = BatchNormalization(axis=-1)(cnn2)
cnn2 = LeakyReLU()(cnn2)
cnn2 = GlobalMaxPooling1D()(cnn2)
cnn3 = Conv1D(50,
kernel_size=5,
strides=1,
padding='same',
kernel_initializer='he_normal')(x)
cnn3 = BatchNormalization(axis=-1)(cnn3)
cnn3 = LeakyReLU()(cnn3)
cnn3 = GlobalMaxPooling1D()(cnn3)
x = concatenate([cnn1, cnn2, cnn3], axis=-1)
x = Dense(units=num_classes, activation='softmax')(x)
model = TFModel(inputs=inputs, outputs=x)
opt = optimizers.rmsprop(lr=0.0001, decay=1e-6)
model.compile(optimizer=opt,
loss="sparse_categorical_crossentropy",
metrics=['acc'])
model.summary()
self._model = model
self.is_init = True
def init_model(self,
input_shape,
num_classes,
**kwargs):
inputs = Input(shape=input_shape)
lstm_1 = CuDNNLSTM(128, return_sequences=True)(inputs)
activation_1 = Activation('tanh')(lstm_1)
if num_classes >= 20:
if num_classes < 30:
dropout1 = SpatialDropout1D(0.5)(activation_1)
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
# no dropout to get more infomation for classifying a large number
# classes
else:
attention_1 = Attention(8, 16)(
[activation_1, activation_1, activation_1])
k_num = 10
kmaxpool_l = Lambda(
lambda x: tf.reshape(tf.nn.top_k(tf.transpose(x, [0, 2, 1]), k=k_num, sorted=True)[0],
shape=[-1, k_num, 128]))(attention_1)
flatten = Flatten()(kmaxpool_l)
dropout2 = Dropout(rate=0.5)(flatten)
else:
dropout1 = SpatialDropout1D(0.5)(activation_1)
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
pool_l = GlobalMaxPool1D()(attention_1)
dropout2 = Dropout(rate=0.5)(pool_l)
dense_1 = Dense(units=256, activation='softplus')(dropout2)
outputs = Dense(units=num_classes, activation='softmax')(dense_1)
model = TFModel(inputs=inputs, outputs=outputs)
optimizer = optimizers.Nadam(
lr=0.002,
beta_1=0.9,
beta_2=0.999,
epsilon=None,
schedule_decay=0.004)
model.compile(
optimizer=optimizer,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self,
input_shape,
num_classes,
**kwargs):
inputs = Input(shape=input_shape)
# bnorm_1 = BatchNormalization(axis=2)(inputs)
sequence_len = input_shape[0]
lstm_units_array = np.array([32, 64, 128, 256, 512])
lstm_units = lstm_units_array[np.argmin(np.abs(lstm_units_array-sequence_len))]
lstm_1 = Bidirectional(CuDNNLSTM(lstm_units, name='blstm_1',
return_sequences=True),
merge_mode='concat')(inputs)
activation_1 = Activation('tanh')(lstm_1)
dropout1 = SpatialDropout1D(0.5)(activation_1)
if lstm_units <=128:
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
else:
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
pool_1 = GlobalMaxPool1D()(attention_1)
dropout2 = Dropout(rate=0.5)(pool_1)
dense_1 = Dense(units=256, activation='relu')(dropout2)
# dense_1 = Dense(units=256, activation='relu',kernel_regularizer=regularizers.l2(0.01),
# activity_regularizer=regularizers.l1(0.01))(dropout2)
#dense_1 = DropConnect(Dense(units=256, activation='relu'), prob=0.5)(dropout2)
outputs = Dense(units=num_classes, activation='softmax')(dense_1)
model = TFModel(inputs=inputs, outputs=outputs)
loss_fun = CategoricalCrossentropy(label_smoothing=0.2)
optimizer = optimizers.Adam(
# learning_rate=1e-3,
lr=1e-3,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0002,
amsgrad=True)
model.compile(
optimizer=optimizer,
loss=loss_fun,
#loss="sparse_categorical_crossentropy",
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, input_shape, num_classes, **kwargs):
# FIXME: keras sequential model is better than keras functional api,
# why???
inputs = Input(shape=input_shape)
# dropout0 = SpatialDropout2D(rate=0.1, data_format='channels_last')(inputs)
min_size = min(input_shape[:2])
pool_l = None
for i in range(5):
if i == 0:
conv_l = Conv2D(64,
3,
input_shape=input_shape,
padding='same',
data_format='channels_last')(inputs)
else:
conv_l = Conv2D(64, 3, padding='same')(pool_l)
activation_l = Activation('relu')(conv_l)
bn_l = BatchNormalization()(activation_l)
pool_l = MaxPooling2D(pool_size=(2, 2))(bn_l)
min_size //= 2
if min_size < 2:
break
avgpool_l = GlobalAveragePooling2D(data_format='channels_last')(pool_l)
maxpool_l = GlobalMaxPooling2D(data_format='channels_last')(pool_l)
concat = Concatenate()([avgpool_l, maxpool_l])
flatten = Flatten()(concat)
bn1 = BatchNormalization()(flatten)
dense1 = Dense(256, activation='relu')(bn1)
bn2 = BatchNormalization()(dense1)
dropout1 = Dropout(rate=0.5)(bn2)
outputs = Dense(num_classes, activation='softmax')(dropout1)
model = TFModel(inputs=inputs, outputs=outputs)
# optimizer = tf.keras.optimizers.Adadelta()
optimizer = tf.keras.optimizers.Adam()
# optimizer = optimizers.SGD(lr=1e-3, decay=2e-4, momentum=0.9, clipvalue=5)
model.compile(loss='sparse_categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
model.summary()
self.is_init = True
self._model = model
def init_model(self,
input_shape,
num_classes,
**kwargs):
layers = 5
filters_size = [64, 128, 256, 512, 512]
kernel_size = (3, 3)
pool_size = [(2, 2), (2, 2), (2, 2), (4, 1), (4, 1)]
freq_axis = 2
channel_axis = 3
channel_size = 128
min_size = min(input_shape[:2])
melgram_input = Input(shape=input_shape)
# x = ZeroPadding2D(padding=(0, 37))(melgram_input)
x = Reshape((input_shape[0], input_shape[1], 1))(melgram_input)
x = BatchNormalization(axis=freq_axis, name='bn_0_freq')(x)
# Conv block 1
x = Convolution2D(
filters=filters_size[0],
kernel_size=kernel_size,
padding='same',
name='conv1')(x)
x = ELU()(x)
x = BatchNormalization(axis=channel_axis, name='bn1')(x)
x = MaxPooling2D(
pool_size=pool_size[0],
strides=pool_size[0],
name='pool1')(x)
x = Dropout(0.1, name='dropout1')(x)
min_size = min_size // pool_size[0][0]
for layer in range(1, layers):
min_size = min_size // pool_size[layer][0]
if min_size < 1:
break
x = Convolution2D(
filters=filters_size[layer],
kernel_size=kernel_size,
padding='same',
name='conv' + str(layer + 1))(x)
x = ELU()(x)
x = BatchNormalization(axis=channel_axis, name='bn'+str(layer + 1)+'')(x)
x = MaxPooling2D(
pool_size=pool_size[layer],
strides=pool_size[layer],
name='pool'+str(layer + 1)+'')(x)
x = Dropout(0.1, name='dropout'+str(layer + 1)+'')(x)
x = Reshape((-1, channel_size))(x)
gru_units = 32
if num_classes > 32:
gru_units = int(num_classes * 1.5)
# GRU block 1, 2, output
x = CuDNNGRU(gru_units, return_sequences=True, name='gru1')(x)
x = CuDNNGRU(gru_units, return_sequences=False, name='gru2')(x)
x = Dropout(0.3)(x)
outputs = Dense(num_classes, activation='softmax', name='output')(x)
model = TFModel(inputs=melgram_input, outputs=outputs)
optimizer = optimizers.Adam(
# learning_rate=1e-3,
lr=1e-3,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=1e-4,
amsgrad=True)
model.compile(
optimizer=optimizer,
loss="sparse_categorical_crossentropy",
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, input_shape, num_classes, **kwargs):
freq_axis = 2
channel_axis = 3
channel_size = 128
min_size = min(input_shape[:2])
melgram_input = Input(shape=input_shape)
# x = ZeroPadding2D(padding=(0, 37))(melgram_input)
# x = BatchNormalization(axis=freq_axis, name='bn_0_freq')(x)
x = Reshape((input_shape[0], input_shape[1], 1))(melgram_input)
# Conv block 1
x = Convolution2D(64, 3, 1, padding='same', name='conv1')(x)
x = BatchNormalization(axis=channel_axis, name='bn1')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='pool1')(x)
x = Dropout(0.1, name='dropout1')(x)
# Conv block 2
x = Convolution2D(channel_size, 3, 1, padding='same', name='conv2')(x)
x = BatchNormalization(axis=channel_axis, name='bn2')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(3, 3), name='pool2')(x)
x = Dropout(0.1, name='dropout2')(x)
# Conv block 3
x = Convolution2D(channel_size, 3, 1, padding='same', name='conv3')(x)
x = BatchNormalization(axis=channel_axis, name='bn3')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(3, min_size / 6),
strides=(3, min_size / 6),
name='pool3')(x)
x = Dropout(0.1, name='dropout3')(x)
# if min_size // 24 >= 4:
# # Conv block 4
# x = Convolution2D(
# channel_size,
# 3,
# 1,
# padding='same',
# name='conv4')(x)
# x = BatchNormalization(axis=channel_axis, name='bn4')(x)
# x = ELU()(x)
# x = MaxPooling2D(pool_size=(4, 4), strides=(4, 4), name='pool4')(x)
# x = Dropout(0.1, name='dropout4')(x)
x = Reshape((-1, channel_size))(x)
avg = GlobalAvgPool1D()(x)
max = GlobalMaxPool1D()(x)
x = concatenate([avg, max], axis=-1)
# x = Dense(max(int(num_classes*1.5), 128), activation='relu', name='dense1')(x)
x = Dropout(0.3)(x)
outputs = Dense(num_classes, activation='softmax', name='output')(x)
model = TFModel(inputs=melgram_input, outputs=outputs)
optimizer = optimizers.Nadam(lr=0.002,
beta_1=0.9,
beta_2=0.999,
epsilon=None,
schedule_decay=0.004)
model.compile(optimizer=optimizer,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, input_shape, num_classes, **kwargs):
freq_axis = 2
channel_axis = 3
channel_size = 128
min_size = min(input_shape[:2])
melgram_input = Input(shape=input_shape)
# x = ZeroPadding2D(padding=(0, 37))(melgram_input)
# x = BatchNormalization(axis=freq_axis, name='bn_0_freq')(x)
x = Reshape((input_shape[0], input_shape[1], 1))(melgram_input)
# Conv block 1
x = Convolution2D(64, 3, 1, padding='same', name='conv1')(x)
x = BatchNormalization(axis=channel_axis, name='bn1')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='pool1')(x)
x = Dropout(0.1, name='dropout1')(x)
# Conv block 2
x = Convolution2D(channel_size, 3, 1, padding='same', name='conv2')(x)
x = BatchNormalization(axis=channel_axis, name='bn2')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(3, 3), name='pool2')(x)
x = Dropout(0.1, name='dropout2')(x)
# Conv block 3
x = Convolution2D(channel_size, 3, 1, padding='same', name='conv3')(x)
x = BatchNormalization(axis=channel_axis, name='bn3')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(4, 4), strides=(4, 4), name='pool3')(x)
x = Dropout(0.1, name='dropout3')(x)
if min_size // 24 >= 4:
# Conv block 4
x = Convolution2D(channel_size, 3, 1, padding='same',
name='conv4')(x)
x = BatchNormalization(axis=channel_axis, name='bn4')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(4, 4), strides=(4, 4), name='pool4')(x)
x = Dropout(0.1, name='dropout4')(x)
x = Reshape((-1, channel_size))(x)
gru_units = 128
if num_classes > gru_units:
gru_units = int(num_classes * 1.5)
# GRU block 1, 2, output
x = CuDNNGRU(gru_units, return_sequences=True, name='gru1')(x)
x = CuDNNGRU(gru_units, return_sequences=False, name='gru2')(x)
# x = Dense(max(int(num_classes*1.5), 128), activation='relu', name='dense1')(x)
x = Dropout(0.3)(x)
outputs = Dense(num_classes, activation='softmax', name='output')(x)
model = TFModel(inputs=melgram_input, outputs=outputs)
optimizer = optimizers.Adam(
# learning_rate=1e-3,
lr=1e-3,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=1e-4,
amsgrad=True)
model.compile(optimizer=optimizer,
loss="sparse_categorical_crossentropy",
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
def init_model(self, input_shape, num_classes, **kwargs):
freq_axis = 2
channel_axis = 3
channel_size = 128
min_size = min(input_shape[:2])
inputs = Input(shape=input_shape)
# x = ZeroPadding2D(padding=(0, 37))(melgram_input)
# x = BatchNormalization(axis=freq_axis, name='bn_0_freq')(x)
x = Reshape((input_shape[0], input_shape[1], 1))(inputs)
# Conv block 1
x = Convolution2D(64, 3, 1, padding='same', name='conv1')(x)
x = BatchNormalization(axis=channel_axis, name='bn1')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(2, 2), strides=(2, 2), name='pool1')(x)
x = Dropout(0.1, name='dropout1')(x)
# Conv block 2
x = Convolution2D(channel_size, 3, 1, padding='same', name='conv2')(x)
x = BatchNormalization(axis=channel_axis, name='bn2')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(3, 3), name='pool2')(x)
x = Dropout(0.1, name='dropout2')(x)
# Conv block 3
x = Convolution2D(channel_size, 3, 1, padding='same', name='conv3')(x)
x = BatchNormalization(axis=channel_axis, name='bn3')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(4, 4), strides=(4, 4), name='pool3')(x)
x = Dropout(0.1, name='dropout3')(x)
if min_size // 24 >= 4:
# Conv block 4
x = Convolution2D(channel_size, 3, 1, padding='same',
name='conv4')(x)
x = BatchNormalization(axis=channel_axis, name='bn4')(x)
x = ELU()(x)
x = MaxPooling2D(pool_size=(4, 4), strides=(4, 4), name='pool4')(x)
x = Dropout(0.1, name='dropout4')(x)
x = Reshape((-1, channel_size))(x)
avg = GlobalAvgPool1D()(x)
max = GlobalMaxPool1D()(x)
x = concatenate([avg, max], axis=-1)
# x = Dense(max(int(num_classes*1.5), 128), activation='relu', name='dense1')(x)
x = Dropout(0.3)(x)
outputs1 = Dense(num_classes, activation='softmax', name='output')(x)
# bnorm_1 = BatchNormalization(axis=2)(inputs)
lstm_1 = Bidirectional(CuDNNLSTM(64,
name='blstm_1',
return_sequences=True),
merge_mode='concat')(inputs)
activation_1 = Activation('tanh')(lstm_1)
dropout1 = SpatialDropout1D(0.5)(activation_1)
attention_1 = Attention(8, 16)([dropout1, dropout1, dropout1])
pool_1 = GlobalMaxPool1D()(attention_1)
dropout2 = Dropout(rate=0.5)(pool_1)
dense_1 = Dense(units=256, activation='relu')(dropout2)
outputs2 = Dense(units=num_classes, activation='softmax')(dense_1)
outputs = Average()([outputs1, outputs2])
model = TFModel(inputs=inputs, outputs=outputs)
optimizer = optimizers.Adam(
# learning_rate=1e-3,
lr=1e-3,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0002,
amsgrad=True)
model.compile(optimizer=optimizer,
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
model.summary()
self._model = model
self.is_init = True
