def RNNmodel(self):
size_1 = self.model_setting['size1']
size_2 = self.model_setting['size2']
p = self.model_setting['dropout']
gru1 = Bidirectional(CuDNNGRU(size_1, return_sequences=True),
merge_mode='ave')
gru2 = Bidirectional(CuDNNGRU(size_2, return_sequences=True),
merge_mode='ave')
layer = {
'gru1': gru1,
'gru2': gru2,
}
embedding_layer = SpatialDropout1D(p)(self.embedding_layer)
x = gru1(embedding_layer)
x = gru2(x) # 200 * 80
x1 = GlobalMaxPooling1D()(x)
x2 = GlobalAveragePooling1D()(x)
self.cat_layers += [x1, x2]
y = Concatenate()(self.cat_layers)
# y = Dense(90,activation='relu')(y)
fc = Dense(6)(y)
result_layer = Activation(activation='sigmoid')(fc)
if self.boost:
result_layer = concatenate([result_layer, self.boost_layer],
axis=-1)
result_layer = boostLayer()(result_layer)
self.result_model = Model(inputs=self.inputs, outputs=result_layer)
print(self.result_model.summary())
if self.lossname == 'rankLoss':
loss_layer = fc
else:
loss_layer = result_layer
self.set_loss(loss_layer)
if self.load_weight:
name = self.lossname
name = 'focalLoss'
gru1_weight = np.load(WEIGHT_FILE + name + 'gru1_weight.npy')
gru1.set_weights(gru1_weight)
gru2_weight = np.load(WEIGHT_FILE + name + 'gru2_weight.npy')
gru2.set_weights(gru2_weight)
return layer
def lrpify_model(model):
'''
This function takes as input user defined keras Model, and replace all LSTM/Bi_LSTM
with equivalent one which have LRP_LSTMCell as core cell
'''
cell_config_keys = ['units', 'activation', 'recurrent_activation', 'use_bias', 'unit_forget_bias', 'kernel_constraint', 'recurrent_constraint', 'bias_constraint']
rnn_config_keys = ['return_sequences', 'return_state', 'go_backwards', 'stateful', 'unroll']
bidirect_config_keys = ['merge_mode']
for i,layer in enumerate(model.layers):
if isinstance(layer,Bidirectional):
weights = layer.get_weights()
inp_shape = layer.input_shape
cell_config = {key:layer.get_config()['layer']['config'][key] for key in cell_config_keys}
rnn_config = {key:layer.get_config()['layer']['config'][key] for key in rnn_config_keys}
bidirect_config = {key:layer.get_config()[key] for key in bidirect_config_keys}
with CustomObjectScope({'LRP_LSTMCell': LRP_LSTMCell}):
cell = LRP_LSTMCell(**cell_config, implementation=1)
bi_lstm = Bidirectional(RNN(cell,**rnn_config),**bidirect_config)
bi_lstm.build(inp_shape)
bi_lstm.call(layer.input)
bi_lstm._inbound_nodes = layer._inbound_nodes
bi_lstm._outbound_nodes = layer._outbound_nodes
bi_lstm.set_weights(weights)
model.layers[i] = bi_lstm
if isinstance(layer,LSTM):
weights = layer.get_weights()
inp_shape = layer.input_shape
cell_config = {key:layer.get_config()[key] for key in cell_config_keys}
rnn_config = {key:layer.get_config()[key] for key in rnn_config_keys}
with CustomObjectScope({'LRP_LSTMCell': LRP_LSTMCell}):
cell = LRP_LSTMCell(**cell_config,implementation=1)
lstm = RNN(cell,**rnn_config)
lstm.build(inp_shape)
lstm.call(layer.input)
lstm.set_weights(weights)
lstm._inbound_nodes = layer._inbound_nodes
lstm._outbound_nodes = layer._outbound_nodes
model.layers[i] = lstm
return model