l2_layers = []
for layer in lasagne.layers.get_all_layers(output_layer):
if isinstance(layer,(DCNN.embeddings.SentenceEmbeddingLayer,DCNN.convolutions.Conv1DLayerSplitted,lasagne.layers.DenseLayer)):
l2_layers.append(layer)
loss_train = lasagne.objectives.aggregate(lasagne.objectives.categorical_crossentropy(lasagne.layers.get_output(output_layer),y_batch),mode='mean') \
+ lasagne.regularization.regularize_layer_params_weighted(dict(zip(l2_layers,hyperparas["L2"])),lasagne.regularization.l2)
# validating/testing
loss_eval = lasagne.objectives.categorical_crossentropy(lasagne.layers.get_output(output_layer, deterministic=True),y_batch)
pred = T.argmax(lasagne.layers.get_output(output_layer , deterministic=True),axis=1)
correct_predictions = T.eq(pred, y_batch)
# In the matlab code, Kalchbrenner works with a adagrad reset mechanism, if the para --adagrad_reset has value 0, no reset will be applied
all_params = lasagne.layers.get_all_params(output_layer)
updates, accumulated_grads = utils.adagrad(loss_train, all_params, hyperparas['learning_rate'])
#updates = lasagne.updates.adagrad(loss_train, all_params, hyperparas['learning_rate'])
train_model = theano.function(inputs=[SENT1_batch,SENT2_batch,y_batch], outputs=loss_train,updates=updates, on_unused_input='warn')
valid_model = theano.function(inputs=[SENT1_batch,SENT2_batch,y_batch], outputs=correct_predictions, on_unused_input='warn')
test_model = theano.function(inputs=[SENT1_batch,SENT2_batch,y_batch], outputs=correct_predictions, on_unused_input='warn')
###############
# TRAIN MODEL #
###############
print('Started training')
# Kalchbrenner uses a fine-grained L2 regularization in the Matlab code, default values taken from Matlab code
# Training objective
l2_layers = []
for layer in lasagne.layers.get_all_layers(output_layer):
if isinstance(layer,(DCNN.embeddings.SentenceEmbeddingLayer,DCNN.convolutions.Conv1DLayerSplitted,lasagne.layers.DenseLayer)):
l2_layers.append(layer)
loss_train = lasagne.objectives.aggregate(lasagne.objectives.categorical_crossentropy(lasagne.layers.get_output(output_layer,X_batch),y_batch),mode='mean')+lasagne.regularization.regularize_layer_params_weighted(dict(zip(l2_layers,hyperparas["L2"])),lasagne.regularization.l2)
# validating/testing
loss_eval = lasagne.objectives.categorical_crossentropy(lasagne.layers.get_output(output_layer,X_batch,deterministic=True),y_batch)
pred = T.argmax(lasagne.layers.get_output(output_layer, X_batch, deterministic=True),axis=1)
correct_predictions = T.eq(pred, y_batch)
# In the matlab code, Kalchbrenner works with a adagrad reset mechanism, if the para --adagrad_reset has value 0, no reset will be applied
all_params = lasagne.layers.get_all_params(output_layer)
updates, accumulated_grads = utils.adagrad(loss_train, all_params, hyperparas['learning_rate'])
#updates = lasagne.updates.adagrad(loss_train, all_params, hyperparas['learning_rate'])
train_model = theano.function(inputs=[X_batch,y_batch], outputs=loss_train,updates=updates)
valid_model = theano.function(inputs=[X_batch,y_batch], outputs=correct_predictions)
test_model = theano.function(inputs=[X_batch,y_batch], outputs=correct_predictions)
###############
# TRAIN MODEL #
###############
print('Started training')