inputs=[cnn_features_idx, caption_begin, caption_end],
outputs=loss,
updates=updates,
givens={
cnn_features:
train_data_iterator.features[cnn_features_idx],
word_indices:
train_data_iterator.flatten_input_captions[caption_begin:caption_end],
true_dist:
train_data_iterator.flatten_output_captions[caption_begin:caption_end]
})
# define valid model
val_word_embedding_layer = EmbeddingLayer.create_copy(word_embedding_layer)
val_cnn_embedding_layer = DenseLayer.create_copy(cnn_embedding_layer)
val_row_stack_layer = RowStackLayer.create_copy(row_stack_layer)
val_embedding_scale_layer = ScaleLayer('embedding_scale',
embedding_dropout_layer.keep_prob)
val_lstm_layer = LstmLayer.create_copy(lstm_layer)
val_hidden_states_scale_layer = ScaleLayer(
'hidden_states_scale', hidden_states_dropout_layer.keep_prob)
val_pre_softmax_layer = DenseLayer.create_copy(pre_softmax_layer)
val_softmax_layer = NonlinearityLayer.create_copy(softmax_layer)
# define forward propagation expression for val model and loss function
val_word_embedings = val_word_embedding_layer.get_output_expr(word_indices)
val_cnn_embedings = val_cnn_embedding_layer.get_output_expr(cnn_features)
val_embedings = val_row_stack_layer.get_output_expr(val_cnn_embedings,
val_word_embedings)
val_scaled_embedings = val_embedding_scale_layer.get_output_expr(val_embedings)
val_h = val_lstm_layer.get_output_expr(val_scaled_embedings)
train_model = theano.function(
inputs=[cnn_features_idx, caption_begin, caption_end],
outputs=loss,
updates=updates,
givens={
cnn_features: train_data_iterator.features[cnn_features_idx],
word_indices: train_data_iterator.flatten_input_captions[caption_begin:caption_end],
true_dist: train_data_iterator.flatten_output_captions[caption_begin:caption_end]
}
)
# define valid model
val_word_embedding_layer = EmbeddingLayer.create_copy(word_embedding_layer)
val_cnn_embedding_layer = DenseLayer.create_copy(cnn_embedding_layer)
val_row_stack_layer = RowStackLayer.create_copy(row_stack_layer)
val_embedding_scale_layer = ScaleLayer('embedding_scale', embedding_dropout_layer.keep_prob)
val_lstm_layer = LstmLayer.create_copy(lstm_layer)
val_hidden_states_scale_layer = ScaleLayer('hidden_states_scale', hidden_states_dropout_layer.keep_prob)
val_pre_softmax_layer = DenseLayer.create_copy(pre_softmax_layer)
val_softmax_layer = NonlinearityLayer.create_copy(softmax_layer)
# define forward propagation expression for val model and loss function
val_word_embedings = val_word_embedding_layer.get_output_expr(word_indices)
val_cnn_embedings = val_cnn_embedding_layer.get_output_expr(cnn_features)
val_embedings = val_row_stack_layer.get_output_expr(val_cnn_embedings, val_word_embedings)
val_scaled_embedings = val_embedding_scale_layer.get_output_expr(val_embedings)
val_h = val_lstm_layer.get_output_expr(val_scaled_embedings)
val_scaled_h = val_hidden_states_scale_layer.get_output_expr(val_h[1:])
val_unnormalized_probs = val_pre_softmax_layer.get_output_expr(val_scaled_h)
