def _lstm_init_params(self, params):
input_size = hidden_size = output_size = self.word_img_embed_hidden_dim
# Recurrent weights: take x_t, h_{t-1}, and bias unit
# and produce the 3 gates and the input to cell signal
params['WLSTM'] = initw(input_size + hidden_size, 4 * hidden_size).astype(config.floatX)
params['bLSTM'] = np.zeros(4 * hidden_size).astype(config.floatX)
# Decoder weights (e.g. mapping to vocabulary)
params['Wd'] = initw(hidden_size, self.vocab_size).astype(config.floatX)
params['bd'] = np.zeros(self.vocab_size).astype(config.floatX)
def init(input_size, hidden_size, output_size):
model = {}
model['WLSTM'] = initw(input_size + hidden_size + 1, 4 * hidden_size)
model['Wd'] = initw(hidden_size, output_size)
model['bd'] = np.zeros((1, output_size))
update = ['WLSTM', 'Wd', 'bd']
return {'model': model, 'update': update}
def init_params(self, model_options): self.image_size = 4096 # size of CNN vectors hardcoded here self.word_img_embed_hidden_dim = model_options['word_img_embed_hidden_dim'] self.vocab_size = model_options['vocab_size'] # Dict name (string) -> numpy ndarray params = OrderedDict() # Ws -> word encoder params['Ws'] = initw(self.vocab_size, self.word_img_embed_hidden_dim).astype(config.floatX) # We, be -> image encoder params['We'] = initw(self.image_size, self.word_img_embed_hidden_dim).astype(config.floatX) params['be'] = initw(self.word_img_embed_hidden_dim).astype(config.floatX) self._lstm_init_params(params) return params
