assert False, 'no activation function' def cost_func(y, y_hat, batch_size, choice='euclidean'): if choice == 'euclidean': return T.sum((y-y_hat)**2) / batch_size else: assert False, 'no cost function' t_start = time.time() sample_file = 'mfcc/train.ark' test_file = 'mfcc/test.ark' label_file = 'label/train.lab' label_map_file = 'phones/48_39.map' DataParser.load(sample_file, label_file, label_map_file) #DataParser.test() dim_x = DataParser.dimension_x dim_y_hat = DataParser.dimension_y batch_size = 21 neuron_num = 64 epoch_cycle = 50 learning_rate = 0.01 lr = theano.shared(learning_rate) lr_decay = 1.0 # e.g. matrix 3*2 dot matrix 2*1 = matrix 3*1 # [[1., 3.], [2., 2.], [3.,1.]] dot [[2], [1]] = [[5.], [6.], [7.]] x = T.matrix('input', dtype='float64') # matrix of dim_x * batch_size y_hat = T.matrix('reference', dtype='float64') # matrix of dim_y_hat * batch_size