8ec91c97eb64bbe3e5030248b08d3870a5b4cb60,theanolm/network/samplingoutputlayer.py,SamplingOutputLayer,_get_sample_tensors,#SamplingOutputLayer#Any#,97

Before Change

        num_sequences = layer_input.shape[1]
        num_samples = self._network.num_noise_samples
        num_classes = numpy.int64(self._network.vocabulary.num_classes())
        random = self._network.random

        minibatch_size = num_time_steps * num_sequences
        if self._network.noise_probs is None:
            // The upper bound is exclusive, so this always creates samples that
            // are < num_classes.
            num_batch_samples = minibatch_size * num_samples
            sample = random.uniform((num_batch_samples,)) * num_classes
            sample = sample.astype("int64")
        else:
            // We repeat the distribution for each mini-batch element, and sample
            // k noise words per mini-batch element. k < number of outpus, so
            // it"s possible without replacement.
            class_probs = self._network.noise_probs[None, :]
            class_probs = tensor.tile(class_probs, [minibatch_size, 1])
            // Since we sample different noise words for different data words, we
            // could set the probability of the correct data words to zero, as
            // suggested in the BlackOut paper. That seems to result in a little
            // bit worse model with NCE and BlackOut.
//            target_class_ids = self._network.target_class_ids.flatten()
//            target_sample_ids = tensor.arange(minibatch_size)
//            class_probs = tensor.set_subtensor(
//                class_probs[(target_sample_ids, target_class_ids)], 0)
//            denominators = class_probs.sum(1)
//            denominators = denominators[:, None]
//            class_probs /= denominators
            sample = multinomial(random, class_probs, num_samples)
            // For some reason (maybe a rounding error) it may happen that the
            // sample contains a very high or negative value.
            sample = tensor.maximum(sample, 0)
            sample = tensor.minimum(sample, num_classes - 1)

        sample = sample.reshape([num_time_steps, num_sequences, num_samples])
        return sample, self._get_target_preact(layer_input, sample)

    def _get_seqshared_sample_tensors(self, layer_input):

After Change

        num_sequences = layer_input.shape[1]
        num_samples = self._network.num_noise_samples
        num_classes = numpy.int64(self._network.vocabulary.num_classes())
        noise_sampler = self._network.noise_sampler

        minibatch_size = num_time_steps * num_sequences
        sample = noise_sampler.sample(minibatch_size, num_samples)
        sample = sample.reshape([num_time_steps, num_sequences, num_samples])
        return sample, self._get_target_preact(layer_input, sample)

    def _get_seqshared_sample_tensors(self, layer_input):

Instances