Python utils.dict_subset方法代码示例

# 需要导入模块: from blocks import utils [as 别名]
# 或者: from blocks.utils import dict_subset [as 别名]
def process_batch(self, batch):
        try:
            input_names = [v.name for v in self.unique_inputs]
            batch = dict_subset(batch, input_names)
        except KeyError:
            reraise_as(
                "Not all data sources required for monitoring were"
                " provided. The list of required data sources:"
                " {}.".format(input_names))
        if self._accumulate_fun is not None:
            numerical_values = self._accumulate_fun(**batch)
            for value, var in zip(numerical_values,self.theano_variables):
                self.data[var.name].append(value) 

# 需要导入模块: from blocks import utils [as 别名]
# 或者: from blocks.utils import dict_subset [as 别名]
def cost_matrix(self, application_call, outputs, mask=None, **kwargs):
        """Adapted from ``BaseSequenceGenerator.cost_matrix``
        """
        # We assume the data has axes (time, batch, features, ...)
        batch_size = outputs.shape[1]

        # Prepare input for the iterative part
        states = dict_subset(kwargs, self._state_names, must_have=False)
        # masks in context are optional (e.g. `attended_mask`)
        contexts = dict_subset(kwargs, self._context_names, must_have=False)
        feedback = self.readout.feedback(outputs)
        inputs = self.fork.apply(feedback, as_dict=True)

        # Run the recurrent network
        results = self.transition.apply(
            mask=mask, return_initial_states=True, as_dict=True,
            **dict_union(inputs, states, contexts))

        # Separate the deliverables. The last states are discarded: they
        # are not used to predict any output symbol. The initial glimpses
        # are discarded because they are not used for prediction.
        # Remember, glimpses are computed _before_ output stage, states are
        # computed after.
        states = {name: results[name][:-1] for name in self._state_names}
        glimpses = {name: results[name][1:] for name in self._glimpse_names}

        # Compute the cost
        feedback = tensor.roll(feedback, 1, 0)
        feedback = tensor.set_subtensor(
            feedback[0],
            self.readout.feedback(self.readout.initial_outputs(batch_size)))
        readouts = self.readout.readout(
            feedback=feedback, **dict_union(states, glimpses, contexts))
        costs = self.readout.cost(readouts, outputs)
        if mask is not None:
            costs *= mask

        for name, variable in list(glimpses.items()) + list(states.items()):
            application_call.add_auxiliary_variable(
                variable.copy(), name=name)

        # This variables can be used to initialize the initial states of the
        # next batch using the last states of the current batch.
        for name in self._state_names:
            application_call.add_auxiliary_variable(
                results[name][-1].copy(), name=name+"_final_value")

        if not self.pruning_variables_initialized:
            self.results = results
            self.pruning_variables_initialized = True
        return costs 

# 需要导入模块: from blocks import utils [as 别名]
# 或者: from blocks.utils import dict_subset [as 别名]
def cost_matrix_nmt(self, application_call, target_char_seq, target_sample_matrix, target_resample_matrix,
                        target_word_mask, target_char_aux, target_prev_char_seq, target_prev_char_aux, **kwargs):
        """Returns generation costs for output sequences.

        See Also
        --------
        :meth:`cost` : Scalar cost.

        """
        # We assume the data has axes (time, batch, features, ...)
        batch_size = target_char_seq.shape[1]

        # Prepare input for the iterative part
        states = dict_subset(kwargs, self._state_names, must_have=False)
        # masks in context are optional (e.g. `attended_mask`)
        contexts = dict_subset(kwargs, self._context_names, must_have=False)
        feedback = self.readout.feedback_apply(target_char_seq, target_sample_matrix, target_char_aux)
        inputs = self.fork.apply(feedback, as_dict=True)

        # Run the recurrent network
        results = self.transition.apply(
            mask=target_word_mask, return_initial_states=True, as_dict=True,
            **dict_union(inputs, states, contexts))

        # Separate the deliverables. The last states are discarded: they
        # are not used to predict any output symbol. The initial glimpses
        # are discarded because they are not used for prediction.
        # Remember, glimpses are computed _before_ output stage, states are
        # computed after.
        states = {name: results[name][:-1] for name in self._state_names}
        glimpses = {name: results[name][1:] for name in self._glimpse_names}

        feedback = tensor.roll(feedback, 1, 0)
        init_feedback = self.readout.single_feedback(self.readout.initial_outputs(batch_size), batch_size)
        if self.trg_dgru_depth == 1:
            feedback = tensor.set_subtensor(feedback[0], init_feedback)
        else:
            feedback = tensor.set_subtensor(feedback[0], init_feedback[-1])

        decoder_readout_outputs = self.readout.readout(
            feedback=feedback, **dict_union(states, glimpses, contexts))
        resampled_representation = tensor.batched_dot(target_resample_matrix,
                                                      decoder_readout_outputs.dimshuffle([1, 0, 2]))
        resampled_readouts = resampled_representation.dimshuffle([1, 0, 2])
        readouts_chars = self.readout.readout_gru(target_prev_char_seq, target_prev_char_aux, resampled_readouts)

        # Compute the cost
        costs = self.readout.cost(readouts_chars, target_char_seq)

        for name, variable in list(glimpses.items()) + list(states.items()):
            application_call.add_auxiliary_variable(
                variable.copy(), name=name)

        # This variables can be used to initialize the initial states of the
        # next batch using the last states of the current batch.
        for name in self._state_names + self._glimpse_names:
            application_call.add_auxiliary_variable(
                results[name][-1].copy(), name=name + "_final_value")

        return costs