rn_rnn_char.py 文件源码

def buildmodelparams(self, hyper, checkpointdir=None):
        """Builds model parameters from given hyperparameters and charset size.
        Optionally saves checkpoint immediately after building if path specified.
        """
        useclass = self.modeltypes[self.modeltype]
        self.model = useclass(hyper)

        if checkpointdir:
            # Compile training functions
            self.model._build_t()

            # Get initial loss estimate
            stderr.write("Calculating initial loss estimate...\n")

            # We don't need anything fancy or long, just a rough baseline
            data_len = self.valid.batchepoch(16)
            loss_len = 20 if data_len >= 20 else data_len
            loss = self.model.calc_loss(self.valid, 0, batchsize=8, num_examples=loss_len)

            stderr.write("Initial loss: {0:.3f}\n".format(loss))
            stderr.write("Initial log loss: {0:.3f}\n".format(log(loss)))

            # Take checkpoint
            self.newcheckpoint(loss, savedir=checkpointdir)