16d2eb3061f1bed8ade390c5e2a2c1de9daa3509,theanolm/optimizers/basicoptimizer.py,BasicOptimizer,__init__,#BasicOptimizer#Any#Any#Any#,16

Before Change

            total_sum = sum(sums)  // sum over parameter variables
            norm = tensor.sqrt(total_sum)
            target_norm = tensor.clip(norm, 0.0, max_norm)
            gradients = [gradient * target_norm / (epsilon + norm)
                         for gradient in gradients]

        self._gradient_exprs = gradients
        self.gradient_update_function = \
            theano.function([self.network.minibatch_input,

After Change

                       for name, value in self.param_init_values.items()}

        // numerical stability / smoothing term to prevent divide-by-zero
        if not "epsilon" in optimization_options:
            raise ValueError("Epsilon is not given in optimization options.")
        self._epsilon = optimization_options["epsilon"]

        // maximum norm for parameter updates
        if "max_gradient_norm" in optimization_options:
            self._max_gradient_norm = optimization_options["max_gradient_norm"]
        else:
            self._max_gradient_norm = None

        // Derive the symbolic expression for log probability of each word.
        logprobs = tensor.log(self.network.prediction_probs)
        // Set the log probability to 0, if the next input word (the one

Instances