1f5bd7028636e6320079fc785c4aca9eda51defa,niftynet/utilities/util_common.py,,average_multi_opt_gradients,#Any#,43

Before Change

            single_opt_multi_device_grad = list()
            // iterate over the gradient dicts of all devices
            for device_gradient in multi_device_gradients:
                if not isinstance(device_gradient, dict):
                    tf.logging.fatal("gradient type miss match")
                    raise ValueError
                if device_gradient.get(optimiser) is not None:
                    single_opt_multi_device_grad.append(device_gradient.get(optimiser))
            averaged_gradients[optimiser] = average_gradients(single_opt_multi_device_grad)
        return averaged_gradients
    return average_gradients(multi_device_gradients)

After Change

        optimiser_names = sorted(multi_device_gradients[0])
        ave_gradients = dict()
        for opt_name in optimiser_names:
            multi_device_grad = [device_gradient.get(opt_name)
                                 for device_gradient in multi_device_gradients]
            ave_gradients[opt_name] = average_gradients(multi_device_grad)
        return ave_gradients
    // multi_device_gradients is a list of N device_gradients, for N devices
    return average_gradients(multi_device_gradients)

Instances