Python torch.nn.parallel._functions.Broadcast.apply() Examples

The following are 30 code examples of torch.nn.parallel._functions.Broadcast.apply(). You can vote up the ones you like or vote down the ones you don't like, and go to the original project or source file by following the links above each example. You may also want to check out all available functions/classes of the module torch.nn.parallel._functions.Broadcast , or try the search function .
Example #1
Source File: batchnorm.py    From BraTS-DMFNet with Apache License 2.0 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #2
Source File: module.py    From openseg.pytorch with MIT License 6 votes vote down vote up 
def batchnormtrain(input, mean, std, gamma, beta):
    r"""Applies Batch Normalization over a 3d input that is seen as a
    mini-batch.

    .. _encoding.batchnormtrain:

    .. math::

        y = \frac{x - \mu[x]}{ \sqrt{var[x] + \epsilon}} * \gamma + \beta

    Shape:
        - Input: :math:`(N, C)` or :math:`(N, C, L)`
        - Output: :math:`(N, C)` or :math:`(N, C, L)` (same shape as input)

    """
    return _batchnormtrain.apply(input, mean, std, gamma, beta)
Example #3
Source File: batchnorm.py    From deeplabv3plus-pytorch with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #4
Source File: batchnorm.py    From Deeplab-v3plus with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #5
Source File: batchnorm.py    From semantic-segmentation-pytorch with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)

        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #6
Source File: batchnorm.py    From BEAL with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #7
Source File: module.py    From openseg.pytorch with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #8
Source File: batchnorm.py    From ViewAL with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #9
Source File: batchnorm.py    From pytorch-deeplab-xception with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #10
Source File: syncbn.py    From TorchSeg with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #11
Source File: wideresnet_utils.py    From LaSO with BSD 3-Clause "New" or "Revised" License 6 votes vote down vote up 
def data_parallel(f, input, params, mode, device_ids, output_device=None):
    assert isinstance(device_ids, list)
    if output_device is None:
        output_device = device_ids[0]

    if len(device_ids) == 1:
        return f(input, params, mode)

    params_all = Broadcast.apply(device_ids, *params.values())
    params_replicas = [{k: params_all[i + j*len(params)] for i, k in enumerate(params.keys())}
                       for j in range(len(device_ids))]

    replicas = [partial(f, params=p, mode=mode)
                for p in params_replicas]
    inputs = scatter([input], device_ids)
    outputs = parallel_apply(replicas, inputs)
    return gather(outputs, output_device)
Example #12
Source File: batchnorm.py    From kaggle-understanding-clouds with BSD 2-Clause "Simplified" License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #13
Source File: batchnorm.py    From pytorch_segmentation with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #14
Source File: batchnorm.py    From lidar-bonnetal with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
    """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

    # Always using same "device order" makes the ReduceAdd operation faster.
    # Thanks to:: Tete Xiao (http://tetexiao.com/)
    intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

    to_reduce = [i[1][:2] for i in intermediates]
    to_reduce = [j for i in to_reduce for j in i]  # flatten
    target_gpus = [i[1].sum.get_device() for i in intermediates]

    sum_size = sum([i[1].sum_size for i in intermediates])
    sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
    mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

    broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

    outputs = []
    for i, rec in enumerate(intermediates):
      outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

    return outputs
Example #15
Source File: batchnorm.py    From reid_baseline_with_syncbn with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #16
Source File: syncbn.py    From TreeFilter-Torch with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #17
Source File: batchnorm.py    From fast-reid with Apache License 2.0 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #18
Source File: batchnorm.py    From pytorch_connectomics with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #19
Source File: batchnorm.py    From Mobile-Image-Matting with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #20
Source File: batchnorm.py    From PointWeb with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #21
Source File: batchnorm.py    From overhaul-distillation with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #22
Source File: syncbn.py    From FNA with Apache License 2.0 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #23
Source File: batchnorm.py    From PVN3D with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #24
Source File: batchnorm.py    From CAG_UDA with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2])))

        return outputs
Example #25
Source File: batchnorm.py    From CAG_UDA with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #26
Source File: batchnorm.py    From CAG_UDA with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #27
Source File: parallel.py    From structure_knowledge_distillation with BSD 2-Clause "Simplified" License 6 votes vote down vote up 
def forward(self, inputs, *targets, **kwargs):
        if not self.device_ids:
            return self.module(inputs, *targets, **kwargs)

        is_target_scattered = kwargs.get('is_target_scattered', False)
        kwargs.pop('is_target_scattered', None)  # this key is unexpected

        if not is_target_scattered:
            targets, kwargs = self.scatter(targets, kwargs, self.device_ids)

        if len(self.device_ids) == 1:
            if is_target_scattered:
                targets = (targets,)
                kwargs = (kwargs,)
            return self.module(inputs, *targets[0], **kwargs[0])

        if is_target_scattered:
            targets = targets[0]

        replicas = self.replicate(self.module, self.device_ids[:len(inputs)])
        outputs = _criterion_parallel_apply(replicas, inputs, targets)
        return Reduce.apply(*outputs) / len(outputs)
Example #28
Source File: batchnorm.py    From bonnetal with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
    """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

    # Always using same "device order" makes the ReduceAdd operation faster.
    # Thanks to:: Tete Xiao (http://tetexiao.com/)
    intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

    to_reduce = [i[1][:2] for i in intermediates]
    to_reduce = [j for i in to_reduce for j in i]  # flatten
    target_gpus = [i[1].sum.get_device() for i in intermediates]

    sum_size = sum([i[1].sum_size for i in intermediates])
    sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
    mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

    broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

    outputs = []
    for i, rec in enumerate(intermediates):
      outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

    return outputs
Example #29
Source File: batchnorm.py    From RMI with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs
Example #30
Source File: syncbn.py    From RMI with MIT License 6 votes vote down vote up 
def _data_parallel_master(self, intermediates):
        """Reduce the sum and square-sum, compute the statistics, and broadcast it."""

        # Always using same "device order" makes the ReduceAdd operation faster.
        # Thanks to:: Tete Xiao (http://tetexiao.com/)
        intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device())

        to_reduce = [i[1][:2] for i in intermediates]
        to_reduce = [j for i in to_reduce for j in i]  # flatten
        target_gpus = [i[1].sum.get_device() for i in intermediates]

        sum_size = sum([i[1].sum_size for i in intermediates])
        sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce)
        mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size)

        broadcasted = Broadcast.apply(target_gpus, mean, inv_std)

        outputs = []
        for i, rec in enumerate(intermediates):
            outputs.append((rec[0], _MasterMessage(*broadcasted[i*2:i*2+2])))

        return outputs