Python comm.reduce_add方法代碼示例

本文整理匯總了Python中torch.cuda.comm.reduce_add方法的典型用法代碼示例。如果您正苦於以下問題：Python comm.reduce_add方法的具體用法？Python comm.reduce_add怎麽用？Python comm.reduce_add使用的例子？那麽, 這裏精選的方法代碼示例或許可以為您提供幫助。您也可以進一步了解該方法所在類torch.cuda.comm的用法示例。

在下文中一共展示了comm.reduce_add方法的5個代碼示例，這些例子默認根據受歡迎程度排序。您可以為喜歡或者感覺有用的代碼點讚，您的評價將有助於係統推薦出更棒的Python代碼示例。

示例1: backward

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import reduce_add [as 別名]
def backward(ctx, dz):
        z, var, weight, bias = ctx.saved_tensors
        dz = dz.contiguous()

        # Undo activation
        _act_backward(ctx, z, dz)

        if ctx.training:
            edz, eydz = _backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps)

            if ctx.is_master:
                edzs, eydzs = [edz], [eydz]
                for _ in range(len(ctx.worker_queues)):
                    edz_w, eydz_w = ctx.master_queue.get()
                    ctx.master_queue.task_done()
                    edzs.append(edz_w)
                    eydzs.append(eydz_w)

                edz = comm.reduce_add(edzs) / (ctx.master_queue.maxsize + 1)
                eydz = comm.reduce_add(eydzs) / (ctx.master_queue.maxsize + 1)

                tensors = comm.broadcast_coalesced((edz, eydz), [edz.get_device()] + ctx.worker_ids)
                for ts, queue in zip(tensors[1:], ctx.worker_queues):
                    queue.put(ts)
            else:
                ctx.master_queue.put((edz, eydz))
                edz, eydz = ctx.worker_queue.get()
                ctx.worker_queue.task_done()
        else:
            edz = dz.new_zeros(dz.size(1))
            eydz = dz.new_zeros(dz.size(1))

        dx, dweight, dbias = _backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps)
        dweight = dweight if ctx.affine else None
        dbias = dbias if ctx.affine else None

        return dx, dweight, dbias, None, None, None, None, None, None, None, None

開發者ID:miraiaroha，項目名稱:ACAN，代碼行數:39，代碼來源:functions.py

示例2: forward

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import reduce_add [as 別名]
def forward(ctx, *inputs):
        ctx.target_gpus = [inputs[i].get_device() for i in range(len(inputs))]
        inputs = sorted(inputs, key=lambda i: i.get_device())
        return comm.reduce_add(inputs)

開發者ID:speedinghzl，項目名稱:pytorch-segmentation-toolbox，代碼行數:6，代碼來源:encoding.py

示例3: backward

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import reduce_add [as 別名]
def backward(ctx, dz):
        x, weight, bias, mean, var = ctx.saved_tensors
        dz = dz.contiguous()

        # 1. compute \sum(\frac{dJ}{dy_i}) and \sum(\frac{dJ}{dy_i}*\hat{x_i})
        sum_dz, sum_dz_xhat = _backend.syncbn_backward_xhat(
            dz, x, mean, var, ctx.eps)
        if ctx.is_master:
            sum_dzs, sum_dz_xhats = [sum_dz], [sum_dz_xhat]
            # master : gatther from slaves
            for _ in range(ctx.master_queue.maxsize):
                sum_dz_w, sum_dz_xhat_w = ctx.master_queue.get()
                ctx.master_queue.task_done()
                sum_dzs.append(sum_dz_w)
                sum_dz_xhats.append(sum_dz_xhat_w)
            # master : compute global stats
            sum_dz = comm.reduce_add(sum_dzs)
            sum_dz_xhat = comm.reduce_add(sum_dz_xhats)
            sum_dz /= ctx.N
            sum_dz_xhat /= ctx.N
            # master : broadcast global stats
            tensors = comm.broadcast_coalesced(
                (sum_dz, sum_dz_xhat), [mean.get_device()] + ctx.worker_ids)
            for ts, queue in zip(tensors[1:], ctx.worker_queues):
                queue.put(ts)
        else:
            # slave : send to master
            ctx.master_queue.put((sum_dz, sum_dz_xhat))
            # slave : get global stats
            sum_dz, sum_dz_xhat = ctx.worker_queue.get()
            ctx.worker_queue.task_done()

        # do batch norm backward
        dx, dweight, dbias = _backend.syncbn_backward(
            dz, x, weight, bias, mean, var, sum_dz, sum_dz_xhat,
            ctx.affine, ctx.eps)

        return dx, dweight, dbias, \
            None, None, None, None, None, None

開發者ID:tamakoji，項目名稱:pytorch-syncbn，代碼行數:41，代碼來源:syncbn.py

示例4: backward

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import reduce_add [as 別名]
def backward(ctx, dz):
        z, var, weight, bias = ctx.saved_tensors
        dz = dz.contiguous()

        # Undo activation
        _act_backward(ctx, z, dz)

        if ctx.training:
            edz, eydz = backend.edz_eydz(z, dz, weight, bias, ctx.affine, ctx.eps)

            if ctx.is_master:
                edzs, eydzs = [edz], [eydz]
                for _ in range(len(ctx.worker_queues)):
                    edz_w, eydz_w = ctx.master_queue.get()
                    ctx.master_queue.task_done()
                    edzs.append(edz_w)
                    eydzs.append(eydz_w)

                edz = comm.reduce_add(edzs) / (ctx.master_queue.maxsize + 1)
                eydz = comm.reduce_add(eydzs) / (ctx.master_queue.maxsize + 1)

                tensors = comm.broadcast_coalesced((edz, eydz), [edz.get_device()] + ctx.worker_ids)
                for ts, queue in zip(tensors[1:], ctx.worker_queues):
                    queue.put(ts)
            else:
                ctx.master_queue.put((edz, eydz))
                edz, eydz = ctx.worker_queue.get()
                ctx.worker_queue.task_done()
        else:
            edz = dz.new_zeros(dz.size(1))
            eydz = dz.new_zeros(dz.size(1))

        dx, dweight, dbias = backend.backward(z, dz, var, weight, bias, edz, eydz, ctx.affine, ctx.eps)
        dweight = dweight if ctx.affine else None
        dbias = dbias if ctx.affine else None

        return dx, dweight, dbias, None, None, None, None, None, None, None, None

開發者ID:BloodAxe，項目名稱:segmentation-networks-benchmark，代碼行數:39，代碼來源:functions.py

示例5: forward

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import reduce_add [as 別名]
def forward(ctx, x, weight, bias, running_mean, running_var,
                extra, compute_stats=True, momentum=0.1, eps=1e-05):
        def _parse_extra(ctx, extra):
            ctx.is_master = extra["is_master"]
            if ctx.is_master:
                ctx.master_queue = extra["master_queue"]
                ctx.worker_queues = extra["worker_queues"]
                ctx.worker_ids = extra["worker_ids"]
            else:
                ctx.master_queue = extra["master_queue"]
                ctx.worker_queue = extra["worker_queue"]
        # Save context
        if extra is not None:
            _parse_extra(ctx, extra)
        ctx.compute_stats = compute_stats
        ctx.momentum = momentum
        ctx.eps = eps
        ctx.affine = weight is not None and bias is not None
        if ctx.compute_stats:
            N = _count_samples(x) * (ctx.master_queue.maxsize + 1)
            assert N > 1
            # 1. compute sum(x) and sum(x^2)
            xsum, xsqsum = _backend.syncbn_sum_sqsum(x.detach())
            if ctx.is_master:
                xsums, xsqsums = [xsum], [xsqsum]
                # master : gatther all sum(x) and sum(x^2) from slaves
                for _ in range(ctx.master_queue.maxsize):
                    xsum_w, xsqsum_w = ctx.master_queue.get()
                    ctx.master_queue.task_done()
                    xsums.append(xsum_w)
                    xsqsums.append(xsqsum_w)
                xsum = comm.reduce_add(xsums)
                xsqsum = comm.reduce_add(xsqsums)
                mean = xsum / N
                sumvar = xsqsum - xsum * mean
                var = sumvar / N
                uvar = sumvar / (N - 1)
                # master : broadcast global mean, variance to all slaves
                tensors = comm.broadcast_coalesced(
                    (mean, uvar, var), [mean.get_device()] + ctx.worker_ids)
                for ts, queue in zip(tensors[1:], ctx.worker_queues):
                    queue.put(ts)
            else:
                # slave : send sum(x) and sum(x^2) to master
                ctx.master_queue.put((xsum, xsqsum))
                # slave : get global mean and variance
                mean, uvar, var = ctx.worker_queue.get()
                ctx.worker_queue.task_done()

            # Update running stats
            running_mean.mul_((1 - ctx.momentum)).add_(ctx.momentum * mean)
            running_var.mul_((1 - ctx.momentum)).add_(ctx.momentum * uvar)
            ctx.N = N
            ctx.save_for_backward(x, weight, bias, mean, var)
        else:
            mean, var = running_mean, running_var

        # do batch norm forward
        z = _backend.syncbn_forward(x, weight, bias, mean, var,
                                    ctx.affine, ctx.eps)
        return z

開發者ID:tamakoji，項目名稱:pytorch-syncbn，代碼行數:63，代碼來源:syncbn.py

注：本文中的torch.cuda.comm.reduce_add方法示例由純淨天空整理自Github/MSDocs等開源代碼及文檔管理平台，相關代碼片段篩選自各路編程大神貢獻的開源項目，源碼版權歸原作者所有，傳播和使用請參考對應項目的License；未經允許，請勿轉載。