Python comm.broadcast_coalesced方法代碼示例

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def forward(ctx, target_gpus, *inputs):
        if not all(input.is_cuda for input in inputs):
            raise TypeError('Broadcast function not implemented for CPU tensors')
        ctx.target_gpus = target_gpus
        if len(inputs) == 0:
            return tuple()
        ctx.num_inputs = len(inputs)
        ctx.input_device = inputs[0].get_device()
        outputs = comm.broadcast_coalesced(inputs, ctx.target_gpus)
        non_differentiables = []
        for idx, input_requires_grad in enumerate(ctx.needs_input_grad[1:]):
            if not input_requires_grad:
                for output in outputs:
                    non_differentiables.append(output[idx])
        ctx.mark_non_differentiable(*non_differentiables)
        return tuple([t for tensors in outputs for t in tensors]) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def _sync_params(self):
        params = [p.data for p in self.module.parameters()]
        result = broadcast_coalesced(params, self.device_ids, self.broadcast_bucket_size)
        for tensors, module in zip(result[1:], self._module_copies[1:]):
            for tensor, param in zip(tensors, module.parameters()):
                param.data.set_(tensor)

        # cross-node buffer sync
        buffers = list(self.module._all_buffers())
        flat_buffers = _flatten_tensors(buffers)
        dist.broadcast(flat_buffers, 0)
        for buf, synced in zip(buffers, _unflatten_tensors(flat_buffers, buffers)):
            buf.copy_(synced)

        # intra-node buffer sync
        result = broadcast_coalesced(buffers, self.device_ids, self.broadcast_bucket_size)
        for tensors, module in zip(result[1:], self._module_copies[1:]):
            for tensor, buf in zip(tensors, module._all_buffers()):
                buf.set_(tensor) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def step(self, l2norm):
    self.optimizer.step()
    group_params = [(i, [(n,m[n]) for n,p in self.param_group]) for i,m in enumerate(self.named_params)]
    group_params = sorted(group_params, key=lambda x:x[0] if x[0]!=self.device else -1)
    params = dict(self.param_group)
    for n,p in group_params[0][1]:
      if p.data.dtype == torch.half:
        p.data.copy_(params[n].data)
      else:
        p.data = params[n].data
    param_list = [[p for n,p in g] for i,g in group_params]
    device_list =[i for i,g in group_params]
    outputs = broadcast_coalesced(param_list[0], device_list)
    for o,p in zip(outputs, param_list):
      for x,y in zip(o, p):
        y.data.copy_(x.data) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [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 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def forward(ctx, num_inputs, *inputs):
        ctx.num_inputs = num_inputs
        ctx.target_gpus = [inputs[i].get_device() for i in range(0, len(inputs), num_inputs)]
        inputs = [inputs[i:i + num_inputs]
                 for i in range(0, len(inputs), num_inputs)]
        # sort before reduce sum
        inputs = sorted(inputs, key=lambda i: i[0].get_device())
        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
        return tuple([t for tensors in outputs for t in tensors]) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def backward(ctx, *inputs):
        inputs = [i.data for i in inputs]
        inputs = [inputs[i:i + ctx.num_inputs]
                 for i in range(0, len(inputs), ctx.num_inputs)]
        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
        return (None,) + tuple([Variable(t) for tensors in outputs for t in tensors]) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def forward(ctx, num_inputs, *inputs):
        ctx.num_inputs = num_inputs
        ctx.target_gpus = [inputs[i].get_device() for i in range(0, len(inputs), num_inputs)]
        inputs = [inputs[i:i + num_inputs]
                  for i in range(0, len(inputs), num_inputs)]
        # sort before reduce sum
        inputs = sorted(inputs, key=lambda i: i[0].get_device())
        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
        return tuple([t for tensors in outputs for t in tensors]) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def backward(ctx, *inputs):
        inputs = [i.data for i in inputs]
        inputs = [inputs[i:i + ctx.num_inputs]
                  for i in range(0, len(inputs), ctx.num_inputs)]
        results = comm.reduce_add_coalesced(inputs, ctx.target_gpus[0])
        outputs = comm.broadcast_coalesced(results, ctx.target_gpus)
        return (None,) + tuple([Variable(t) for tensors in outputs for t in tensors]) 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [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 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def backward(ctx, dz):
        x, _ex, _exs, gamma, beta = ctx.saved_tensors
        dz = dz.contiguous()

        # BN backward
        if dz.is_cuda:
            dx, _dex, _dexs, dgamma, dbeta = \
                lib.gpu.batchnorm_backward(dz, x, _ex, _exs, gamma, beta, ctx.eps)
        else:
            raise NotImplemented

        if ctx.training:
            if ctx.sync:
                if ctx.is_master:
                    _dex, _dexs = [_dex.unsqueeze(0)], [_dexs.unsqueeze(0)]
                    for _ in range(ctx.master_queue.maxsize):
                        _dex_w, _dexs_w = ctx.master_queue.get()
                        ctx.master_queue.task_done()
                        _dex.append(_dex_w.unsqueeze(0))
                        _dexs.append(_dexs_w.unsqueeze(0))

                    _dex = comm.gather(_dex).mean(0)
                    _dexs = comm.gather(_dexs).mean(0)

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

            if x.is_cuda:
                dx_ = lib.gpu.expectation_backward(x, _dex, _dexs)
            else:
                raise NotImplemented
            dx = dx + dx_

        return dx, dgamma, dbeta, None, None, None, None, None, None, None, None, None 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def backward(ctx, dz):
        x, _ex, _exs, gamma, beta = ctx.saved_tensors
        dz = dz.contiguous()

        # BN backward
        if dz.is_cuda:
            dx, _dex, _dexs, dgamma, dbeta = lib.gpu.batchnorm_backward(dz, x, _ex, _exs, gamma, beta, ctx.eps)
        else:
            raise NotImplemented

        if ctx.training:
            if ctx.sync:
                if ctx.is_master:
                    _dex, _dexs = [_dex.unsqueeze(0)], [_dexs.unsqueeze(0)]
                    for _ in range(ctx.master_queue.maxsize):
                        _dex_w, _dexs_w = ctx.master_queue.get()
                        ctx.master_queue.task_done()
                        _dex.append(_dex_w.unsqueeze(0))
                        _dexs.append(_dexs_w.unsqueeze(0))

                    _dex = comm.gather(_dex).mean(0)
                    _dexs = comm.gather(_dexs).mean(0)

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

            if x.is_cuda:
                dx_ = lib.gpu.expectation_backward(x, _dex, _dexs)
            else:
                raise NotImplemented
            dx = dx + dx_

        return dx, dgamma, dbeta, None, None, None, None, None, None, None, None, None 

# 需要導入模塊: from torch.cuda import comm [as 別名]
# 或者: from torch.cuda.comm import broadcast_coalesced [as 別名]
def backward(ctx, dz):
        x, _ex, _exs, gamma, beta = ctx.saved_tensors
        dz = dz.contiguous()

        # BN backward
        dx, _dex, _dexs, dgamma, dbeta = _C.batchnorm_backward(dz, x, _ex, _exs, gamma, beta, ctx.eps)

        if ctx.training:
            if ctx.sync:
                if ctx.is_master:
                    _dex, _dexs = [_dex.unsqueeze(0)], [_dexs.unsqueeze(0)]
                    for _ in range(ctx.master_queue.maxsize):
                        _dex_w, _dexs_w = ctx.master_queue.get()
                        ctx.master_queue.task_done()
                        _dex.append(_dex_w.unsqueeze(0))
                        _dexs.append(_dexs_w.unsqueeze(0))

                    _dex = comm.gather(_dex).mean(0)
                    _dexs = comm.gather(_dexs).mean(0)

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

            dx_ = _C.expectation_backward(x, _dex, _dexs)
            dx = dx + dx_

        return dx, dgamma, dbeta, None, None, None, None, None, None, None, None, None