本文整理汇总了Python中nervanagpu.NervanaGPU.add方法的典型用法代码示例。如果您正苦于以下问题:Python NervanaGPU.add方法的具体用法?Python NervanaGPU.add怎么用?Python NervanaGPU.add使用的例子?那么恭喜您, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在类nervanagpu.NervanaGPU
的用法示例。
在下文中一共展示了NervanaGPU.add方法的2个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。
示例1: GPU
# 需要导入模块: from nervanagpu import NervanaGPU [as 别名]
# 或者: from nervanagpu.NervanaGPU import add [as 别名]
#.........这里部分代码省略.........
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
layer (Layer): The layer object.
"""
self.ng.dot(weights, inputs, out)
def bprop_fc(self, out, weights, deltas, layer=None):
"""
Backward propagate the error through a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
self.ng.dot(weights.T, deltas, out)
def update_fc(self, out, inputs, deltas, layer=None):
"""
Compute the updated gradient for a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the updated gradient value.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
self.ng.dot(deltas, inputs.T, out)
def fprop_conv(self, out, inputs, weights, ofmshape, ofmsize, ofmlocs,
ifmshape, links, nifm, padding, stride, ngroups, fpropbuf,
local=False):
"""
Forward propagate the inputs of a convolutional network layer to
produce output pre-activations (ready for transformation by an
activation function).
Arguments:
out (GPUTensor): Where to store the forward propagated results.
inputs (GPUTensor): Will be either the dataset input values (first
layer), or the outputs from the previous layer.
weights (GPUTensor): The weight coefficient values for this layer.
ofmshape (tuple): Dimensions of each output feature map (typically
number of height and width neurons).
ofmsize (int): Total size of each output feature map.
ofmlocs (GPUTensor): Indices giving the location of each element
in each output feature map stored in out.
ifmshape (tuple): Dimensions of each input feature map (typically
number of height and width neurons). For this
backend we expect these values to be square.
links (GPUTensor): Input receptive field indices.
nifm (int): Total number of input feature maps.
padding (int): Number of additional elements to include along each
dimension of each local receptive field during the
convolution operation.
stride (int): Number of neurons to shift the filter at each step.
ngroups (int): Number of groups.
fpropbuf (GPUTensor): Temporary storage buffer used to hold the
convolved outputs for a single receptive
field. Not used for this backend.
local (bool, optional): Whether to do local filtering (True) or
convolution (False, the default)
"""
示例2: MGPU
# 需要导入模块: from nervanagpu import NervanaGPU [as 别名]
# 或者: from nervanagpu.NervanaGPU import add [as 别名]
#.........这里部分代码省略.........
assert hbuf.dtype == dbuf.dtype
ndata = dbuf.size
starts = [i * ndata for i in range(self.num_dev)]
for dest, strm, ctx, doff in zip(dbuf.tlist, self.strms, self.ctxs,
starts):
src = hbuf.reshape((hbuf.size))[doff:(doff + ndata)]
ctx.push()
drv.memcpy_htod_async(dest.ptr, src, strm)
ctx.pop()
self.synchronize()
def fprop_fc(self, out, inputs, weights, layer=None):
"""
In this case, the weights are shards, the acts are replicas
ubuf should be of size nout/num_dev x mbsz
"""
ubuf = layer.mempool[0]
assert ubuf.shape == (weights.shape[0], inputs.shape[1])
if layer.use_biases:
biases = layer.biases.tlist
else:
biases = [None for i in range(self.num_dev)]
for dbuf, ibuf, wt, bs, strm, ctx in zip(ubuf.tlist, inputs.tlist,
weights.tlist, biases,
self.strms, self.ctxs):
ctx.push()
self.ng.stream = strm
self.ng.dot(wt, ibuf, dbuf)
if layer.use_biases:
self.ng.add(dbuf, bs, out=dbuf)
ctx.pop()
# Note, should be safe not to sync because each fragment is computed
# on the same stream that originates the copy
# self.synchronize()
self.fragment_to_replica(ubuf, out)
def bprop_fc(self, out, weights, deltas, layer=None):
"""
Backward propagate the error through a fully connected network layer.
Arguments:
out (GPUTensor): Where to store the backward propagated errors.
weights (GPUTensor): The weight coefficient values for this layer.
deltas (GPUTensor): The error values for this layer
layer (Layer): The layer object.
"""
ubuf = layer.mempool[1]
wtsz = weights.shape[0]
starts = [i * wtsz for i in range(self.num_dev)]
assert out.shape == (weights.shape[1], deltas.shape[1])
assert ubuf.shape == out.shape
for dbuf, ibuf, wt, strm, ctx, off in zip(out.tlist, deltas.tlist,
weights.tlist, self.strms,
self.ctxs, starts):
ctx.push()
self.ng.stream = strm
self.ng.dot(wt.T, ibuf[off:(off + wtsz)], dbuf)
ctx.pop()
# Note, should be safe not to sync because each fragment is computed