Python NervanaGPU.max方法代碼示例

# 需要導入模塊: from nervanagpu import NervanaGPU [as 別名]
# 或者: from nervanagpu.NervanaGPU import max [as 別名]
            glops = max(glops16, glops32, glops64, glops128)

            if glops16 == glops:
                fastest = 16
            elif glops32 == glops:
                fastest = 32
            elif glops64 == glops:
                fastest = 64
            else:
                fastest = 128

            glopsref = cublas_dot(devA2, devB2, devC2, repeat=repeat)

            partial1 = ng.empty((devC1.shape[0],1), dtype=np.float32)
            partial2 = partial1[0:1,0:1]

            diff = ng.max(abs(devC2 - devC1), partial=partial1, out=partial2).get()[0,0]
            mean = ng.mean(abs(devC2), partial=partial1, out=partial2).get()[0,0]

            flops_diff = glops - glopsref

            note = "**************" if flops_diff <= 0 else ""
            
            print "Faster: %.0f gflops Choice: %d Error: %.3f%%%s" % (flops_diff, fastest, 100 * diff / mean, note)

        print "--------------------------------------------------------------------------------"


cublas.cublasDestroy(handle)

# 需要導入模塊: from nervanagpu import NervanaGPU [as 別名]
# 或者: from nervanagpu.NervanaGPU import max [as 別名]

#.........這裏部分代碼省略.........
                                 each output feature map stored in out.
            ifmshape (tuple): Dimensions of each input feature map (typically
                              height and width).
            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.
            fwidth (int): Filter width.
            updatebuf (GPUTensor): Temporary storage buffer used to hold the
                                   updated gradient for a single receptive
                                   field
            local (bool, optional): Whether to do local filtering (True) or
                                    convolution (False, the default)
            layer (Layer): The layer object.
        """
        self.ng.update_conv(layer=updatebuf, I=inputs, E=deltas, grad_F=out,
                            alpha=1.0, repeat=1)

    def fprop_pool(self, out, inputs, op, ofmshape, ofmsize, ofmlocs, fshape,
                   ifmshape, links, nifm, padding, stride, fpropbuf):
        """
        Forward propagate the inputs of a Pooling 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.
            op (string): The type of pooling operation to apply.  We support
                         "max", "avg", "l2" currently.
            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.
            fshape (tuple): Dimensions of each filter (typically height and
                            width).
            ifmshape (tuple): Dimensions of each input feature map (typically
                              number of height and width neurons).
            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
                           pooling operation.
            stride (int): Number of neurons to shift the filter at each step.
            fpropbuf (GPUTensor): Temporary storage buffer used to hold the
                                  pooled outputs for a single receptive field.
        """
        op = op.lower()
        if op == "max":
            self.ng.fprop_pool(layer=fpropbuf, I=inputs, O=out, repeat=1)
        else:
            raise AttributeError("unexpected pooling op type: %s", op)

    def bprop_pool(self, out, fouts, inputs, deltas, op, ofmshape, ofmsize,
                   ofmlocs, fshape, fpsize, ifmshape, links, nifm, padding,
                   stride, bpropbuf):
        """
        Backward propagate the error through a pooling network layer.

        Arguments:
            out (GPUTensor): Where to store the backward propagated errors.

# 需要導入模塊: from nervanagpu import NervanaGPU [as 別名]
# 或者: from nervanagpu.NervanaGPU import max [as 別名]
    ng.fprop_conv (conv, nlI, nlF, nlO, alpha=alpha, repeat=repeat)
    ng.bprop_conv (conv, nlF, nlE, nlB, alpha=alpha, repeat=repeat)
    ng.update_conv(conv, nlI, nlE, nlU, alpha=alpha, repeat=repeat)

    nlI = nlF = nlE = None

    print "\ncudnn vs nervanaLib:"

    parO = ng.empty((N,1), dtype=np.float32)
    parB = ng.empty((N,1), dtype=np.float32)
    parU = ng.empty((K,1), dtype=np.float32)
    maxO = parO[0:1,0:1]
    maxB = parB[0:1,0:1]
    maxU = parU[0:1,0:1]

    maxo  = ng.max(abs(cuO - nlO.T), partial=parO, out=maxO).get()[0,0]
    maxb  = ng.max(abs(cuB - nlB.T), partial=parB, out=maxB).get()[0,0]
    maxu  = ng.max(abs(cuU - nlU.T), partial=parU, out=maxU).get()[0,0]

    meano = ng.mean(abs(cuO), partial=parO, out=maxO).get()[0,0]
    meanb = ng.mean(abs(cuB), partial=parB, out=maxB).get()[0,0]
    meanu = ng.mean(abs(cuU), partial=parU, out=maxU).get()[0,0]

    print "        maxerr   mean   pct"
    print "fprop: %7.5f %6.2f %5.3f" % (maxo, meano, 100*maxo/meano)
    print "bprop: %7.5f %6.2f %5.3f" % (maxb, meanb, 100*maxb/meanb)
    print "updat: %7.5f %6.2f %5.3f" % (maxu, meanu, 100*maxu/meanu)

    # free up memory from this layer before proceeding
    cuB  = cuU  = cuO  = None
    nlB  = nlU  = nlO  = None