Python NervanaGPU.fprop_conv方法代码示例

# 需要导入模块: from nervanagpu import NervanaGPU [as 别名]
# 或者: from nervanagpu.NervanaGPU import fprop_conv [as 别名]

#.........这里部分代码省略.........
            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.
            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)

# 需要导入模块: from nervanagpu import NervanaGPU [as 别名]
# 或者: from nervanagpu.NervanaGPU import fprop_conv [as 别名]
def run():
    ng = NervanaGPU(stochastic_round=False)

    dt = np.float32
    # N: Number of images in mini-batch
    # C: Number of input feature maps
    # K: Number of output feature maps

    # D: Depth  of input image
    # H: Height of input image
    # W: Width  of input image

    # T: Depth  of filter kernel
    # R: Height of filter kernel
    # S: Width  of filter kernel
    # 
    # * images:      (numColors, imgSizeY, imgSizeX, numImages) with stride given
    # * filters:     (numColors, filterPixels, numFilters) if conv
    # *              (numModules, numColors, filterPixels, numFilters) otherwise
    # *
    # * targets:     (numFilters, numModulesY, numModulesX, numImages)

    N = 128
    C = 3
    K = 64

    D = 1
    H = 64
    W = 64

    T = 1
    R = 8
    S = 8

    pad_h = pad_w = 0
    str_h = str_w = 4

    layer = ng.conv_layer(dt, N, C, K,
            D=D, H=H, W=W,
            T=T, R=R, S=S,
            pad_d=0, pad_h=pad_h, pad_w=pad_w,
            str_d=1, str_h=str_h, str_w=str_w,
            grid_P=0, grid_Q=0, update_size=None)

    numImages = N 
    numFilters = K

    numModulesY = int(math.ceil(float(H - R + 1 + 2*pad_h) / str_h))
    numModulesX = int(math.ceil(float(W - S + 1 + 2*pad_w) / str_w))

    print "Num Modules ", numModulesX, numModulesY


    # Set up images, filters, and outputs
    # imgd = np.loadtxt("im1.txt")
    # img = np.zeros((64, 64, 3))
    # print imgd.shape
    # for i in range(3):
    #     img[:, :, i] = imgd[i*64:(i+1)*64, :]
    # hostImages = np.tile(img)

    hostImages = np.random.rand(C, H, W, N)
    hostFilters = np.random.uniform(low=0.0, high=1.0, size=(C, S*R, numFilters)) #np.ones((C, S*R, numFilters)) #
    hostOutputs = np.zeros((numFilters, numModulesY, numModulesX, N))

    print "Input sum", np.sum(hostImages)

    # Run cc2 kernel    
    devI = ng.array(hostImages, dtype=dt)
    devF = ng.array(hostFilters, dtype=dt)
    devO = ng.array(hostOutputs, dtype=dt)

    ng.fprop_cuda_conv(layer, devI, devF, devO)

    print "CC2 input sum: ", np.sum(devI.asnumpyarray())
    print "CC2 output sum: ", np.sum(devO.asnumpyarray())

    # Run maxwel kernel
    # images: (C * H * W, N)
    # filters:  (C * S * R , numFilters)
    # outputs:  (numFilters * numModulesX * numModulesY, N)
    devI = ng.array(hostImages.reshape((C*H*W, N)), dtype=dt)
    devF = ng.array(hostFilters.reshape((C*S*R, numFilters)), dtype=dt)
    devO2 = ng.array(hostOutputs.reshape(numFilters*numModulesX*numModulesY, N), dtype=dt)

    ng.fprop_conv(layer, devI, devF, devO2)
    print "NG input sum: ", np.sum(devI.asnumpyarray())
    print "NG output sum: ", np.sum(devO2.asnumpyarray())

    hostOutputs1 = np.reshape(devO.asnumpyarray(), devO2.shape)
    hostOutputs2 = devO2.asnumpyarray()

    for i in xrange(hostOutputs1.shape[0]):
       for j in xrange(hostOutputs1.shape[1]):
           assert(abs(hostOutputs1[i, j] - hostOutputs2[i, j]) < 1e-4)

# 需要导入模块: from nervanagpu import NervanaGPU [as 别名]
# 或者: from nervanagpu.NervanaGPU import fprop_conv [as 别名]
    cuI = None

    nlF = ng.empty(dimF, dtype=dtype)
    nlF[:] = cuF.T
    cuF = None

    nlE = ng.empty(dimO, dtype=dtype)
    nlE[:] = cuE.T
    cuE = None

    nlB = ng.empty(dimI, dtype=dtype)
    nlU = ng.empty(dimF, dtype=dtype)
    nlO = ng.empty(dimO, dtype=dtype)
    #print drv.mem_get_info()

    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]

# 需要导入模块: from nervanagpu import NervanaGPU [as 别名]
# 或者: from nervanagpu.NervanaGPU import fprop_conv [as 别名]
# cpu output arrays
cpuO = np.zeros(dimO,             dtype=np.float32)
cpuB = np.zeros(slicable(dimI,1), dtype=np.float32)
cpuU = np.zeros(slicable(dimF),   dtype=np.float32)

# give gpu the input array without zero padding (not needed)
devI = ng.array(cpuI[:-1,:].reshape(dimI), dtype=dtype)
devF = ng.array(cpuF.reshape(dimF), dtype=dtype)
devE = ng.array(cpuE, dtype=dtype)

devO = devB = devU = 0

if "fprop"  in ops:
    devO = ng.empty(dimO, dtype=dtype)
    ng.fprop_conv(conv,  devI, devF, devO, alpha=1.0, repeat=repeat)

if "bprop"  in ops:
    devB = ng.empty(dimI, dtype=dtype)
    ng.bprop_conv(conv,  devF, devE, devB, alpha=1.0, repeat=repeat)

if "update" in ops:
    devU = ng.empty(dimF, dtype=dtype)
    ng.update_conv(conv, devI, devE, devU, alpha=1.0, repeat=repeat)


def pixel_indices(mt, pr, qs):

    T,R,S = conv.TRS
    D,H,W = conv.DHW
    C     = conv.C