Python cupy.ElementwiseKernel方法代碼示例

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def backward_gpu(self, inputs, gys):
        if not self.gpu_optim:
            return self.backward_cpu(inputs,  gys)
        xp = cuda.get_array_module(*inputs)
        x, gamma, beta = inputs
        gy, = gys
        g_beta = xp.sum(gy, axis=0, keepdims=True)
        g_gamma = xp.sum(gy*self.normalized, axis=0, keepdims=True)
        
        gy2 = gy*gamma
        gy_centered = gy2 - xp.mean(gy2, axis=1, keepdims=True)
        sc_prod = xp.sum(gy_centered * self.normalized, axis = 1, keepdims=True)
        
        H = x.shape[1]
#         ga = backprop_scale(self.inv_norm, gy_centered, self.normalized, sc_prod/H)
        ga = cp.ElementwiseKernel(
         'T inv_norm, T gy_centered, T normalized, T sc_prod',
         'T z',
          '''
              z = inv_norm *(gy_centered - normalized * (sc_prod/%f));
         '''%H,
         'backprop_scale')(self.inv_norm, gy_centered, self.normalized, sc_prod)
        
        return ga, g_gamma, g_beta 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def _get_map_kernel(ndim, large_int, yshape, mode, cval=0.0, order=1,
                    integer_output=False):
    in_params = 'raw X x, raw W coords'
    out_params = 'Y y'
    operation, name = _generate_interp_custom(
        coord_func=_get_coord_map,
        ndim=ndim,
        large_int=large_int,
        yshape=yshape,
        mode=mode,
        cval=cval,
        order=order,
        name='shift',
        integer_output=integer_output,
    )
    return cupy.ElementwiseKernel(in_params, out_params, operation, name) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def _get_shift_kernel(ndim, large_int, yshape, mode, cval=0.0, order=1,
                      integer_output=False):
    in_params = 'raw X x, raw W shift'
    out_params = 'Y y'
    operation, name = _generate_interp_custom(
        coord_func=_get_coord_shift,
        ndim=ndim,
        large_int=large_int,
        yshape=yshape,
        mode=mode,
        cval=cval,
        order=order,
        name='shift',
        integer_output=integer_output,
    )
    return cupy.ElementwiseKernel(in_params, out_params, operation, name) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def _get_zoom_shift_kernel(ndim, large_int, yshape, mode, cval=0.0, order=1,
                           integer_output=False):
    in_params = 'raw X x, raw W shift, raw W zoom'
    out_params = 'Y y'
    operation, name = _generate_interp_custom(
        coord_func=_get_coord_zoom_and_shift,
        ndim=ndim,
        large_int=large_int,
        yshape=yshape,
        mode=mode,
        cval=cval,
        order=order,
        name='zoom_shift',
        integer_output=integer_output,
    )
    return cupy.ElementwiseKernel(in_params, out_params, operation, name) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def _get_zoom_kernel(ndim, large_int, yshape, mode, cval=0.0, order=1,
                     integer_output=False):
    in_params = 'raw X x, raw W zoom'
    out_params = 'Y y'
    operation, name = _generate_interp_custom(
        coord_func=_get_coord_zoom,
        ndim=ndim,
        large_int=large_int,
        yshape=yshape,
        mode=mode,
        cval=cval,
        order=order,
        name='zoom',
        integer_output=integer_output,
    )
    return cupy.ElementwiseKernel(in_params, out_params, operation, name) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def _kernel_finalize():
    return cupy.ElementwiseKernel(
        'int32 maxlabel', 'raw int32 labels, raw Y y',
        '''
        if (y[i] < 0) {
            y[i] = 0;
            continue;
        }
        int yi = y[i];
        int j_min = 0;
        int j_max = maxlabel - 1;
        int j = (j_min + j_max) / 2;
        while (j_min < j_max) {
            if (yi == labels[j]) break;
            if (yi < labels[j]) j_max = j - 1;
            else j_min = j + 1;
            j = (j_min + j_max) / 2;
        }
        y[i] = j + 1;
        ''',
        'cupyx_nd_label_finalize') 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def test_manual_indexing(self, n=100):
        in1 = cupy.random.uniform(-1, 1, n).astype(cupy.float32)
        in2 = cupy.random.uniform(-1, 1, n).astype(cupy.float32)
        uesr_kernel_1 = cupy.ElementwiseKernel(
            'T x, T y',
            'T z',
            '''
                z = x + y;
            ''',
            'uesr_kernel_1')
        out1 = uesr_kernel_1(in1, in2)

        uesr_kernel_2 = cupy.ElementwiseKernel(
            'raw T x, raw T y',
            'raw T z',
            '''
                z[i] = x[i] + y[i];
            ''',
            'uesr_kernel_2')
        out2 = uesr_kernel_2(in1, in2, size=n)

        testing.assert_array_equal(out1, out2) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def test_python_scalar(self):
        for typ in (int, float, bool):
            dtype = numpy.dtype(typ).type
            in1_cpu = numpy.random.randint(0, 1, (4, 5)).astype(dtype)
            in1 = cupy.array(in1_cpu)
            scalar_value = typ(2)
            uesr_kernel_1 = cupy.ElementwiseKernel(
                'T x, T y',
                'T z',
                '''
                    z = x + y;
                ''',
                'uesr_kernel_1')
            out1 = uesr_kernel_1(in1, scalar_value)

            expected = in1_cpu + dtype(2)
            testing.assert_array_equal(out1, expected) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def __init__(self):
        self.unpack_kernel = cupy.ElementwiseKernel(
            'raw T vec, int32 matrix_size',
            'raw T mat',
            """
            int x = i % matrix_size;
            int y = i / matrix_size;
            if( x < y ) {
                int tmp = y;
                y = x;
                x = tmp;
            }
            mat[i] = vec[matrix_size * y - y * (y + 1) / 2 + x];
            """,
            'unpack'
        ) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def execute(cls, ctx, op):
        import cupy as cp

        chunk = op.outputs[0]
        func = cp.ElementwiseKernel(*_evaluate(chunk))
        ctx[chunk.key] = func(*[ctx[i.key] for i in op.inputs]) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def test_kernel(self):
        import cupy as cp
        x = cp.arange(6, dtype='f').reshape(2, 3)
        y = cp.arange(3, dtype='f')
        kernel = cp.ElementwiseKernel(
            'float32 x, float32 y', 'float32 z',
            '''if (x - 2 > y) {
                z = x * y;
            } else {
                z = x + y;
            }''',
            'my_kernel')
        r = kernel(x, y)
        
        self.assertEqual((2, 3), r.shape) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def col2im_gpu(col, sy, sx, ph, pw, h, w, dy=1, dx=1):
    n, c, kh, kw, out_h, out_w = col.shape
    img = cp.empty((n, c, h, w), dtype=col.dtype)
    cp.ElementwiseKernel(
        'raw T col, int32 h, int32 w, int32 out_h, int32 out_w,'
        'int32 kh, int32 kw, int32 sy, int32 sx, int32 ph, int32 pw,'
        'int32 dx, int32 dy',
        'T img',
        '''
           int c0 = i / (h * w);
           int y  = i / w % h;
           int x  = i % w;
           T val = 0;
           for (int ky = 0; ky < kh; ++ky) {
             int out_y = (y + ph - ky * dy);
             if (0 > out_y || out_y >= out_h * sy) continue;
             if (out_y % sy != 0) continue;
             out_y /= sy;
             for (int kx = 0; kx < kw; ++kx) {
               int out_x = (x + pw - kx * dx);
               if (0 > out_x || out_x >= out_w * sx) continue;
               if (out_x % sx != 0) continue;
               out_x /= sx;
               int k = out_y + out_h * (kx + kw * (ky + kh * c0));
               val = val + col[out_x + out_w * k];
             }
           }
           img = val;
        ''',
        'col2im')(col.reduced_view(),
                  h, w, out_h, out_w, kh, kw, sy, sx, ph, pw, dx, dy, img)
    return img 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def elementwise(in_params, out_params, operation, name, **kwargs):
    """Creates an elementwise kernel function.

    This function uses :func:`~chainer.backends.cuda.memoize` to cache the
    kernel object, i.e. the resulting kernel object is cached for each argument
    combination and CUDA device.

    The arguments are the same as those for
    :class:`cupy.ElementwiseKernel`, except that the ``name`` argument is
    mandatory.

    """
    check_cuda_available()
    return cupy.ElementwiseKernel(
        in_params, out_params, operation, name, **kwargs) 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def get_label_lengths(self, labels):
        if self.xp == numpy:
            label_lengths = self.xp.zeros(len(labels))

            for i in range(len(labels)):
                for j in range(len(labels[i])):
                    if labels.data[i][j] == self.blank_symbol:
                        label_lengths[i] = j
                        break
        else:
            import cupy
            label_length_kernel = cupy.ElementwiseKernel(
                'raw T labels, int32 blank_symbol, int32 num_labels',
                'T length',
                '''
                    for (int j = 0; j < num_labels; ++j) {
                        T label_value = labels[i * num_labels + j];
                        if (label_value == blank_symbol) {
                            length = j;
                            break;
                        }
                    }
                ''',
                'get_label_lengths'
            )
            label_lengths = label_length_kernel(labels.data, self.blank_symbol, labels.shape[1], size=len(labels))
        return label_lengths 

# 需要導入模塊: import cupy [as 別名]
# 或者: from cupy import ElementwiseKernel [as 別名]
def _call_kernel(kernel, input, weights, output, structure=None,
                 weights_dtype=cupy.float64, structure_dtype=cupy.float64):
    """
    Calls a constructed ElementwiseKernel. The kernel must take an input image,
    an optional array of weights, an optional array for the structure, and an
    output array.

    weights and structure can be given as None (structure defaults to None) in
    which case they are not passed to the kernel at all. If the output is given
    as None then it will be allocated in this function.

    This function deals with making sure that the weights and structure are
    contiguous and float64 (or bool for weights that are footprints)*, that the
    output is allocated and appriopately shaped. This also deals with the
    situation that the input and output arrays overlap in memory.

    * weights is always cast to float64 or bool in order to get an output
    compatible with SciPy, though float32 might be sufficient when input dtype
    is low precision. If weights_dtype is passed as weights.dtype then no
    dtype conversion will occur. The input and output are never converted.
    """
    args = [input]
    if weights is not None:
        weights = cupy.ascontiguousarray(weights, weights_dtype)
        args.append(weights)
    if structure is not None:
        structure = cupy.ascontiguousarray(structure, structure_dtype)
        args.append(structure)
    output = _get_output(output, input)
    needs_temp = cupy.shares_memory(output, input, 'MAY_SHARE_BOUNDS')
    if needs_temp:
        output, temp = _get_output(output.dtype, input), output
    args.append(output)
    kernel(*args)
    if needs_temp:
        temp[...] = output[...]
        output = temp
    return output