Spaces:
Sleeping
Sleeping
File size: 8,208 Bytes
dc2106c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 |
# Copyright (c) ONNX Project Contributors
# SPDX-License-Identifier: Apache-2.0
import numpy as np
from onnx.reference.op_run import OpRun
from onnx.reference.ops._op_common_indices import _get_indices, _is_out
def _col2im_shape_check_2d(X, output_shape, kernel_shape, dilations, pads, strides): # type: ignore
output_height, output_width = output_shape
kernel_height, kernel_width = kernel_shape
dilation_height, dilation_width = dilations
stride_height, stride_width = strides
ndim = len(X.shape)
if not (
(ndim == 2 and X.shape[0] != 0 and X.shape[1] != 0)
or (ndim == 3 and X.shape[1] != 0 and X.shape[2] != 0)
):
raise ValueError(
"Expected 2D or 3D (batch mode) tensor for input with possibly 0 batch size and non-zero dimensions for input."
)
batch_dim = 0 if len(X.shape) == 3 else -1
n_input_plane = X.shape[batch_dim + 1]
if n_input_plane % (kernel_width * kernel_height) != 0:
raise ValueError(
f"Expected size of input's dimension 1 to be divisible by the "
f"product of kernel_size, but got input.size(1)={n_input_plane} "
f"and kernel_size={kernel_shape}."
)
input_length = X.shape[batch_dim + 2]
n_blocks_height = (
output_height + pads[0, :].sum() - dilation_height * (kernel_height - 1) - 1
) // stride_height + 1
n_blocks_width = (
output_width + pads[1, :].sum() - dilation_width * (kernel_width - 1) - 1
) // stride_width + 1
if input_length != (n_blocks_height * n_blocks_width):
raise ValueError(
f"Given batch_dim={batch_dim}, n_input_plane={n_input_plane}, X.shape={X.shape}, "
f"output_shape={output_shape}, kernel_shape={kernel_shape}, "
f"dilations={dilations}, pads={pads}, strides={strides}, "
f"expected size of input's dimension 2 to match the calculated number of ",
f"sliding blocks {n_blocks_height} * {n_blocks_width} = {n_blocks_height * n_blocks_width}, "
f"but got input.size(2)={input_length}.",
)
if not (n_blocks_height >= 1 and n_blocks_width >= 1):
raise ValueError(
f"Given batch_dim={batch_dim}, n_input_plane={n_input_plane}, X.shape={X.shape}, "
f"output_shape={output_shape}, kernel_shape={kernel_shape}, "
f"dilations={dilations}, pads={pads}, strides={strides}, "
f"calculated shape of the array of sliding blocks as ({n_blocks_height}, {n_blocks_width}), "
f"which is too small (non-positive)."
)
def _col2im_naive_implementation_2d(res, image_shape, kernel_shape, dilations, pads, strides): # type: ignore
# source: https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/im2col.h
n_dims = len(pads) // 2
new_pads = np.array([(pads[i], pads[i + n_dims]) for i in range(n_dims)])
_col2im_shape_check_2d(res, image_shape, kernel_shape, dilations, new_pads, strides)
data_col = res.ravel()
data_im = np.zeros(image_shape, dtype=res.dtype).flatten()
kernel_h, kernel_w = kernel_shape
channels_col = kernel_h * kernel_w
stride_h, stride_w = strides
dilation_h, dilation_w = dilations
pad_h, pad_w = new_pads[:, 0]
height, width = image_shape
output_height, output_width = image_shape
height_col = (
output_height + new_pads[0, :].sum() - (dilation_h * (kernel_h - 1) + 1)
) // stride_h + 1
width_col = (
output_width + new_pads[1, :].sum() - (dilation_w * (kernel_w - 1) + 1)
) // stride_w + 1
for c_col in range(channels_col):
w_offset = c_col % kernel_w
h_offset = (c_col // kernel_w) % kernel_h
c_im = c_col // (kernel_h * kernel_w)
for h_col in range(height_col):
h_im = h_col * stride_h - pad_h + h_offset * dilation_h
for w_col in range(width_col):
w_im = w_col * stride_w - pad_w + w_offset * dilation_w
if 0 <= h_im < height and 0 <= w_im < width:
i_im = (c_im * height + h_im) * width + w_im
i_col = (c_col * height_col + h_col) * width_col + w_col
if 0 <= i_col < data_col.shape[0]:
data_im[i_im] += data_col[i_col]
return data_im.reshape(image_shape)
def _col2im_shape_check(X, output_shape, kernel_shape, dilations, pads, strides): # type: ignore
n_input_plane = X.shape[0]
kernel_size = np.prod(kernel_shape)
if n_input_plane % kernel_size != 0:
raise ValueError(
f"Expected size of input's dimension 1 to be divisible by the "
f"product of kernel_size={kernel_size}, "
f"but got input.size(1)={n_input_plane} "
f"and kernel_shape={kernel_shape}, X.shape={X.shape}, output_shape={output_shape}."
)
input_length = X.shape[1]
n_dims = len(output_shape)
n_blocks = []
for i in range(n_dims):
n_block = (
output_shape[i]
+ pads[i, :].sum()
- dilations[i] * (kernel_shape[i] - 1)
- 1
) // strides[i] + 1
n_blocks.append(n_block)
block_size = np.prod(n_blocks)
if input_length != block_size:
raise ValueError(
f"Given n_input_plane={n_input_plane}, X.shape={X.shape}, "
f"output_shape={output_shape}, kernel_shape={kernel_shape}, "
f"dilations={dilations}, pads={pads}, strides={strides}, "
f"expected size of input's dimension 2 to match the calculated number of "
f"sliding blocks {n_blocks} = {block_size}, "
f"but got input.size(2)={input_length}.",
)
def col2im_naive_implementation(data, image_shape, kernel_shape, dilations, pads, strides): # type: ignore
"""Naive implementation for `col2im`."""
n_dims = len(pads) // 2
new_pads = np.array([(pads[i], pads[i + n_dims]) for i in range(n_dims)])
_col2im_shape_check(data, image_shape, kernel_shape, dilations, new_pads, strides)
data_col = data
data_im = np.zeros(image_shape, dtype=data.dtype)
dim_col = []
for i in range(n_dims):
col = (
image_shape[i]
+ new_pads[i, :].sum()
- (dilations[i] * (kernel_shape[i] - 1) + 1)
) // strides[i] + 1
dim_col.append(col)
kernel_size = np.prod(kernel_shape)
col_size = np.prod(dim_col)
for c_col in range(kernel_size):
offset = _get_indices(c_col, kernel_shape)
for col in range(col_size):
ind_col = _get_indices(col, dim_col)
ind_im = []
for i in range(n_dims):
ind = (
ind_col[i] * strides[i] - new_pads[i, 0] + offset[i] * dilations[i]
)
ind_im.append(ind)
if not _is_out(ind_im, data_im.shape):
data_im[tuple(ind_im)] += data_col[c_col, col]
return data_im
class Col2Im(OpRun):
def _run(self, data, image_shape, block_shape, dilations=None, pads=None, strides=None): # type: ignore
if dilations is None:
dilations = [1 for s in image_shape]
if pads is None:
pads = [0 for s in image_shape] * 2
if strides is None:
strides = [1 for s in image_shape]
bl = np.prod(block_shape)
C = data.shape[1] // bl
data = data.reshape(data.shape[:1] + (C,) + (bl,) + data.shape[2:])
ks = tuple(block_shape)
res = None
for n in range(data.shape[0]):
for c in range(data.shape[1]):
out = col2im_naive_implementation(
data[n, c, ...], image_shape, ks, dilations, pads, strides
)
if res is None:
new_shape = data.shape[:2] + out.shape
res = np.empty(new_shape, dtype=data.dtype)
res[n, c, ...] = out
return (res,) # type: ignore
|