MCPcopy Index your code
hub / github.com/apache/tvm / conv2d_hwcn_python

Function conv2d_hwcn_python

python/tvm/topi/testing/conv2d_hwcn_python.py:26–79  ·  view source on GitHub ↗

Convolution operator in HWCN layout. Parameters ---------- a_np : numpy.ndarray 4-D with shape [in_height, in_width, in_channel, batch] w_np : numpy.ndarray 4-D with shape [filter_height, filter_width, in_channel, num_filter] stride : int or a list/tuple of two

(a_np, w_np, stride, padding)

Source from the content-addressed store, hash-verified

24
25
26def conv2d_hwcn_python(a_np, w_np, stride, padding):
27 """Convolution operator in HWCN layout.
28
29 Parameters
30 ----------
31 a_np : numpy.ndarray
32 4-D with shape [in_height, in_width, in_channel, batch]
33
34 w_np : numpy.ndarray
35 4-D with shape [filter_height, filter_width, in_channel, num_filter]
36
37 stride : int or a list/tuple of two ints
38 Stride size, or [stride_height, stride_width]
39
40 padding : int or str or a list/tuple of 2 or 4 ints
41 Padding size, or ['VALID', 'SAME'], or
42 [pad_height, pad_width] for 2 ints, or
43 [pad_top, pad_left, pad_bottom, pad_right] for 2 ints
44
45 Returns
46 -------
47 b_np : np.ndarray
48 4-D with shape [out_height, out_width, out_channel, batch]
49 """
50 in_height, in_width, in_channel, batch = a_np.shape
51 kernel_h, kernel_w, _, num_filter = w_np.shape
52 if isinstance(stride, int):
53 stride_h = stride_w = stride
54 else:
55 stride_h, stride_w = stride
56
57 pad_top, pad_left, pad_bottom, pad_right = get_pad_tuple(padding, (kernel_h, kernel_w))
58 pad_h = pad_top + pad_bottom
59 pad_w = pad_left + pad_right
60 # compute the output shape
61 out_channel = num_filter
62 out_height = (in_height - kernel_h + pad_h) // stride_h + 1
63 out_width = (in_width - kernel_w + pad_w) // stride_w + 1
64 # change the layout from HWCN to NCHW
65 at = a_np.transpose((3, 2, 0, 1))
66 wt = w_np.transpose((3, 2, 0, 1))
67 bt = np.zeros((batch, out_channel, out_height, out_width))
68 # computation
69 for n in range(batch):
70 for f in range(out_channel):
71 for c in range(in_channel):
72 if pad_h > 0 or pad_w > 0:
73 apad = np.zeros((in_height + pad_h, in_width + pad_w))
74 apad[pad_top : pad_top + in_height, pad_left : pad_left + in_width] = at[n, c]
75 else:
76 apad = at[n, c]
77 out = scipy.signal.convolve2d(apad, np.rot90(np.rot90(wt[f, c])), mode="valid")
78 bt[n, f] += out[::stride_h, ::stride_w]
79 return bt.transpose((2, 3, 1, 0))

Callers

nothing calls this directly

Calls 3

get_pad_tupleFunction · 0.90
transposeMethod · 0.45
zerosMethod · 0.45

Tested by

no test coverage detected

Used in the wild real call sites across dependent graphs

searching dependent graphs…