Convolution operator in NCHW layout. Parameters ---------- a_np : numpy.ndarray 4-D with shape [batch, in_channel, in_height, in_width] w_np : numpy.ndarray 4-D with shape [num_filter, in_channel, filter_height, filter_width] stride : int or a list/tuple of two
(a_np, w_np, stride, padding)
| 25 | |
| 26 | |
| 27 | def _conv2d_nchw_python(a_np, w_np, stride, padding): |
| 28 | """Convolution operator in NCHW layout. |
| 29 | |
| 30 | Parameters |
| 31 | ---------- |
| 32 | a_np : numpy.ndarray |
| 33 | 4-D with shape [batch, in_channel, in_height, in_width] |
| 34 | |
| 35 | w_np : numpy.ndarray |
| 36 | 4-D with shape [num_filter, in_channel, filter_height, filter_width] |
| 37 | |
| 38 | stride : int or a list/tuple of two ints |
| 39 | Stride size, or [stride_height, stride_width] |
| 40 | |
| 41 | padding : int or str or a list/tuple of 2 or 4 ints |
| 42 | Padding size, or ['VALID', 'SAME'], or |
| 43 | [pad_height, pad_width] for 2 ints, or |
| 44 | [pad_top, pad_left, pad_bottom, pad_right] for 2 ints |
| 45 | |
| 46 | Returns |
| 47 | ------- |
| 48 | b_np : np.ndarray |
| 49 | 4-D with shape [batch, out_channel, out_height, out_width] |
| 50 | """ |
| 51 | batch, in_channel, in_height, in_width = a_np.shape |
| 52 | num_filter, _, kernel_h, kernel_w = w_np.shape |
| 53 | if isinstance(stride, int): |
| 54 | stride_h = stride_w = stride |
| 55 | else: |
| 56 | stride_h, stride_w = stride |
| 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 | b_np = np.zeros((batch, out_channel, out_height, out_width), dtype=a_np.dtype) |
| 65 | # computation |
| 66 | for n in range(batch): |
| 67 | for f in range(out_channel): |
| 68 | for c in range(in_channel): |
| 69 | if pad_h > 0 or pad_w > 0: |
| 70 | apad = np.zeros((in_height + pad_h, in_width + pad_w), dtype=a_np.dtype) |
| 71 | apad[pad_top : pad_top + in_height, pad_left : pad_left + in_width] = a_np[n, c] |
| 72 | else: |
| 73 | apad = a_np[n, c] |
| 74 | |
| 75 | out = _conv2d_hw(apad, w_np[f, c]) |
| 76 | b_np[n, f] += out[::stride_h, ::stride_w] |
| 77 | return b_np |
| 78 | |
| 79 | |
| 80 | def _conv2d_hw(apad, w_np_fc): |
no test coverage detected
searching dependent graphs…