|
| 1 | +# Copyright 2024 Arm Limited and/or its affiliates. |
| 2 | +# |
| 3 | +# This source code is licensed under the BSD-style license found in the |
| 4 | +# LICENSE file in the root directory of this source tree. |
| 5 | + |
| 6 | +import logging |
| 7 | + |
| 8 | +import torch |
| 9 | + |
| 10 | +logger = logging.getLogger(__name__) |
| 11 | + |
| 12 | + |
| 13 | +def _print_channels(result, reference, channels_close, C, H, W, rtol, atol): |
| 14 | + |
| 15 | + output_str = "" |
| 16 | + for c in range(C): |
| 17 | + if channels_close[c]: |
| 18 | + continue |
| 19 | + |
| 20 | + max_diff = torch.max(torch.abs(reference - result)) |
| 21 | + exp = f"{max_diff:2e}"[-3:] |
| 22 | + output_str += f"channel {c} (e{exp})\n" |
| 23 | + |
| 24 | + for y in range(H): |
| 25 | + res = "[" |
| 26 | + for x in range(W): |
| 27 | + if torch.allclose(reference[c, y, x], result[c, y, x], rtol, atol): |
| 28 | + res += " . " |
| 29 | + else: |
| 30 | + diff = (reference[c, y, x] - result[c, y, x]) / 10 ** (int(exp)) |
| 31 | + res += f"{diff: .2f} " |
| 32 | + |
| 33 | + # Break early for large widths |
| 34 | + if x == 16: |
| 35 | + res += "..." |
| 36 | + break |
| 37 | + |
| 38 | + res += "]\n" |
| 39 | + output_str += res |
| 40 | + |
| 41 | + return output_str |
| 42 | + |
| 43 | + |
| 44 | +def _print_elements(result, reference, C, H, W, rtol, atol): |
| 45 | + output_str = "" |
| 46 | + for y in range(H): |
| 47 | + res = "[" |
| 48 | + for x in range(W): |
| 49 | + result_channels = result[:, y, x] |
| 50 | + reference_channels = reference[:, y, x] |
| 51 | + |
| 52 | + n_errors = 0 |
| 53 | + for a, b in zip(result_channels, reference_channels): |
| 54 | + if not torch.allclose(a, b, rtol, atol): |
| 55 | + n_errors = n_errors + 1 |
| 56 | + |
| 57 | + if n_errors == 0: |
| 58 | + res += ". " |
| 59 | + else: |
| 60 | + res += f"{n_errors} " |
| 61 | + |
| 62 | + # Break early for large widths |
| 63 | + if x == 16: |
| 64 | + res += "..." |
| 65 | + break |
| 66 | + |
| 67 | + res += "]\n" |
| 68 | + output_str += res |
| 69 | + |
| 70 | + return output_str |
| 71 | + |
| 72 | + |
| 73 | +def print_error_diffs( |
| 74 | + result: torch.Tensor | tuple, |
| 75 | + reference: torch.Tensor | tuple, |
| 76 | + quantization_scale=None, |
| 77 | + atol=1e-03, |
| 78 | + rtol=1e-03, |
| 79 | + qtol=0, |
| 80 | +): |
| 81 | + """ |
| 82 | + Prints the error difference between a result tensor and a reference tensor in NCHW format. |
| 83 | + Certain formatting rules are applied to clarify errors: |
| 84 | +
|
| 85 | + - Batches are only expanded if they contain errors. |
| 86 | + -> Shows if errors are related to batch handling |
| 87 | + - If errors appear in all channels, only the number of errors in each HW element are printed. |
| 88 | + -> Shows if errors are related to HW handling |
| 89 | + - If at least one channel is free from errors, or if C==1, errors are printed channel by channel |
| 90 | + -> Shows if errors are related to channel handling or single errors such as rounding/quantization errors |
| 91 | +
|
| 92 | + Example output of shape (3,3,2,2): |
| 93 | +
|
| 94 | + ############################ ERROR DIFFERENCE ############################# |
| 95 | + BATCH 0 |
| 96 | + . |
| 97 | + BATCH 1 |
| 98 | + [. . ] |
| 99 | + [. 3 ] |
| 100 | + BATCH 2 |
| 101 | + channel 1 (e-03) |
| 102 | + [ 1.85 . ] |
| 103 | + [ . 9.32 ] |
| 104 | +
|
| 105 | + MEAN MEDIAN MAX MIN (error as % of reference output range) |
| 106 | + 60.02% 55.73% 100.17% 19.91% |
| 107 | + ########################################################################### |
| 108 | +
|
| 109 | +
|
| 110 | + """ |
| 111 | + |
| 112 | + if isinstance(reference, tuple): |
| 113 | + reference = reference[0] |
| 114 | + if isinstance(result, tuple): |
| 115 | + result = result[0] |
| 116 | + |
| 117 | + if not result.shape == reference.shape: |
| 118 | + raise ValueError("Output needs to be of same shape") |
| 119 | + shape = result.shape |
| 120 | + |
| 121 | + match len(shape): |
| 122 | + case 4: |
| 123 | + N, C, H, W = (shape[0], shape[1], shape[2], shape[3]) |
| 124 | + case 3: |
| 125 | + N, C, H, W = (1, shape[0], shape[1], shape[2]) |
| 126 | + case 2: |
| 127 | + N, C, H, W = (1, 1, shape[0], shape[1]) |
| 128 | + case 1: |
| 129 | + N, C, H, W = (1, 1, 1, shape[0]) |
| 130 | + case _: |
| 131 | + raise ValueError("Invalid tensor rank") |
| 132 | + |
| 133 | + if quantization_scale is not None: |
| 134 | + atol += quantization_scale * qtol |
| 135 | + |
| 136 | + # Reshape tensors to 4D NCHW format |
| 137 | + result = torch.reshape(result, (N, C, H, W)) |
| 138 | + reference = torch.reshape(reference, (N, C, H, W)) |
| 139 | + |
| 140 | + output_str = "" |
| 141 | + for n in range(N): |
| 142 | + output_str += f"BATCH {n}\n" |
| 143 | + result_batch = result[n, :, :, :] |
| 144 | + reference_batch = reference[n, :, :, :] |
| 145 | + is_close = torch.allclose(result_batch, reference_batch, rtol, atol) |
| 146 | + if is_close: |
| 147 | + output_str += ".\n" |
| 148 | + else: |
| 149 | + channels_close = [None] * C |
| 150 | + for c in range(C): |
| 151 | + result_hw = result[n, c, :, :] |
| 152 | + reference_hw = reference[n, c, :, :] |
| 153 | + |
| 154 | + channels_close[c] = torch.allclose(result_hw, reference_hw, rtol, atol) |
| 155 | + |
| 156 | + if any(channels_close) or len(channels_close) == 1: |
| 157 | + output_str += _print_channels( |
| 158 | + result[n, :, :, :], |
| 159 | + reference[n, :, :, :], |
| 160 | + channels_close, |
| 161 | + C, |
| 162 | + H, |
| 163 | + W, |
| 164 | + rtol, |
| 165 | + atol, |
| 166 | + ) |
| 167 | + else: |
| 168 | + output_str += _print_elements( |
| 169 | + result[n, :, :, :], reference[n, :, :, :], C, H, W, rtol, atol |
| 170 | + ) |
| 171 | + |
| 172 | + reference_range = torch.max(reference) - torch.min(reference) |
| 173 | + diff = torch.abs(reference - result).flatten() |
| 174 | + diff = diff[diff.nonzero()] |
| 175 | + if not len(diff) == 0: |
| 176 | + diff_percent = diff / reference_range |
| 177 | + output_str += "\nMEAN MEDIAN MAX MIN (error as % of reference output range)\n" |
| 178 | + output_str += f"{torch.mean(diff_percent):<8.2%} {torch.median(diff_percent):<8.2%} {torch.max(diff_percent):<8.2%} {torch.min(diff_percent):<8.2%}\n" |
| 179 | + |
| 180 | + # Over-engineer separators to match output width |
| 181 | + lines = output_str.split("\n") |
| 182 | + line_length = [len(line) for line in lines] |
| 183 | + longest_line = max(line_length) |
| 184 | + title = "# ERROR DIFFERENCE #" |
| 185 | + separator_length = max(longest_line, len(title)) |
| 186 | + |
| 187 | + pre_title_length = max(0, ((separator_length - len(title)) // 2)) |
| 188 | + post_title_length = max(0, ((separator_length - len(title) + 1) // 2)) |
| 189 | + start_separator = ( |
| 190 | + "\n" + "#" * pre_title_length + title + "#" * post_title_length + "\n" |
| 191 | + ) |
| 192 | + output_str = start_separator + output_str |
| 193 | + end_separator = "#" * separator_length + "\n" |
| 194 | + output_str += end_separator |
| 195 | + |
| 196 | + logger.info(output_str) |
| 197 | + |
| 198 | + |
| 199 | +if __name__ == "__main__": |
| 200 | + import sys |
| 201 | + |
| 202 | + logging.basicConfig(stream=sys.stdout, level=logging.INFO) |
| 203 | + |
| 204 | + """ This is expected to produce the example output of print_diff""" |
| 205 | + torch.manual_seed(0) |
| 206 | + a = torch.rand(3, 3, 2, 2) * 0.01 |
| 207 | + b = a.clone().detach() |
| 208 | + logger.info(b) |
| 209 | + |
| 210 | + # Errors in all channels in element (1,1) |
| 211 | + a[1, :, 1, 1] = 0 |
| 212 | + # Errors in (0,0) and (1,1) in channel 1 |
| 213 | + a[2, 1, 1, 1] = 0 |
| 214 | + a[2, 1, 0, 0] = 0 |
| 215 | + |
| 216 | + print_error_diffs(a, b) |
0 commit comments