| 8 | |
| 9 | |
| 10 | def macro_f1(pred, targ, num_classes=None): |
| 11 | pred = th.max(pred, 1)[1] |
| 12 | tp_out = [] |
| 13 | fp_out = [] |
| 14 | fn_out = [] |
| 15 | if num_classes is None: |
| 16 | num_classes = sorted(set(targ.cpu().numpy().tolist())) |
| 17 | else: |
| 18 | num_classes = range(num_classes) |
| 19 | for i in num_classes: |
| 20 | tp = ((pred == i) & (targ == i)).sum().item() # 预测为i，且标签的确为i的 |
| 21 | fp = ((pred == i) & (targ != i)).sum().item() # 预测为i，但标签不是为i的 |
| 22 | fn = ((pred != i) & (targ == i)).sum().item() # 预测不是i，但标签是i的 |
| 23 | tp_out.append(tp) |
| 24 | fp_out.append(fp) |
| 25 | fn_out.append(fn) |
| 26 | |
| 27 | eval_tp = np.array(tp_out) |
| 28 | eval_fp = np.array(fp_out) |
| 29 | eval_fn = np.array(fn_out) |
| 30 | |
| 31 | precision = eval_tp / (eval_tp + eval_fp) |
| 32 | precision[np.isnan(precision)] = 0 |
| 33 | precision = np.mean(precision) |
| 34 | |
| 35 | recall = eval_tp / (eval_tp + eval_fn) |
| 36 | recall[np.isnan(recall)] = 0 |
| 37 | recall = np.mean(recall) |
| 38 | |
| 39 | f1 = 2 * (precision * recall) / (precision + recall) |
| 40 | return f1, precision, recall |
| 41 | |
| 42 | |
| 43 | def accuracy(pred, targ): |