| 82 | |
| 83 | |
| 84 | class EarlyStopping(Callback): |
| 85 | def __init__( |
| 86 | self, |
| 87 | monitor="val_loss", |
| 88 | min_delta=0, |
| 89 | patience=0, |
| 90 | verbose=0, |
| 91 | mode="auto", |
| 92 | baseline=None, |
| 93 | restore_best_weights=False, |
| 94 | ): |
| 95 | """Create an early-stopping callback.""" |
| 96 | super(EarlyStopping, self).__init__() |
| 97 | self.monitor = monitor |
| 98 | self.min_delta = abs(min_delta) |
| 99 | self.patience = patience |
| 100 | self.verbose = verbose |
| 101 | self.mode = mode |
| 102 | self.baseline = baseline |
| 103 | self.restore_best_weights = restore_best_weights |
| 104 | |
| 105 | if mode not in {"auto", "min", "max"}: |
| 106 | raise ValueError("mode should be one of {'auto', 'min', 'max'}") |
| 107 | |
| 108 | if mode == "min": |
| 109 | self.monitor_op = np.less |
| 110 | elif mode == "max": |
| 111 | self.monitor_op = np.greater |
| 112 | else: |
| 113 | if "acc" in self.monitor or self.monitor.endswith("auc") or self.monitor.startswith("fmeasure"): |
| 114 | self.monitor_op = np.greater |
| 115 | else: |
| 116 | self.monitor_op = np.less |
| 117 | |
| 118 | def on_train_begin(self, logs=None): |
| 119 | self.wait = 0 |
| 120 | self.stopped_epoch = 0 |
| 121 | self.best_weights = None |
| 122 | if self.baseline is not None: |
| 123 | self.best = self.baseline |
| 124 | else: |
| 125 | self.best = np.inf if self.monitor_op == np.less else -np.inf |
| 126 | |
| 127 | def _is_improvement(self, current, best): |
| 128 | if self.monitor_op == np.less: |
| 129 | return current < (best - self.min_delta) |
| 130 | return current > (best + self.min_delta) |
| 131 | |
| 132 | def on_epoch_end(self, epoch, logs=None): |
| 133 | logs = logs or {} |
| 134 | current = logs.get(self.monitor) |
| 135 | if current is None: |
| 136 | return |
| 137 | |
| 138 | if self._is_improvement(current, self.best): |
| 139 | self.best = current |
| 140 | self.wait = 0 |
| 141 | if self.restore_best_weights and self.model is not None: |
no outgoing calls