(self, data_set, return_pred=False)
| 144 | return total_loss |
| 145 | |
| 146 | def test_epoch(self, data_set, return_pred=False): |
| 147 | self.model.eval() |
| 148 | self.tra.eval() |
| 149 | data_set.eval() |
| 150 | |
| 151 | preds = [] |
| 152 | metrics = [] |
| 153 | for batch in tqdm(data_set): |
| 154 | data, label, index = batch["data"], batch["label"], batch["index"] |
| 155 | |
| 156 | feature = data[:, :, : -self.tra.num_states] |
| 157 | hist_loss = data[:, : -data_set.horizon, -self.tra.num_states :] |
| 158 | |
| 159 | with torch.no_grad(): |
| 160 | hidden = self.model(feature) |
| 161 | pred, all_preds, prob = self.tra(hidden, hist_loss) |
| 162 | |
| 163 | L = (all_preds - label[:, None]).pow(2) |
| 164 | |
| 165 | L -= L.min(dim=-1, keepdim=True).values # normalize & ensure positive input |
| 166 | |
| 167 | data_set.assign_data(index, L) # save loss to memory |
| 168 | |
| 169 | X = np.c_[ |
| 170 | pred.cpu().numpy(), |
| 171 | label.cpu().numpy(), |
| 172 | ] |
| 173 | columns = ["score", "label"] |
| 174 | if prob is not None: |
| 175 | X = np.c_[X, all_preds.cpu().numpy(), prob.cpu().numpy()] |
| 176 | columns += ["score_%d" % d for d in range(all_preds.shape[1])] + [ |
| 177 | "prob_%d" % d for d in range(all_preds.shape[1]) |
| 178 | ] |
| 179 | |
| 180 | pred = pd.DataFrame(X, index=index.cpu().numpy(), columns=columns) |
| 181 | |
| 182 | metrics.append(evaluate(pred)) |
| 183 | |
| 184 | if return_pred: |
| 185 | preds.append(pred) |
| 186 | |
| 187 | metrics = pd.DataFrame(metrics) |
| 188 | metrics = { |
| 189 | "MSE": metrics.MSE.mean(), |
| 190 | "MAE": metrics.MAE.mean(), |
| 191 | "IC": metrics.IC.mean(), |
| 192 | "ICIR": metrics.IC.mean() / metrics.IC.std(), |
| 193 | } |
| 194 | |
| 195 | if return_pred: |
| 196 | preds = pd.concat(preds, axis=0) |
| 197 | preds.index = data_set.restore_index(preds.index) |
| 198 | preds.index = preds.index.swaplevel() |
| 199 | preds.sort_index(inplace=True) |
| 200 | |
| 201 | return metrics, preds |
| 202 | |
| 203 | def fit(self, dataset, evals_result=dict()): |
no test coverage detected