Getting the negative loss in fastflow

[Haibara-z]

Epoch 0: 67%|██████▋ | 2/3 [00:14<00:07, 7.09s/it, loss=2.14e+05, train_loss_step=1.94e+5] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 0: 100%|██████████| 3/3 [00:22<00:00, 7.52s/it, loss=2.14e+05, train_loss_step=1.94e+5, pixel_F1Score=0.163, pixel_AUROC=0.966] Epoch 1: 67%|██████▋ | 2/3 [00:15<00:07, 7.59s/it, loss=1.84e+05, train_loss_step=1.4e+5, pixel_F1Score=0.163, pixel_AUROC=0.966, train_loss_epoch=2.19e+5] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 1: 100%|██████████| 3/3 [00:24<00:00, 8.13s/it, loss=1.84e+05, train_loss_step=1.4e+5, pixel_F1Score=0.184, pixel_AUROC=0.972, train_loss_epoch=2.19e+5] Epoch 2: 67%|██████▋ | 2/3 [00:15<00:07, 7.67s/it, loss=1.58e+05, train_loss_step=9.02e+4, pixel_F1Score=0.184, pixel_AUROC=0.972, train_loss_epoch=1.56e+5] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 2: 100%|██████████| 3/3 [00:23<00:00, 7.93s/it, loss=1.58e+05, train_loss_step=9.02e+4, pixel_F1Score=0.194, pixel_AUROC=0.973, train_loss_epoch=1.56e+5] Epoch 3: 67%|██████▋ | 2/3 [00:15<00:07, 7.51s/it, loss=1.33e+05, train_loss_step=4.77e+4, pixel_F1Score=0.194, pixel_AUROC=0.973, train_loss_epoch=1.09e+5] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 3: 100%|██████████| 3/3 [00:23<00:00, 7.87s/it, loss=1.33e+05, train_loss_step=4.77e+4, pixel_F1Score=0.205, pixel_AUROC=0.974, train_loss_epoch=1.09e+5] Epoch 4: 67%|██████▋ | 2/3 [00:14<00:07, 7.41s/it, loss=1.11e+05, train_loss_step=1.51e+4, pixel_F1Score=0.205, pixel_AUROC=0.974, train_loss_epoch=6.39e+4] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 4: 100%|██████████| 3/3 [00:23<00:00, 7.90s/it, loss=1.11e+05, train_loss_step=1.51e+4, pixel_F1Score=0.206, pixel_AUROC=0.973, train_loss_epoch=6.39e+4] Epoch 5: 67%|██████▋ | 2/3 [00:15<00:07, 7.66s/it, loss=9.02e+04, train_loss_step=-2.12e+4, pixel_F1Score=0.206, pixel_AUROC=0.973, train_loss_epoch=2.4e+4] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 5: 100%|██████████| 3/3 [00:24<00:00, 8.20s/it, loss=9.02e+04, train_loss_step=-2.12e+4, pixel_F1Score=0.205, pixel_AUROC=0.972, train_loss_epoch=2.4e+4] Epoch 6: 67%|██████▋ | 2/3 [00:15<00:07, 7.77s/it, loss=7.06e+04, train_loss_step=-5.37e+4, pixel_F1Score=0.205, pixel_AUROC=0.972, train_loss_epoch=-1.29e+4] Validation: 0it [00:00, ?it/s] Validation: 0%| | 0/1 [00:00<?, ?it/s] Validation DataLoader 0: 0%| | 0/1 [00:00<?, ?it/s] Epoch 6: 100%|██████████| 3/3 [00:24<00:00, 8.30s/it, loss=7.06e+04, train_loss_step=-5.37e+4, pixel_F1Score=0.198, pixel_AUROC=0.970, train_loss_epoch=-1.29e+4] Epoch 6: 100%|██████████| 3/3 [00:24<00:00, 8.30s/it, loss=7.06e+04, train_loss_step=-5.37e+4, pixel_F1Score=0.198, pixel_AUROC=0.970, train_loss_epoch=-4.54e+4]

[jpcbertoldo]

I also noticed anomaly_maps with negative values using patchcore.

Looking at the loss and the anomaly map code, they both use hidden_variables as input.

https://github.com/openvinotoolkit/anomalib/blob/52bbb0b2417461ff097a819a06cc7ac3ff408149/src/anomalib/models/fastflow/loss.py#L12

        loss = torch.tensor(0.0, device=hidden_variables[0].device)  # pylint: disable=not-callable
        for hidden_variable, jacobian in zip(hidden_variables, jacobians):
            loss += torch.mean(0.5 * torch.sum(hidden_variable**2, dim=(1, 2, 3)) - jacobian)
        return loss

https://github.com/openvinotoolkit/anomalib/blob/52bbb0b2417461ff097a819a06cc7ac3ff408149/src/anomalib/models/fastflow/anomaly_map.py#L21

        flow_maps: list[Tensor] = []
        for hidden_variable in hidden_variables:
            log_prob = -torch.mean(hidden_variable**2, dim=1, keepdim=True) * 0.5
            prob = torch.exp(log_prob)
            flow_map = F.interpolate(
                input=-prob,
                size=self.input_size,
                mode="bilinear",
                align_corners=False,
            )
            flow_maps.append(flow_map)
        flow_maps = torch.stack(flow_maps, dim=-1)
        anomaly_map = torch.mean(flow_maps, dim=-1)

I guess the loss could be negative because of the - jacobian ? Or am i missing some (assumed) property of the jacobian and the other term?

However I don't see why there is that - prob (instead of just prob) in the anomaly map interpolation, which seems to match the original implementation:

https://github.com/gathierry/FastFlow/blob/2cf1f2f4c562a7f13cfb1959e3afe5df2f2d2565/fastflow.py#L145

            anomaly_map_list = []
            for output in outputs:
                log_prob = -torch.mean(output**2, dim=1, keepdim=True) * 0.5
                prob = torch.exp(log_prob)
                a_map = F.interpolate(
                    -prob,
                    size=[self.input_size, self.input_size],
                    mode="bilinear",
                    align_corners=False,
                )
                anomaly_map_list.append(a_map)
            anomaly_map_list = torch.stack(anomaly_map_list, dim=-1)
            anomaly_map = torch.mean(anomaly_map_list, dim=-1)
            ret["anomaly_map"] = anomaly_map

[TSTB-dev]

I've encountered an issue where FastFlow outputs negative loss values. Since the loss is essentially the "Negative Log-likelihood," its value should be positive. I suspect that the precise computation in the Normalizing Flow is not completely invertible(May be invertible1x1conv?).

Interestingly, while NFs produce positive values, the performance of anomaly detection remains impressively high.