We list some common troubles faced by many users and their corresponding solutions here. Feel free to enrich the list if you find any frequent issues and have ways to help others to solve them. If the contents here do not cover your issue, please create an issue using the [provided templates](https://github.com/open-mmlab/mmsegmentation/blob/master/.github/ISSUE_TEMPLATE/error-report.md/) and make sure you fill in all required information in the template.
- Infer from the name of the config file of the model. You can refer to the `Config Name Style` part of [Learn about Configs](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/tutorials/config.md). For example, for config file with name `segformer_mit-b0_8x1_1024x1024_160k_cityscapes.py`, `8x1` means training the model corresponding to it needs 8 GPUs, and the batch size of each GPU is 1.
- Infer from the log file. Open the log file of the model and search `nGPU` in the file. The number of figures following `nGPU` is the number of GPUs needed to train the model. For instance, searching for `nGPU` in the log file yields the record `nGPU 0,1,2,3,4,5,6,7`, which indicates that eight GPUs are needed to train the model.
Briefly, it is a deep supervision trick to improve the accuracy. In the training phase, `decode_head` is for decoding semantic segmentation output, `auxiliary_head` is just adding an auxiliary loss, the segmentation result produced by it has no impact to your model's result, it just works in training. You may read this [paper](https://arxiv.org/pdf/1612.01105.pdf) for more information.
In the train script, we call `get_root_logger`at Line 167, and `get_root_logger` in mmseg calls `get_logger` in mmcv, mmcv will return the same logger which has been initialized in 'mmsegmentation/tools/train.py' with the parameter `log_file`. There is only one logger (initialized with `log_file`) during training.
MMSegmentation uses `num_classes` and `out_channels` to control output of last layer `self.conv_seg`. More details could be found [here](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/models/decode_heads/decode_head.py).
`num_classes` should be the same as number of types of labels, in binary segmentation task, dataset only has two types of labels: foreground and background, so `num_classes=2`. `out_channels` controls the output channel of last layer of model, it usually equals to `num_classes`.
But in binary segmentation task, there are two solutions:
- Set `out_channels=2`, using Cross Entropy Loss in training, using `F.softmax()` and `argmax()` to get prediction of each pixel in inference.
- Set `out_channels=1`, using Binary Cross Entropy Loss in training, using `F.sigmoid()` and `threshold` to get prediction of each pixel in inference. `threshold` is set 0.3 as default.
In summary, to implement binary segmentation methods users should modify below parameters in the `decode_head` and `auxiliary_head` configs. Here is a modification example of [pspnet_unet_s5-d16.py](https://github.com/open-mmlab/mmsegmentation/blob/master/configs/_base_/models/pspnet_unet_s5-d16.py):
- (1) `num_classes=2`, `out_channels=2` and `use_sigmoid=False` in `CrossEntropyLoss`.
When [loading annotation](https://github.com/open-mmlab/mmsegmentation/blob/master/mmseg/datasets/pipelines/loading.py#L91) in MMSegmentation, `reduce_zero_label (bool)` is provided to determine whether reduce all label value by 1:
```python
if self.reduce_zero_label:
# avoid using underflow conversion
gt_semantic_seg[gt_semantic_seg == 0] = 255
gt_semantic_seg = gt_semantic_seg - 1
gt_semantic_seg[gt_semantic_seg == 254] = 255
```
**Noted:** Please pay attention to label numbers of dataset when using `reduce_zero_label`. If dataset only has two types of labels (i.e., label 0 and 1), it needs to close `reduce_zero_label`, i.e., set `reduce_zero_label=False`.
