Training Problem about cellseg_models.pytorch HOT 15 CLOSED

Hi,

This is a pytorch-lightning error. Basically, you just need to pass n_devices argument to the pl.Trainer to fix this.

from cellseg_models.pytorch.

Thank you for your help. But I seem to have encountered another problem. How to import a trained model into prediction。
When I use the following code
experiment = experiment.load_state_dict(torch.load('./version_0/checkpoints/epoch=149-step=49800.ckpt'))

from cellseg_models.pytorch.

Can you be a bit more specific so I can get a hang of what you're trying to achieve. In general, you should not try to load the state dict to theSegmentationExperiment object. Instead, you should load the weights into a nn.Module object and use that as an input to the SegmentationExperiment

from cellseg_models.pytorch.

I have already solved this problem. I would like to ask how to save the best model results on the validation set. In addition, may I ask you how to validate a test set? Thank you very much.

from cellseg_models.pytorch.

You can save your checkpoints with pytorch-lightning ModelCheckpoint class. Ory ou can just add a ckpt_path to the trainenr.fit method: trainer.fit(model, ckpt_path="some/path/to/my_checkpoint.ckpt"). Check out the lightning checkpointing documentation: https://lightning.ai/docs/pytorch/stable/common/checkpointing_intermediate.html

from cellseg_models.pytorch.

For benchmarking: see this benchmarking notebook: https://github.com/okunator/cellseg_models.pytorch/blob/main/examples/pannuke_cellpose_benchmark.ipynb

from cellseg_models.pytorch.

Thank you very much. I'm going to learn now

from cellseg_models.pytorch.

Hello, I'm training pannuke_nuclei_segmentation_stardist.ipynb, I found that you have enhanced the data used for images and masks. The first dimension of the mask is a cell count map, and the second dimension is a nuclear category map. However, the first dimension is used in your code. Is there a problem?

from cellseg_models.pytorch.

Should we use a second dimensional cell category mask instead of a cell technology mask?

from cellseg_models.pytorch.

I trained 150 epochs using default parameters, and the test results seemed very unsatisfactory.pannuke_nuclei_segmentation_stardist.ipynb

from cellseg_models.pytorch.

Hello, I'm training pannuke_nuclei_segmentation_stardist.ipynb, I found that you have enhanced the data used for images and masks. The first dimension of the mask is a cell count map, and the second dimension is a nuclear category map. However, the first dimension is used in your code. Is there a problem?

No there is not a problem, the data flow is such that the instance mask and the semantic masks are transformed separately since they require different kinds of transformations. I'm not quite sure what you mean about enhancing data but applying augmentations during the data loading is a routine practice and if you wish to not use augmentations you don't have to

from cellseg_models.pytorch.

I trained 150 epochs using default parameters, and the test results seemed very unsatisfactory.pannuke_nuclei_segmentation_stardist.ipynb

Stardist is not easy to train on multi-class problems and requires a lot of regularization or other bias mitigation since the learning of the radial distance maps dominate the gradient updates during training. You might want to try out cellpose or hover-net instead for pannuke data.

from cellseg_models.pytorch.

Thank you for your help

from cellseg_models.pytorch.

Hello, but I used the training set to test indicators and found that there are two categories that are also 0. Is there any problem?

from cellseg_models.pytorch.

No, it just mean that the model did not learn to recognize those categories.

from cellseg_models.pytorch.