Mohammed Innat

**Update** It happens in a special case. Let's say I've the following branches: - main - branch_1 - branch_2 - branch_3 And, currently, I'm on `branch_2`. But next, I want...

It's `git version 2.20.1`

```yaml Config option `kernel_spec_manager_class` not recognized by `ListLabExtensionsApp`. [W 2022-04-13 11:20:05.946 LabApp] Config option `kernel_spec_manager_class` not recognized by `LabApp`. JupyterLab v3.2.9 /opt/conda/share/jupyter/labextensions nbdime-jupyterlab v2.1.1 enabled OK jupyterlab-jupytext v1.3.8 enabled OK...

Hardly anyone can help with the provided information. However, a few things you should do such as - grayscale your images. - Use learning rate let's say 0.001 (don't panic,...

IMO, this feature is not needed as we can implement gradient accumulation in the custom training in `tf 2`. [Link](https://stackoverflow.com/a/62683800/9215780).

That should also possible to achieve by overriding the `train_step` method, customizing the `.fit` function.

If you want to plug and play, then try this. https://github.com/CyberZHG/keras-gradient-accumulation

Overriding the `train_step` doesn't necessarily refer to a custom training loop, [link](https://keras.io/guides/customizing_what_happens_in_fit/). This is the leverage we get from new `tf.keras`, it should be adopted. Being too much plug-and-play stuff...

@dathudeptrai understood. It sounds great then. However, I'm facing some issues with implementing GA by customizing the `.fit`. In case you're interested, please have a look here https://github.com/tensorflow/tensorflow/issues/47578. ### update...

Yeah, looks so. Models like patch-core, or fast-flow (transformer based) left less room for improvement on this dataset.