mewc-train v2.0.6

binding Dockerfile to mewc-flow patch version (e.g. v2.0.3) instead o…

mewc-train v2.0.5

Bugfix for unfreeze layers

Another missed occurrence of "vit" rather than "vt". Also fixed to look for first two characters in name rather than anywhere in name, as per the new standard across the code.

I think I've got all of those cases now!

mewc-train v2.0.4

Standarised DNN model names

Standardised all models to a 3-letter code. This impacts lib_common.py in mewc-flow and lib_model.py and config.yaml (comments) in mewc-train.

mewc-train v2.0.3

Efficiency fixes and refinements

- Refinement to model top.
- Fix for image size for base models loaded from file (see also mewc-flow).
- Removed unneccessary and confusing saving of final-epoch model (best model should always be used to avoid overfitting).
- Freezes best model after reloading in confusion matrix eval. This keeps the model in inference mode resulting in more efficient inference with a much smaller file size.

mewc-train v2.0.2

remove comment from CLASS_SAMPLES_SPECIFIC so env vars can be read co…

mewc-train v2.0.0

Version 2 Updates

The mewc-train Docker image has been updated to version 2. Key updates include:

• Base Image: Uses the new mewc-flow base image featuring tensorflow/tensorflow:2.16.1-gpu, CUDA, cuDNN, and JAX.
• Hugging Face Base Models: Start training off a base model from Hugging Face.
• Improved Training Control: New user configurable options to fine tune model training.
• Optional Validation Path: There is an option to specify separate paths for data|validation VAL_PATH and data|test TEST_PATH. Doing so will keep test data isolated from hyper-parameter tuning leakage.