Hybrid Model (CNN + Transformer)


Recently, I’ve made some revisions one of my previous project (Hybrid EfficientNet Swin-Transformer), hosted on kaggle; to the overall implementation and training strategies.

You may try :slightly_smiling_face:
(Recommend kaggle environment for efficient and smooth UX.)

Open In Colab

Some highlights

  • Enhanced code quality.
  • Applied augmentation from KerasCV (ported due to tf 2.6).
  • Applied Jax based augmentation in tf.data API.
  • Applied Gradient Accumulation techniques, cosine annealed warm -restart learning schedulers, etc.
  • Illustrate saving and reloading in SaveModel format.
  • Conversion in TF-Lite
  • Code demonstration on Kaggle, Colab, and Deepnote.
  • Quick Inference on Gradio (Huggingface Space).