ML snippets

This is where I save some useful ML code and snippets from different sources.

1. Huggingface nlp+transformers+imdb->sentiment analysis

This is based on:

https://www.youtube.com/watch?v=G3pOvrKkFuk&list=PL1v8zpldgH3pQwRz1FORZdChMaNZaR3pu&index=17

And the address:

https://github.com/yk/huggingface-nlp-demo/blob/master/demo.py

This is a useful demo for huggingface tools, including:

• nlp for datasets
• transformers for model zoo

Current Performance ~90% validation acc

2. DETR Facebook object detection

This is based on:

https://www.youtube.com/watch?v=LfUsGv-ESbc

And:

https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_demo.ipynb#scrollTo=bawoKsBC7oBh

This official DETR github:

https://github.com/facebookresearch/detr

What is interesting:

• Cnn+transformers
• Get around those computational extensive bounding boxes per pixel calculation and use transformers to propose bounding boxes

Limitation:

• Need to have a score threshold

3. DETR Panoptic segmentation detection

This is based on:

https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/DETR_panoptic.ipynb

Need to see that is the difference between this and Umap, but it looks fantastic!

• Detectron2 is broken on colab even with its original colab code, I'll try to use numpy and matplotlib to fix this

4. Generate Fake pandas dataframe

This is a short snippet I found on:

https://zhuanlan.zhihu.com/p/394857784 This is usefull since it is short and easily customizable. It can be used as:

generate_fake_dataframe(1000,cols='ifcd')

5. DEMO in openCV for HDR

This is a short DEMO from the openCV online tutorial:

https://opencv.org/opencv-python-free-course/

This shows how HDR works, with the builtin openCV processing alglrithms

6. DEEP INTEREST NETWORK in Pytorch

Official github link: https://github.com/zhougr1993/DeepInterestNetwork I use code from here as a reference: https://github.com/fanoping/DIN-pytorch