How to Quantize a ResNet from Scratch! Full Coding Tutorial (Eager Mode)

If you need help with anything quantization or ML related (e.g. debugging code) feel free to book a 30 minute consultation session! https://calendly.com/oscar-savolainen

I'm also available for long-term freelance work, e.g. for training / productionizing models, teaching AI concepts, etc.

Video Summary:
In this first of its kind video, we take a publicly available PyTorch ResNet model, and statistically quantize it entirely from scratch! We go through the coding steps of making architecture changes (QuantStubs/DeQuantStubs, FloatFunctionals), fusing modules, assigning qconfigs, preparing the model for fake-quant, and finally converting the model to a "true" int8 model. We find and squash some nasty bugs, go through some of the important documentation, and gain an understanding of how Eager Mode, Static Quantization works.

Timestamps:
00:00 Intro
03:19 Setting Up
06:35 Step 1: Architecture changes
10:11 Step 2: Fusing modules
31:35 Step 3: Assigning QConfigs
38:13 Step 4: Preparing the model for fake-quant
46:23 Evaluating our models
51:00 Solving the fusing bug
54:50 Static quantization, turning off observers
59:09 Step 5: Converting the model
01:00:38 Solving the conversion bug: adding FloatFunctionals

Links:
See this video for the theory on how to quantize a PyTorch model in Eager mode:    • How to statically quantize a PyTorch model...  
Github for the code from this tutorial: https://github.com/OscarSavolainen/Qu...
Useful PyTorch quantization Github link: https://github.com/pytorch/pytorch/tr...
NVIDIA white paper: https://arxiv.org/abs/2004.09602