# **Lab 3: Quantize MobileNet**

<span style="color:Red;">**Due Date: 5/9 23:55**</span>

## Introduction

This lab aims to quantize the MobileNetV2 model, resulting in reduced model size and potentially improved inference speed without significant loss in accuracy.

* Please download the provided Jupyter Notebook file using the link below. Follow the prompts and hints provided within the notebook to fill in the empty blocks and answer the questions.

> [lab3.ipynb](https://colab.research.google.com/drive/1uO8nSfVe5SbY-DR0M4o3wBuhTwJZr4gG?usp=drive_link)

## Part 1: Linear Quatization Implementation (70%)

In this part, you will learn how linear quantization works and implement the quantized version of **Fully Connected Layer** (Linear) and **Convolution Layer** (Conv2d).

Refer to **"Part1"** in the provided notebook (ipynb).

## Part 2: Quantize MobileNetV2 and Export (30%)

You will quantize MobileNetV2 using XNNPack library, and you may try to deploy the quantized model on to Raspberry Pi.

Refer to **"Part2"** in the provided notebook (ipynb).

* Below is a MobileNetV2 with 96.3% accuarcy on CIFAR10, finetuned by TAs. You will be using the model's weights a starting point for quantization:

> [mobilenetv2_0.963.pth](https://drive.google.com/file/d/1k89xAqC1FETperw11xvpxSPGcEwMfZJh/view)

You can load the above model with the following snippet:

    import torch
    from torchvision.models.quantization import mobilenet_v2

    model = torch.load('./mobilenetv2_for_quant.pth', map_location="cpu")

:::success
We will not calculate your score by executorch pte, but you can deploy it onto Raspberry Pi to compare the execution time with the model before being quantized.
:::

## Hand-In Policy

You will need to hand-in:
* Your quantized model ***mobilenet_quantized.pth***
* Fill out ***lab3.ipynb***, and rename it to ***```<YourID>```.ipynb***

Please organize your submission files into a zip archive structured as follows:

YourID.zip
    ├── model/
    │     └── mobilenet_quantized.pth
    └── YourID.ipynb

## Evaluation Criteria

Upon receiving your zip file:
1. We will use following code to load your quantized model

    // Load the saved ExportedProgram
    loaded_quantized_ep = torch.export.load(pt2e_quantized_model_file_path)
    loaded_quantized_model = loaded_quantized_ep.module()

2. Evaluate the accuracy of your quantized model using the following code:

    acc = evaluate_model(loaded_quantized_model, test_loader, device)

$$
Score = (10 \times Step function(Accuracy-0.88) + 20 \times \dfrac{Accuracy - 0.88}{0.96 - 0.88})
$$
The reported accuracy must be higher than *88%* to obtain full score.