# **Lab 2: Prune MobileNet**

### <span style="color:Red;">* **04/02 Update:**</span>
The accuracy drops due to the executor runner using a different interpolation method.
We will test out your model's accuracy through the *.pth* model you have submitted instead.

#### **Revised Submission Guidelines**

Now you only need to hand-in:
* Your pruned model named ***mobilenet.pth***
* A filled out ***lab2.ipynb***, renamed to ***```<YourID>```.ipynb***

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

YourID.zip
    ├── model/
    │     └── mobilenet.pth (sparsified model from part 2)
    └── YourID.ipynb

For those who have already submitted their assignments following the previously outlined format, please be advised that there is no requirement for resubmission.

Your work will be accepted as is, and there will be no penalization for adhering to the initial guidelines.

---

<span style="color:Red;">**Due Date: 4/4 23:55**</span>

:::danger
**Reminder:** Part 2 may require a decent amount of time on experimenting and retraining model. Start early!
:::

## Introduction

This lab aims to prune the MobileNetV2 model, resulting in a significant speedup when running on Raspberry Pi while maintaining satisfactory 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.

> [lab2.ipynb](https://colab.research.google.com/drive/1_9Mo0qVsx4qZZUFiRbgYJpXeonoQqbWD?usp=sharing)

* Below is a MobileNetV2 with 96.3% accuarcy on CIFAR10, finetuned by TAs. Feel free to use it as a starting point for pruning. It's also fine if you want to use your own model:

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

You can load the above model with the following snippet:

    import torch
    from torchvision.models import mobilenet_v2

    model = torch.load('./mobilenetv2_0.963.pth', map_location="cpu")

## Part 1: Finegrained-Pruning (Unstructered Pruning) (50%)

In this part, you will learn the pruning basics by implementing various methods to sparsify the model.

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

## Part 2: Channel Pruning using Torch-Pruning (50% +)

You will apply modern channel pruning techniques on MobileNetV2 using the [Torch-Pruning library](https://github.com/VainF/Torch-Pruning). This library captures the dependencies between adjacent model layers, which allows us to perform consistent structural pruning easily.


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

## Hand-In Policy

You will need to hand-in:
* Two formats of your pruned model (***.pth***, ***.pte***)
* Your ***xnn_executor_runner*** built from the previous lab
* Fill out ***lab2.ipynb***, and rename it to ***```<YourID>```.ipynb***

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

YourID.zip
    ├── model/
    │     ├── mobilenet.pth (sparsified model from part 2)
    │     └── mobilenet.pte (sparsified model from part 2)
    ├── xnn_executor_runner (from the previous lab)
    └── YourID.ipynb

## Evaluation Criteria

Upon receiving your zip file, we will:
1. Unzip and add the dataset folder, ```data/test_batch.bin```, ensuring the correct folder structure:

    YourID
        ├── data/
        │     └── test_batch.bin
        ├── model/
        ├── xnn_executor_runner
        └── YourID.ipynb

2. Verify the structure of your pruned model by inspecting ```mobilenet.pth``` using [Netron](https://netron.app/). Your pruned model should originate from MobileNetV2 or you will get no points.
3. Obtain the number of MFLOPs of your pruned model using the provided code snippet:

    import torch
    import torch_pruning as tp
    from torchvision.models import mobilenet_v2

    model = torch.load('./model/mobilenet.pth', map_location="cpu")
    example_inputs = torch.zeros(1, 3, 224, 224).to(args.device)
    ops, size = tp.utils.count_ops_and_params(model, example_inputs=example_inputs)

    MFLOPs = ops/1e6
    print(MFLOPs)

4. Evaluate the accuracy of your pruned model using the following command:

    ./xnn_executor_runner --model_path ./model/mobilenet.pte

Accuracy and MFLOPs obtained above will be used in the grading formula for **Part 2**:

$$
Score = (\dfrac{200 - MFLOPs}{200 - 45} - 10\times Clamp(0.92-Accuracy,\ 0,\ 1)) \times 45\%
$$

- It is possible to obtain higher than 45%.

:::info
:warning: Note: Accuracy beyond 0.92 won't contribute to a higher score!
:::

:::danger
**Reminder:** Go through the evaluation steps yourself before submission.
:::

## Tips and Recommendations

* Retraining after pruning is essential to recover model's accuracy.
* A moderate number of epochs (around 10) should suffice for finetuning/retraining.
* Experiment with multiple pruning and retraining iterations.
* Experiment with learning rate of different order of magnitude.