Video Classification on COIN Dataset

This recipe is also explained in Medium Blog.

Introduction

In this recipe, we will classify cooking and decoration video clips with Pytorch.

Dataset

I selected 2 categories from COIN dataset. There are also sub-categories in primary categories. Selected categories:

Cooking: MakeSandwich, CookOmelet, MakePizza, MakeYoutiao, MakeBurger, MakeFrenchFries Decoration: AssembleBed, AssembleSofa, AssembleCabinet, AssembleOfficeChair

It is very easy to train with different data setups. You should only change target_label_list variable. Default configuration is:

target_label_list = [
    ["MakeSandwich", "CookOmelet", "MakePizza", "MakeYoutiao", "MakeBurger", "MakeFrenchFries"],
    ["AssembleBed", "AssembleSofa", "AssembleCabinet", "AssembleOfficeChair"],
]

There are 2 lists in target_label_list which are classes to classify. Internal lists contain actions in taxonomy.xlsx.

How to Run

You may changes a few lines in train.py

1. If you want, change target_label_list to set data classes.
2. If you want, change experiment path logger = Logger(exp_path="exps/exp1")
3. Change number of out dimension in model.blocks[5].proj = nn.Linear(2048, 2).to(device) according to step 1.
4. RUN python train.py