This recipe is also explained in Medium Blog.
In this recipe, we will classify cooking and decoration video clips with Pytorch.
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.
You may changes a few lines in train.py
- If you want, change
target_label_list
to set data classes. - If you want, change experiment path
logger = Logger(exp_path="exps/exp1")
- Change number of out dimension in
model.blocks[5].proj = nn.Linear(2048, 2).to(device)
according to step 1. - RUN
python train.py