Learnable PINs: Cross-Modal Embeddings for Person Identity
This code is based on the Self-Lifting project which enables train a model just in minutes.
Note that the dataset splitting is VGG-Vox style, which is different from the original paper,
but you can still know how the Curriculum-based Mining is implemented 😜 (utils/pair_selection_util.py
).
The dataset is the same as the Self-Lifting project. If you already have it, you can just create a soft link in the project root:
ln -s Your-Self-Lifting-Project-Root/dataset ./dataset
Or you need to download it by referring to Self-Lifting.
Just Run: python 1_pins.py
You also can use wandb to view the training process:
-
Create
wb_config.json
file in the./configs
folder, using the following content:{ "WB_KEY": "Your wandb auth key" }
-
add
--dryrun=False
to the training command, for example:python 1_pins.py --dryrun=False
Because the Backbone structure and test script are different from the original paper, the scores behave much higher.