PIDM
PIDM copied to clipboard
ankanbhunia
Using the pre-trained model for market-1501 dataset
Hope this message finds you well!
I wanted to try this model with the Market 1501 dataset, but I don't want to train the model from scratch.
I was reading that I can use the concept of transfer learning/fine-tune/retraining. Could you please help me with the steps I need to follow to be able to do that.
I'm just confused about which code I have to use to do the fine tuning as the model consist of multiple parts. Also, how I can prepare the new dataset in terms of keypoints and target poses.
I'm asking these questions since you have already experienced the model with this dataset as mentioned in the paper
Your help will be highly appreciated!
Market-1501 has different image-size than the deepfashion dataset with significantly lower resolution images. Not only that it also has complex real-world background. So, I am not sure if fine-tune would help here. But, its doable.
You can start by creating the dataset first. The codes for that will be available in this repo soon. For now you can follow GFLA repository and associated codebase (here) to adapt the dataloader module.
Then in the train.py file, you can uncomment the following line and add the checkpoint path of the pretrained model.
https://github.com/ankanbhunia/PIDM/blob/f769c6256020fa196d7193ae959163a5f6cbed9d/train.py#L334
Thank you so much for your information. But it is still unclear to me how to create the train_pairs.txt and the test_pairs.txt specially in the case of market 1501 dataset. I'm not able to follow if these files are created by the code or just called by it.
Again, thank you so much for your effort.
Also, when I used openpose, which output format have you used? when I tried the json format there are many parameters and it doesn't look like the one available in pose.rar
@Maram-Helmy @ankanbhunia how to create the train_pairs.txt and the test_pairs ？
@ankanbhunia can you share the pre-trained weights for Market1501 dataset?
