Image Captioning with Attention : How to Increase Speed and Quality of X-Ray Diagnostics
Here we have three main architectures implemented.
In the Show-Attend-Tell folder is the implementation of https://arxiv.org/pdf/1502.03044v1.pdf paper and it's versions combined with GPT-2 models.
In the On-the-Automatic-Generation-of-Medical-Imaging-Reports is the implementation of https://arxiv.org/pdf/1711.08195.pdf paper.
In the Transformer-Based-Generation folder we have Transformer-based implementation using the fairseq github.
All the models are implemented using the CheXNet weights. The weights that are used are at DGX at /raid/data/cxr14-2/DenseNet121_aug4_pretrain_WeightBelow1_1_0.829766922537.pkl. If you dont use the DGX server you should change the weights path in the following files:
- Transformer-Based-Generation/preprocess/preprocess_images_cxr.py
- Show-Attend-Tell/models.py
- On-the-Automatic-Generation-of-Medical-Imaging-Reports/models.py