Scaling mega-pixel images classification via attention sampling

In this repository can be found the code to reproduce the code for the case study Lymphoma subtype classification which uses attention sampling to drastically reduce the computational cost of mega pixel image classification

Highlight

The network is composed by 3 stages:

• Attention network: sees a downscaled version of the image as says where the information for the classification is likely to be located
• Patch sampling according to the distribution given by the attention network
• Classification network: unitary logits vector are extracted by the sampled patches and the final classification is given by a weighted sum of such logits passed though a softmax function

The attention network is trained using REINFORCE with baseline with and entropy regularizer to overcome the exploration/exploitation dilemma.

Structure

|- presentation.pdf : short presentation of the project
|- report.pdf : the full report of the work done
|- evaluation.ipynb : evaluation of the model
|- project.ipynb : code for the traning
|- model : folder with weights of the model

Dataset

Lymphoma subtype classification dataset downloadable at this link

Credits

Processing Megapixel Images with Deep Attention-Sampling Models , arXiv:1905.03711