Pytorch implement for Character-independent font identification
Dataset:
- Images: 1016 font of 64 character images from google font
- Train csv: pair images for training
- Val csv: pair images for validation
- Test csv: pair images for testing
Config file:
- 2flow.yaml: original paper. Two images forward a shared weights network
- 1flow.yaml: I modify the flow that the combine pair images before forwarding the network
-
Instead only use the Conv, I use sequential of [Conv, BatchNorm, ReLU layer] and the data padded to same size 64x64 pixel
-
Combine 1flow data pipeline and sequential layer, it push the accuracy from 0.82 -> 0.93 (for my dataset)