search

faustomorales / vit-keras

Keras implementation of ViT (Vision Transformer)
Apache License 2.0
329 stars 78 forks source link

Minimum image size to use pretrained weights #41

Closed mb1069 closed 8 months ago

mb1069 commented 8 months ago

Hi, I have some very small images from an optical system (15x15 px), and realise I can get a huge performance boost by transfer learning from pretrained weights.

Is the minimum size 128x128? Or can smaller sizes be used?

mb1069 commented 8 months ago

For anyone with the same problem: I ended up using pre-trained weights and an example image. By downsizing the image, I found the model still made the same prediction using pre-trained weights, until the image was so downsized it was no longer recognisable. So, I believe the minimum size is significantly smaller than 128x128 (64x64 still worked).