search

amirhnazerii / Transformer_OD_TPU

Exploration of TPU Architectures for Optimized Transformer Performance in Image Detection of Drainage Crossings
Apache License 2.0
0 stars 0 forks source link

finetuning causes no detection #26

Open amirhnazerii opened 6 months ago

amirhnazerii commented 6 months ago

I have finetuned the model for 10 epochs. But it seems like the loss do not converge.

image

Would you please share the results of the script below with me: checkpoint['model']['class_embed.weight'] & checkpoint['model']['class_embed.weight'].size()

image

DWGodwin commented 5 months ago

Hi @amirhnazerii, as we discussed in our call, what we are doing with the model isn't quite fine-tuning - we are training the entire model from pretrained weights with no frozen weights.

For the sake of completeness, these weights are for class 1, which we have defined as drainage crossings: image

The result of checkpoint['model']['class_embed.weight'].size() is torch.Size([92, 256])