grit

grit copied to clipboard

Hello author, I have some questions about freezing training and hope to get your reply I noticed that you mentioned freezing the backbone network and detector, I want to ask what the specific purpose of this is, in order to understand that you can give some specific instructions for freezing the backbone and detector, when reading the code I found that the backbone is included in the model, what is the purpose of freezing the backbone in this case

Thanks for your question. There are two main purposes:

1. To show that trade-off between finetuning the entire model (including the backbone) and finetuning the model partially (excluding the backbone). It is seen that finetuning the entire model improves the overall performance.
2. Although freezing the backbone decreases the captioning scores, the decrease is acceptable or the overall performance is acceptable (~140 CIDEr). Freezing the backbone also SPEED UP the training phase, which is suitable when you don't have so much compute.