Learning rate mismatch between supplementary and code (encoder LR / decoder LR)

[yslee94]

Hi, thanks for sharing the code and the paper!

I'm trying to reproduce the results and noticed a small mismatch between the paper(supplementary) and the code.

In the supplementary:
(p.2 bottom, Section 1.2 "Implementation Details of Report Generation")

In the first stage, following [8], we set the learning rate for the image encoder at 5e−4 and for the encoder projection layer and text decoder at 5e−5.

• encoder learning rate = 5e-4
• decoder learning rate = 5e-5

But in the code, the defaults seem swapped:

parser.add_argument("--encoder_lr", type=float, default=5e-5)
parser.add_argument("--decoder_lr", type=float, default=5e-4)

Could you please confirm which setting was used for the main experiments?🥹

Thank you!

[fuying-wang]

Hi,

Please use our released code, as the paper may contain an error.

Thank you.