Request for InternVideo2 Stage2 1B Hyperparameters

Hi, appreciate for your great work and thanks for releasing the code. I get some confusion when reproduce the InternVideo2-1B-s2, I am finding close to 42% (~10% lower) performance (R@1) than what is reported in the project. When I check the InternVideo2-stage2_1b-224p-f4.pt checkpoint, I have found some parameters to be different from those in the project.

• I found that the epoch of checkpoint is 3, and the global_samples is 4177920. So is it mean the InternVideo2-stage2_1b-224p-f4.pt checkpoint is only trained by ~1.3M data (25.5M reported in the project)?
• And the global_steps is 2040, so is the batch size should be 1392640 samples / (680 steps * 32 GPUs) = 64? So the global batch size is 64 * 32? I used 64 * 128 global batch size when reproducing. Is the difference between global batch size learning to the lower performance? I am wondering that how many GPUs you used in the stage2 pretrain of InternVideo2 1B.
• Is all the hyperparameters in [config] same as the setting when training InternVideo2-stage2_1b-224p-f4.pt.(https://github.com/OpenGVLab/InternVideo/blob/main/InternVideo2/multi_modality/scripts/pretraining/stage2/1B/config.py)

I would greatly appreciate your assistance in shedding light on this confusion.Thank you very much.

Start with InternVideo2/multi_modality/scripts/pretraining/stage2/1B/config.py and inspect the metadata reported for InternVideo2-stage2_1b-224p-f4.pt. Compare the checkpoint epoch, global_samples, global_steps, GPU count, batch size, and configuration against the project’s reported training setup. Done means documenting the exact Stage 2 1B hyperparameters and resolving which settings reproduce the reported performance.