salesforce / alpro Goto Github PK

In the video encoder part, the output is {v_cls, v_1, ..., v_k} (so the dimension is (k+1)*d)
therefore, the dimension of multi-modal video-text encoder is (k+N_t+1)*d
but according to paper: you claim that the dimension of multi-modal video-text encoder is (N_v+N_t+1)*d
I'm confused about this...

I could not download the DiDeMo dataset from https://drive.google.com/drive/u/1/folders/1_oyJ5rQiZboipbMl6tkhY8v0s9zDkvJc ，could you help me to find a easier way to download it? Thanks!!

Hi ,

I saw "We pre-train ALPRO for 100k iterations, roughly equivalent to 10 epochs" in your paper. So there will be ten checkpoints, which one is the zero-shot checkpoint for testing MSRVTT, and how to choose?

Looking forward to your reply, thanks!

Sorry, I didn't find out the description of the dataset that was used to pre-train the prompter in the paper.

As you mentioned in pretraining data preparation "we downsample webvid2m videos to 10% of the original FPS to speed-up video loading", can you elaborate on how to achieve this?

Thank you

Hi,

Congratulations on the amazing work. Will there be any difference in performance if I use just a single GPU and what are the changes to be made in eg: msvd_qa.json?

Thank you.

when you use vit to extract the video feature, may I know what is the input video feature dimension?

Hi, I have tried to use the provided pretrained checkpoint of ALPRO to finetune on MSRVTT.

Before start training, it would test on the validation set, and it gives me the results:

{'text2video': {'r1': 17.9, 'r5': 40.0, 'r10': 50.9, 'medianR': 10.0, 'meanR': 57.259}, 'video2text': {'r1': 16.0, 'r5': 34.6, 'r10': 46.0, 'medianR': 13.0, 'meanR': 61.448}

I think it does not match the results presented in the paper which achieves R@1 with 24.1, R@5 with 44.7. I am wondering why it would happen?

Thanks for the sharing code! I was trying to set up the environment but met with some problem especially on installing apex; I wonder if it is possible to provide a .yaml file that can used to create the environment using only Conda? or a docker container for setting up the environment? Thanks!

Hi,

In the inference we always load the best model. However, after fine-tuning there is no checkpoint named $OUTPUT_DIR/ckpt/model_step_best.pt. Can you point to the line in the code where the best checkpoint is saved?

Thank you.

Hello,

Thank you very much for open-sourcing this code. I would like to try running it locally, but I only have one GPU, and I've encountered several issues while trying to install the Horovod library, making it impossible for me to proceed. Could you please let me know if there is a version for single GPU testing that doesn't require Horovod installation? Alternatively, could you provide instructions for setting up the environment directly through Conda?

Looking forward to your reply, Very thanks!

Hi, as stated in the issue, the ALPRO does use weight decay. But I did not find the process that passing the parameter "weight_decay" during the optimizer initialization.

optimizer = OptimCls(model.parameters(), lr=opts.learning_rate, betas=opts.betas)

Hi, I am trying to pretrain the model using CC3M dataset,

After the first iteration (batch), the program would stuck, and give the following warning.
horovod/common/stall_inspector.cc:105] One or more tensors were submitted to be reduced, gathered or broadcasted by subset of ranks and are waiting for remainder of ranks for more than 60 seconds. This may indicate that different ranks are trying to submit different tensors or that only subset of ranks is submitting tensors, which will cause deadlock.

Is there any way that I can avoid this?

Hi,
Thanks for sharing the code!

I saw "use 7k videos for training and report results on the 1k test split" in your paper. When I downloaded the MSR-VTT dataset, there are only 7K train sets and 3K test sets, but no val dataset. Could you share the code for dividing the dataset to avoid discrepancies in results?

Looking forward to your reply.