tjdevworks / teasel Goto Github PK

Hi, Thanks for sharing your work! I couldn't find your model after clicking the link on your readme. When will you make it available?

Nice work and nice repository ! But I still have some doubts about the repository~

1. Can I run it on a single GPU? Although, I run your work(python fine_tune_mosi.py --config config_mosi.yaml) in two GPU, it'll always return CUDA out of memory.My GPU is Tesla V100, which shows RuntimeError: CUDA out of memory. Tried to allocate 376.00 MiB (GPU 0; 15.78 GiB total capacity; 14.18 GiB already allocated; 61.50 MiB free. Of course, I also tried to set the batchsize to 1 or 2.
   Unlike the train module, I don't quite understand why the finetune module needs such a large memory.
   Can you tell me how I should go about changing the batchsize or other modifications to support training on a single gpu?
   For example, how many epochs is recommended when trained on a single gpu, in order to reach pretraining convergence?
   I see in code, you always make opt as adamw, did you select apex with fp16 for training or inference? And are there other tricks in the training?

2. Can I still run the code after removing the code that calls the wandb interface?
   I'm so sorry!!! I'm a novice in deep learning and don't understand the built-in mechanism of wandb. Therefore, wandb cannot be used skillfully. If I try to submit tasks to a multi GPU cluster, I don't know how to enter the API_ Key, which will also lead to wandb errors. UsageError: api_ key not configured (no-tty). call wandb. login(key=[your_api_key])， so I only choose wandb: (3) Don't visualize my results.

I'm fairly new to this, and I appreciate the help. Thank you

The raw dataset CMU-MOSI.zip of fine-tune task is unavailable.
It seems that the source file on the server has been moved.

Not Found
The requested URL /raw_datasets/ was not found on this server.
Apache/2.4.7 (Ubuntu) Server at immortal.multicomp.cs.cmu.edu Port 80

Could you please re-upload the raw dataset file? Thanks!

Nice work for you! but I have some question about the dataset. The datasets you use is raw data for MOSI? and i can't reproduce because i don't know which data to use. or can you share the MOSI datasets you use? Thanks!

While fine-tuning the network (on 4 V100 GPUs) for text to SC, I always face the issue of CUDA OUT OF MEMORY.
If the batch size is reduced to 2 (from 16), this issue is faced again.

I am currently using -

pytorch 1.10.2
transformers 3.4.0
wandb==0.12.16
huggingface-hub==0.6.0
batch_size: 2
num_workers: 8

E.g.:
RuntimeError: CUDA out of memory. Tried to allocate 194.00 MiB (GPU 5; 23.59 GiB total capacity; 22.18 GiB already allocated;111.19 MiB free; 23.37 GiB reserved in total by PyTorch).
Did you face any of these issues during fine-tuning?