search

henghuiding / ReLA

[CVPR2023 Highlight] GRES: Generalized Referring Expression Segmentation
https://henghuiding.github.io/GRES/
MIT License
656 stars 19 forks source link

About the training time #6

Closed Nastu-Ho closed 1 year ago

Nastu-Ho commented 1 year ago

According to your supplementary file, the model is trained for 150,000 iterations with a batch size of 24 on four 32G V100 GPUs. How long does it take to complete a training?

changliu19 commented 1 year ago

Hi,

Please note that we have further fine-tuned some hyperparameters for better performance. The current checkpoint was trained using the default configuration files mentioned in the README. Thank you for your reminder, and we have added this information to the documentation. For training with 8 GPUs and a batch size of 48, it typically takes about 3-4 days for 30k iterations.

henghuiding commented 1 year ago

We provide results based on resource-friendly ResNet-50 and Swin-Tiny backbones. Please kindly also consider reporting results on these small backbones as an alternative choice.

PanXiongAdam commented 1 year ago

Please note that we have further fine-tuned some hyperparameters for better performance. The current checkpoint was trained using the default configuration files mentioned in the README. Thank you for your reminder, and we have added this information to the documentation. For training with 8 GPUs and a batch size of 48, it typically takes about 3-4 days for 30k iterations.

Hello! you said that using the current performance configuration would result in better performance. Can you add the current performance indicators to readme.md? What I see is still the indicator in the article about swin-base. And is it convenient to provide the original configuration file? Can be used as a comparative experiment for fairness.

qiulesun commented 1 year ago

@henghuiding @ntuLC

How many epochs correspond to 30k iterations?