search

ViTAE-Transformer / DeepSolo

The official repo for [CVPR'23] "DeepSolo: Let Transformer Decoder with Explicit Points Solo for Text Spotting" & [ArXiv'23] "DeepSolo++: Let Transformer Decoder with Explicit Points Solo for Multilingual Text Spotting"
Other
250 stars 34 forks source link

how to reduce gpu memory usage? #29

Closed megyux closed 1 year ago

megyux commented 1 year ago

Hello, I'm trying to train the model on a custom toy dataset using google's colab, the free version of colab "usually" offers t4 gpu (16gb memory), using a batch size of 8 images produces the cuda out of memory error, even when changed it to only 2 images per patch I still run into that error after ~200 iters, what can I do to reduce the memory usage ? 2 images per patch will make the loss fluctuate too much (I guess) Edit: even when using resnet50

ymy-k commented 1 year ago

Under the default setting of training image size, batching 2 images in one gpu requires about 17GB. You can try to change the image size during training and testing. The performance will be affected by using batchsize 2.

megyux commented 1 year ago

thanks, reducing image size works