What GPUs were used for training on ICDAR2017 HTR?

IntuitionMachines / OrigamiNet

Public implementation of our CVPR Paper "OrigamiNet: Weakly-Supervised, Segmentation-Free, One-Step, Full Page TextRecognition by learning to unfold"

143 stars 39 forks source link

Closed IuliiaSaveleva closed 4 years ago

IuliiaSaveleva commented 4 years ago

How many GPUs were used during training on ICDAR2017 HTR and what was their memory capacity or other important stats (if it is necessary to mention)?

ASDen commented 4 years ago

For the largest model and best results we used 8x 16GB V100s, but it is important to mention:

You can already do very well with smaller models (see the readme)
You can also do very well using a big model but small input resolution
The code supports gradient checkpointing and gradient accumulation to simplify using smaller number of gpus with less amount of available memory