Cadene / murel.bootstrap.pytorch

MUREL (CVPR 2019), a multimodal relational reasoning module for VQA
https://arxiv.org/abs/1902.09487
BSD 3-Clause "New" or "Revised" License
194 stars 24 forks source link

How long does it take to train the models? #3

Closed gaopeng-eugene closed 5 years ago

Cadene commented 5 years ago

For the complete murel model with the pairwise module and three steps, it can take 1 day on two gpus.

gaopeng-eugene commented 5 years ago

Hi

Will you be able to share the training log for TDIUC and VQACP2?

gaopeng-eugene commented 5 years ago

Hi, I can train TDIUC and VQA2 with your code but not for VQACP2. Could you share VQACP2 log first, if possible. I will share the TDIUC log file after finished training.

Cadene commented 5 years ago

@gaopeng-eugene Did you solve your issue?

What was the issue with VQACP2?

Zzmonica commented 5 years ago

For the complete murel model with the pairwise module and three steps, it can take 1 day on two gpus.

I use the same configuration, it takes almost 4.5 hours per epoch, and you know here is 25 epoches in total for training procedure,and I use 4 GeForce GTX 1080 GPU for training. Is that normal?

Einstone-rose commented 3 years ago

For the complete murel model with the pairwise module and three steps, it can take 1 day on two gpus.

I use the same configuration, it takes almost 4.5 hours per epoch, and you know here is 25 epoches in total for training procedure,and I use 4 GeForce GTX 1080 GPU for training. Is that normal?

Hi, i have some questions. When training based on its default parameters ( dropout_input: 0.1, dropout_pre_lin: 0.0, dropout_output: 0.0), Isn't that overfitting? when I training the model , the val score is only 0.24.