chenxi116
Error with main.py
main.py --train --exp lr7e-3 --epochs 50 --base_lr 0.007
raise ValueError('Expected more than 1 value per channel when training, got input size {}'.format(size)) ValueError: Expected more than 1 value per channel when training, got input size [1, 256, 1, 1]
This seems to happen when a particular data batch has size 1. Then BN breaks down.
Which dataset are you training on? And what is your batch size?
This seems to happen when a particular data batch has size 1. Then BN breaks down.
Which dataset are you training on? And what is your batch size?
dataset: pascal voc 2012 batchsize: 16
Interesting. I was using this code very recently, but didn't encounter this problem.
Does your problem occur for the first batch? Or the last batch of the epoch?
Also can you confirm len(dataset) == 10582?
You did not answer my last question, which is about dataset length.
You need to measure the last batch's batch size, which may not equal to 16. My guess is that it is somehow 1, which creates the error.
I just ran the code, the len(dataset) == 10582
Interesting! When I changed the CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 to CUDA_VISIBLE_DEVICES=4,5,6,7. the code ran successfully....
interesting. after running the main.py with 4 GPUs the mean IOU is only 75.75% not 77.14%...
I recommend training with one GPU. This is because the "DataParallel" in pytorch does not synchronize BN statistics. So by using 4 GPUs with batch size 16, you are effectively computing BN statistics using batch size 16/4 = 4, and we all know the larger the batch size, the better BN gets.
I have used this code recently, and if you use one GPU, this number should be at least 76.50%.
interesting. after running the main.py with 4 GPUs the mean IOU is only 75.75% not 77.14%...
Can i have your qq number , I really want your help, thank u very much
