sandeep-krishnamurthy commented on a change in pull request #7795: Fix
build-server test running out of GPU memory (CUDNN allocation failure)
URL: https://github.com/apache/incubator-mxnet/pull/7795#discussion_r137663736
##########
File path: tools/caffe_converter/test_converter.py
##########
@@ -89,18 +89,20 @@ def main():
args = parser.parse_args()
if args.cpu:
gpus = [-1]
- batch_size = 32
+ default_batch_size = 32
else:
gpus = mx.test_utils.list_gpus()
assert gpus, 'At least one GPU is needed to run test_converter in GPU
mode'
- batch_size = 32 * len(gpus)
+ default_batch_size = 32 * len(gpus)
models = ['bvlc_googlenet', 'vgg-16', 'resnet-50']
val = download_data()
for m in models:
test_model_weights_and_outputs(m, args.image_url, gpus[0])
- test_imagenet_model_performance(m, val, gpus, batch_size)
+ # Build/testing machines tend to be short on GPU memory
+ this_batch_size = default_batch_size / 4 if m == 'vgg-16' else
default_batch_size
+ test_imagenet_model_performance(m, val, gpus, this_batch_size)
Review comment:
1. Will reducing batch size has impact on other parameters tested in this
test case? I am assuming not, seeing the changes here.
2. Will it pass on bigger machine? Just to confirm we are not missing out on
an underlying issues?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
