Ross Wightman

@rom1504 close to 36 hours now with no errors / failures using the storage API so all failures do appear to have been caused by piping gsutil via popen. Unfortunately,...

@rom1504 last update on this until I have some time to try some performance improvements... the instances that I have the Python client running on have been running since last...

@tmbdev yeah this behaviour is puzzling The googleapis (http) could work, but I can't do unauthenticated so I'd need to hack in a token to the curl args (or use...

@rmcavoy yes, there are no convs in the feature map level rescalling unless there is a channel dim mismatch to bridge, so only for P6, not P7. For the level...

I do agree that the paper was lacking some detail to fully reproduce the model without making some wrong guesses. I suspect even the official impl may have changed a...

FYI, I decided to get a version of EfficientDet running now that an official reference is available. I saw this issue searching through the pre-existing PyTorch impls out there for...

ResNet50 in Keras is acc1: 75.00%, acc5: 92.16% w/ the interpolation set to bicubic instead of bilinear I'd create a per framework per model scale factor and interpolation method to...

@cgnorthcutt I understand what you mean wrt your goal. Ultimately up to you. You can provide the code that does the eval as it should be, and provide a reference...

There are a lot of factors at play for a given result. Pytorch version, CUDA, PIL, etc. Even changing the image scaling between bicubic and bilinear can have a notable...

ResNet50 on PyTorch 1.0.1.post2 and CUDA 10 w/ bilinear instead of bicubic, Prec@1 76.138, Prec@5 92.864 ... matches your numbers @ankmathur96