search

conservationtechlab / animl-r

Animl comprises a variety of machine learning tools for analyzing ecological data. The package includes a set of functions to classify subjects within camera trap field data and can handle both images and videos.
MIT License
19 stars 5 forks source link

MD resize #13

Open tkswanson opened 2 years ago

tkswanson commented 2 years ago

If MD takes in full image, give the option to resize to help smaller GPUS, make sure we are using the same size for classification training

matobler commented 2 years ago

MD seems to resize the image as a preprocessing step. This preprocessing step is part of the model but we don't know what the output size is. If we know the output size used by MD internally, we can resize the image before sending it to MD without any loss in performance. Need to find that size and compare results from full and reduced-size images. Test on GPU with low memory (e.g. 4GB). The size we use for classification will be different and is given by the model we use..

matobler commented 2 years ago

I exported the model graph and it looks like MD does NOT resize the image internally, just converts it to float. Wonder where I thought I read that. Might still be worthwhile evaluating a resize option for systems with little GPU memory and compare performance but lower priority.

image