This directory contains examples scripts.
The examples in this directory assume that the DUNEdn package with the
example extensions is installed in the current environment.
This can be done running pip install -e .[example] from the package root
directory.
The dataset sample and checkpoints must be downloaded to be able to run the examples. Refere to the main README for more information.
This folder contains a comparison between PyTorch and Onnx model inference.
The examples can be executed as scripts with:
python onnx_accuracy_example.pyAnd as a jupyter notebook.
The performance examples can be executed in the same manner through the
onnx_performance_example files in this folder.
GPU memory is precious and OOM errors are extremely annoying.
This section gives an idea of the memory consumption for each model, so that the user can decide the best strategy based on the available hardware.
The CNN Network acts on image crops of (32,32) pixels resolution.
The GPU memory used for a forward pass depends linearly on the batch size:
CNN usage ~ batch_size * 46 MB
The GCNN Network acts on image crops of (32,32) pixels resolution.
The GPU memory used for a forward pass depends linearly on the batch size:
GCNN usage ~ batch_size * 84 MB
The USCG Network typically acts on image windows of different pixels resolution.
Default values for induction windows are (800,2000), while collection ones are
(960,2000).
The memory consuptions are listed in the following table
| Plane | Resolution | GPU Memory |
|---|---|---|
| Induction | 2000x800 | 3.67 GB |
| Collection | 2000x960 | 4.40 GB |
The two results are consistent.