PyTorch and CUDA
If the Nvidia CUDA driver is already installed on the system:
What is my CUDA version?
nvidia-smi
returns version 12.2.
PyTorch CUDA 12.2 was not available at the time of writing, but the PyTorch CUDA 12.1 version works fine with
CUDA v12.2:
https://stackoverflow.com/questions/76678846/pytorch-version-for-cuda-12-2
Accordingly, set the option to 12.1 in the following command:
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
To test:
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Available Device : {device}")
Prediction
From the example at
https://www.kaggle.com/code/tirdadlajmouraki/transfer-learning-with-pytorch
applied to the example at
https://learn-pytorch.oneoffcoder.com/model-persistence.html
val_loader = DataLoader(val_image_folder, batch_size=4, shuffle=False)
will provide 4 inputs to the predictor at a time (batch_size=4).
outputs = model(inputs)
generates the predictions.
Since there are three shapes, one of the outputs in the loop will be
[-1.1430, 2.0940, -1.7959],
[ 0.3532, -0.4959, -0.7845],
[ 1.2950, -1.5365, -0.6781],
[-0.3655, 1.1077, -1.8694]
meaning that the probabilities for the first input are such the the second shape is the most likely.
_, predicted = torch.max(outputs.data, 1)
sets predicted to the most likely shape in the first batch of 4:
[1, 0, 0, 1]
