Calculate Model Accuracy for one of the Kaggle Notebook

Question

How to calculate accuracy for PyTorch model used inhttps://www.kaggle.com/code/andradaolteanu/whales-dolphins-effnet-train-rapids-clusters

I tried using some methods like -

1. Tried this → correct += (out == targets).float().sum() accuracy = 100 * correct / len(test) This is giving weird errors ,i.e,

   RuntimeError Traceback (most recent call last) /tmp/ipykernel_33/896230284.py in ----> 1 train_pipeline(train) /tmp/ipykernel_33/2939766947.py in train_pipeline(train) 62 # print(targets) 63 ---> 64 correct += (out == targets).float().sum() 65 # num_corrects = num_corrects + torch.sum(out.argmax(1) == targets) 66 RuntimeError: The size of tensor a (15587) must match the size of tensor b (32) at non-singleton dimension 1

   2. Tried this → num_corrects = num_corrects + torch.sum(out.argmax(1) == targets) print(“Accuracy = {}”.format(train_accuracy)) This is giving 0 as accuracy and printing out and targets giving me this

   out tensor([[-13.0885, -15.4269, -15.7011, ..., -18.2123, -21.8597, -19.1524], [-13.3293, -12.2672, -14.0315, ..., -17.7681, -22.0472, -18.1889], [-14.6508, -13.7065, -16.5062, ..., -19.2380, -21.8163, -19.5815], ..., [-15.3698, -16.0767, -14.8986, ..., -18.8057, -22.7389, -19.4303], [-13.5040, -14.4153, -15.7141, ..., -19.5422, -22.9803, -19.5026], [-13.8497, -15.1122, -14.3879, ..., -19.5625, -22.6934, -18.4038]], device='cuda:0', grad_fn=) targets tensor([ 2842, 1353, 118, 108, 425, 168, 2123, 9758, 6036, 748, 443, 1, 102, 12102, 2115, 450, 96, 5147, 998, 1227, 220, 177, 8876, 32, 79, 9920, 65, 43, 1468, 84, 8002, 13], device='cuda:0') out tensor([[-14.9549, -14.1026, -15.7504, ..., -19.3176, -22.6721, -18.4659], [-11.7263, -12.5097, -15.5440, ..., -19.5593, -22.9155, -20.5102], [ -4.3821, -9.7733, -3.9627, ..., -11.9340, -11.7187, -12.0451], [ -6.1771, -7.6977, -5.3471, ..., -11.8034, -11.9683, -13.8230], [-15.3451, -16.2378, -16.4004, ..., -20.1262, -22.6535, -20.3700], [-14.6429, -15.8015, -15.3850, ..., -19.8169, -22.2832, -18.9964]], device='cuda:0', grad_fn=) targets tensor([ 8923, 2130, 4607, 12931, 341, 47], device='cuda:0')

I am adding this accuracy thing in train_pipeline(train)

Please could anyone help me with it,

Answer 1

Since you haven't provided details about the exact shapes and datatypes of out and targets tensors, I can only provide the approach, you have to figure out the exact code.

If you are using softmax at the end of your model, then out should be a tensor of size (n_classes, 1) containing the probabilities of the datapoint belonging to each class. You have to convert that vector of probabilities to the corresponding integer class label, and store it in (say) out_labels.

Also ensure that targets is an integer tensor, and contains the correct integer class labels.

After above steps, you can compute accuracy like this -

accuracy = (out_labels == targets).sum() / len(targets)

(Make sure both out_labels and targets are integer tensors of the same shapes to avoid running into any errors).