Hard Coded Transformations

Hi Gabriel! We shouldn't be applying hard coded transformations as such while testing, rather allow the user to define one and raise assertion error if they didn't :)

https://github.com/GabrielePicco/deep-learning-flower-identifier/blob/master/test_model_pytorch_facebook_challenge.py#L33-L38

Also somehow the code for calculating Accuracy wasn't working for me [PIL Register Decode Error even after applying the fix..], don't know why, so we can replace that with this (self curated version plus someone shared this and i modified to meet my necessities)

data_dir = 'google_test_data'

#sample transformations
transformation = transforms.Compose([
    transforms.Resize((299,299)),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

image_dataset = datasets.ImageFolder(data_dir, transform= transformation)
dataloader = torch.utils.data.DataLoader(image_dataset, batch_size= 128, shuffle=True)

#######validation##########
# track valid loss
valid_loss = 0.0
class_correct = list(0. for i in range(102))
class_total   = list(0. for i in range(102))

model.eval().to(device)
# iterate over valid data
for data, target in dataloader:
    # move tensors to GPU if CUDA is available
    data, target = data.to(device), target.to(device)
    # forward pass: compute predicted outputs by passing inputs to the model
    output = model(data)
    _, pred = torch.max(output, 1)
    # compare predictions to true label
    correct_tensor = pred.eq(target.data.view_as(pred))
    correct = np.squeeze(correct_tensor.numpy()) if not device else np.squeeze(correct_tensor.cpu().numpy())
    # calculate valid accuracy for each object class
    for i in range(len(data)):
        label = target.data[i]
        class_correct[label] += correct[i].item()
        class_total[label] += 1

for i in range(102):
    if class_total[i] > 0:
        print('Validation Accuracy of %5s %s %s: %2d%% (%2d/%2d)' % (
            model_pred_to_folder[i],'->',final[i], 100 * class_correct[i] / class_total[i],
            np.sum(class_correct[i]), np.sum(class_total[i])))

print('\nValidation Accuracy (Overall): %2d%% (%2d/%2d)' % (
    100. * np.sum(class_correct) / np.sum(class_total),
    np.sum(class_correct), np.sum(class_total)))

Sample Output

Validation Accuracy of     1 -> pink primrose: 60% ( 6/10)
Validation Accuracy of    10 -> globe thistle: 90% ( 9/10)
Validation Accuracy of   100 -> blanket flower: 80% ( 8/10)
Validation Accuracy of   101 -> trumpet creeper: 80% ( 8/10)
Validation Accuracy of   102 -> blackberry lily: 90% ( 9/10)
Validation Accuracy of    11 -> snapdragon: 80% ( 8/10)

(PS We can replace the Word Validation with Test) Again Thanks a lot for the dataset !!

GabrielePicco / deep-learning-flower-identifier

Hard Coded Transformations #1