Strange network output when dataloader is using ImageFolder

vision
1

Hello, I have strange behavior on the neural network output when I load custom images and test them through the neural network. I am supposed to get one vector with number of predictions which is the same as number of classes, but I got multiple vectors outputs for one image. I have neural network that has 10 classes and is trained on CIFAR10 dataset.
This is the output of the last layer of neural network for two images.

C:\ProgramData\Anaconda3\envs\env3\python.exe C:/Users/Dejana/Box/Research/Projects/OOD-Activations/Test/test2.py
torch.Size([1, 3, 64, 64])
tensor([[ 2.0437, -1.3588,  1.1609,  0.8730,  0.2099, -2.9088, -0.6814,  1.2066,
          1.0654, -1.7905],
        [ 4.6609, -0.2636,  0.1245,  1.5712,  0.3961, -0.7396, -3.2168,  2.9341,
         -3.2016, -2.0086],
        [-0.5989, -2.4766,  0.7316,  3.6030, -0.0383,  2.3087,  3.0540, -3.0784,
         -0.5181, -3.2308],
        [-0.5472, -8.9718, -1.1429,  5.1866,  3.1074,  4.4514,  1.0680,  3.5349,
         -2.6262, -5.3872]], grad_fn=<AddmmBackward>)
tensor([[ 1.4245, -4.3988,  2.3064,  4.1217,  1.0324,  0.1579, -1.8399,  0.2804,
          1.5780, -4.6716],
        [ 0.7175, -2.7187,  4.3732,  1.6181, -0.5130,  0.1256, -0.5808,  1.7235,
         -3.3944, -1.6412],
        [-1.0284,  1.4488,  1.8022,  3.1178, -2.2113,  0.8049,  2.2102,  4.5608,
         -5.7092, -5.1130],
        [ 1.2146, -8.5995,  0.8800, 10.9451,  0.4852,  0.7585,  0.8278,  3.1131,
         -7.2019, -2.5458]], grad_fn=<AddmmBackward>)

Please see code. I used two different ways to obtain data through the dataloader.

import matplotlib.pyplot as plt
import torch
import torchvision
import torchvision.transforms as transforms
import Utils.densenetBlock as db

transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

#testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=transform)
#testloader = torch.utils.data.DataLoader(testset, batch_size=1, shuffle=False, num_workers=0)

testset = torchvision.datasets.ImageFolder(root=r"..\Data\tiny-imagenet-10000\t1", transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=1, shuffle=False, num_workers=0)

net = db.DenseNet3(100, 10)
net.load_state_dict(torch.load('..\Data\densenet10.pth', map_location=torch.device('cpu')), strict=False)
net.eval()

#first_way
i=0
for data in testloader:
    inputs, labels = data[0], data[1]
    output = net(inputs)
    print(output)
    i = i + 1

###second way
dataiter = iter(testloader)
lst = [x for i, x in enumerate(dataiter) if i % 9 < 500]
lstiter = iter(lst)
j = 0
while j < 2:
    images1, labels1 = next(lstiter)
    outputs = net(images1)
    print(outputs)
    j = j + 1

The datasets I am trying to test against my neural network are: Tiny-ImageNet (structure for testing purposes shown on image) and ImageNet.
Screenshot_37
When I am loading CIFAR10 images, the output seems legit:

torch.Size([1, 3, 32, 32])
tensor([[ -3.7279,  -7.4436,  -5.9792,  40.7376,  -8.7128,   9.2873,   4.3494,
         -11.3954,  -3.0350, -14.5300]], grad_fn=<AddmmBackward>)

Thank you.

2

I have resolved. It was regarding the size of the images: The network was trained on torch.Size([1, 3, 32, 32]) and validation imageswere size of torch.Size([1, 3, 64, 64]). This means that neural network treated validation images as batches of 4 images and that’s reason for the output that dimensions.

1 Like