AlexNet With ImageNet Testing

1

This is probably a human error but i would like to note down accuracy of AlexNet with already trained networks and then replace conv layers with my custom layers and note down results again.

I can find AlexNet and pre_trained weights here [AlexNet]

The Datasets are downloaded from here [AT]

Main Folder Name : imagenet2012
Sub Folder 1: ILSVRC2012_img_train Contains different folder (n01443537,n01484850… 15 of them with images inside)
Sub Folder 2: ILSVRC2012_img_val (ILSVRC2012_val_00000001.JPEG etc…Contains all images)

import torch.nn as nn
from torch.hub import load_state_dict_from_url
import torchvision
from torchvision import transforms
from torch.utils import data
from torchvision.datasets import ImageFolder
import torch
__all__ = ['AlexNet', 'alexnet']


data_transforms = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])


data_dir = "/home/Sami/Documents/imagenet2012/"

model_urls = {
    'alexnet': 'https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth',
}

train_dataset = torchvision.datasets.ImageFolder(root=val_dir, transform=data_transforms, target_transform=None, is_valid_file=None)
test_dataset  = torchvision.datasets.ImageFolder(root=val_dir, transform=data_transforms, is_valid_file=None)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset,  batch_size=100, shuffle=True)
test_ds  = torch.utils.data.Subset(test_loader.dataset, range(0,50))
test_loader_1 = torch.utils.data.DataLoader(dataset=test_ds,  batch_size=100, shuffle=True)




class AlexNet(nn.Module):

    def __init__(self, num_classes=1000):
        super(AlexNet, self).__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),
            nn.Conv2d(64, 192, kernel_size=5, padding=2),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),
            nn.Conv2d(192, 384, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(384, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.Conv2d(256, 256, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(kernel_size=3, stride=2),
        )
        self.avgpool = nn.AdaptiveAvgPool2d((6, 6))
        self.classifier = nn.Sequential(
            nn.Dropout(),
            nn.Linear(256 * 6 * 6, 4096),
            nn.ReLU(inplace=True),
            nn.Dropout(),
            nn.Linear(4096, 4096),
            nn.ReLU(inplace=True),
            nn.Linear(4096, num_classes),
        )

    def forward(self, x):
        x = self.features(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.classifier(x)
        return x

model = AlexNet()
state_dict = load_state_dict_from_url(model_urls['alexnet'],  progress=True)
model.load_state_dict(state_dict)
print(model)



with torch.no_grad():
    correct = 0
    total = 0
    
    for images,labels in test_loader_1:
        out = model(images)
        _,predicted = torch.max(out.data,1)
        total += labels.size(0)
        correct += (predicted==labels).sum().item()
    
    
print('Accuracy of the network on the 50 test images: {} %'.format(100 * correct / total))

But the accuracy stays 0.0 something like that.

2

Could you please check, if correct and total are stores as float values or as int?
In the latter case, your accuracy might get rounded down to zero.

3

You are right it was int, after changing it to float i get 2.0 % on 50 images is this alright ?

4

2% accuracy on 50 images would be a single correct sample.
It depends what you define as “alright”, but I would say it’s a bit low.

5

I am not sure but i was thinking may be around 56-57% would be alright out of 1000 images Comparisons or Accuracy. ILSVRC2012_img_train/ folder has only 15 sub-folders with in.

May be my data loaders have some problems in them ? Directory structure is as follows

imagenet2012/
     ILSVRC2012_img_train/
                         n01443537/
                                  n01443537_2.JPEG
                                  n01443537_16.JPEG
                                  .
                                  .
                         n01484850/
                                  n01484850_17.JPEG
                                  n01484850_76.JPEG
                                  .
                                  .
                                  .
                                  .
     ILSVRC2012_img_val 
         ILSVRC2012_val_00000001.JPEG
         ILSVRC2012_val_00000002.JPEG
                                  .
                                  .
                                  .
6

Based on your code it looks like you are using ImageFolder on your validation directory, which seem to contain only the images without any subfolders.
ImageFolder creates the targets based on subfolders, so your current Datasets might contains only a single class label. Could you check that?

7

Yes you are right. That seems to be the problem.
In reply to the commands

print(test_dataset.classes)
print(test_dataset.class_to_idx)

Yields

[‘ILSVRC2012_img_train’, ‘ILSVRC2012_img_val’]

{‘ILSVRC2012_img_train’: 0, ‘ILSVRC2012_img_val’: 1}

Do i need to have some official dataset ? As i cannot seem to download it when i use this one

torchvision.datasets.ImageNet(root, split='train', download=True, **kwargs)
8

I think the download URL used in datasets.ImageNet was disabled (as described here), so you would need to register on their site and download the data separately.

9

Thank you very much. I have already applied to register on their site. I will update this post as soon as i move ahead. In the meantime i will try this dataset [kaggle] and let you know. Thanks a ton really

Update:
Above kaggle dataset works fine i have used this script sh to put images in folders for “val” so that pytorch’s dataloader can easily identify labels. For test data folder may be a similar script is needed to put them in folders.

Obtained Accuracy on Val Set 70.1% (1000 images)

10

Hi,
Can you please share more details on the structure of the “test data folder” and test-dataloader for ImageNet dataset? Thanks.

11

Hi, are you asking about this one .

_all__ = ['AlexNet', 'alexnet']


data_transforms = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])


val_dirr = "~/Alexnet/val"

model_urls = {
    'alexnet': 'https://download.pytorch.org/models/alexnet-owt-4df8aa71.pth',
}

#train_dataset = torchvision.datasets.ImageFolder(root=..., transform=data_transforms, target_transform=None, is_valid_file=None)
test_dataset  = torchvision.datasets.ImageFolder(root=val_dirr, transform=data_transforms, is_valid_file=None)

test_loader = torch.utils.data.DataLoader(dataset=test_dataset,  batch_size=100, shuffle=True)

#If making subsets
#test_ds  = torch.utils.data.Subset(test_loader.dataset, range(0,2))
#test_loader_1 = torch.utils.data.DataLoader(dataset=test_ds,  batch_size=1, shuffle=True)