My pytorch model outputs a value for one of my classes, but always gives 0 for the other

1

My model is meant to classify 446x2048 images as either defects or non-defects, but it gives me a strange output:

Defect: 0.0
No_Defect: 144.98797607421875
Defect: 0.0
No_Defect: 136.9902801513672
Defect: 0.0
No_Defect: 136.9709014892578
Defect: 0.0
No_Defect: 127.5387191772461
Defect: 0.0
No_Defect: 147.4520721435547
Defect: 0.0
No_Defect: 125.58473205566406
Defect: 0.0
No_Defect: 143.82965087890625

I’m expecting a probability for each label, but I’m getting a number too large for no_defect, and only 0 for defect
Here is my code for training the model:

import os
import torch
import torch.nn as nn
import torchvision
from torchvision import datasets, transforms
from torchvision.transforms import transforms
from torch.utils.data import DataLoader
from torch.optim import Adam
import matplotlib.pyplot as plt
import numpy as np
import torch.nn.functional as F
import onnx

os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:64"
torch.cuda.empty_cache()

# Loading and normalizing the data.
# Define transformations for the training and test sets
transformations = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

batch_size = 10
number_of_labels = 2

data_dir = "D:/Cropped Copper Plates/Edge Strips"

# Load training data
train_dir = os.path.join(data_dir, "Training")
train_set = datasets.ImageFolder(train_dir, transform=transformations)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=0)

# Load testing data
test_dir = os.path.join(data_dir, "Testing")
test_set = datasets.ImageFolder(test_dir, transform=transformations)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=0)

print("The number of images in a training set is: ", len(train_loader)*batch_size)
print("The number of images in a test set is: ", len(test_loader)*batch_size)

print("The number of batches per epoch is: ", len(train_loader))
classes = ('Defect', 'No_Defect')


class Network(nn.Module):
    def __init__(self):
        super(Network, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3)
        self.bn1 = nn.BatchNorm2d(16)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=3)
        self.drop = nn.Dropout2d()
        self.bn2 = nn.BatchNorm2d(32)
        self.fc1 = nn.Linear(7187840, 2)

    def forward(self, x):

        x = F.relu(self.bn1(self.conv1(x)))
        x = self.pool(x)
        x = F.relu(self.bn2(self.drop(self.conv2(x))))
        x = self.drop(x)
        x = torch.flatten(x, 1)
        #x = x.view(x.size(0), -1)
        #print(x.shape)
        x = F.relu(self.fc1(x))
        x = F.dropout(x, training = self.training)
        #x = self.fc2(x)
        return x

# Instantiate a neural network model 
model = Network()

 
# Define the loss function with Classification Cross-Entropy loss and an optimizer with Adam optimizer
pos_weight = torch.full([10, 2], 8.5238)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
loss_fn = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
loss_fn = loss_fn.to(device)
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, weight_decay=0.001)

from torch.autograd import Variable

# Function to save the model
def saveModel():
    path = "./myFirstModel.pth"
    torch.save(model.state_dict(), path)

# Function to test the model with the test dataset and print the accuracy for the test images
def testAccuracy():
    
    model.eval()
    accuracy = 0.0
    total = 0.0
    
    with torch.no_grad():
        for data in test_loader:
            images, labels = data
            device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
            images, labels = images.to(device), labels.to(device)
            images = images.float()
            labels = labels.float()
            # run the model on the test set to predict labels
            outputs = model(images)
            # the label with the highest energy will be our prediction
            _, predicted = torch.max(outputs.data, 1)
            total += labels.size(0)
            accuracy += (predicted == labels).sum().item()
    
    # compute the accuracy over all test images
    accuracy = (100 * accuracy / total)
    return(accuracy)


# Training function. We simply have to loop over our data iterator and feed the inputs to the network and optimize.
def train(num_epochs):
    
    best_accuracy = 0.0

    # Define your execution device
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    print("The model will be running on", device, "device")
    # Convert model parameters and buffers to CPU or Cuda
    model.to(device)

    for epoch in range(num_epochs):  # loop over the dataset multiple times
        running_loss = 0.0
        running_acc = 0.0
        accum_iter = 4;
        for i, (images, labels) in enumerate(train_loader, 0):
            
            # get the inputs
            images = images.to(device)
            labels = labels.to(device)

            # predict classes using images from the training set
            outputs = model(images)
            labels = F.one_hot(labels, num_classes=2)
            labels = labels.float()
            outputs = outputs.float()
            # compute the loss based on model output and real labels
            loss = loss_fn(outputs, labels)
            loss = loss/accum_iter
            # backpropagate the loss
            loss.backward()
            if((i+1) % accum_iter == 0) or (i+1 == len(train_loader)):
                # adjust parameters based on the calculated gradients
                optimizer.step()
                # zero the parameter gradients
                optimizer.zero_grad()

            running_loss += loss.item()     # extract the loss value
            if i % 10 == 9:    
                # print every 100 (twice per epoch) 
                print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 100))
                # zero the loss
                running_loss = 0.0

        # Compute and print the average accuracy fo this epoch when tested over all test images
        accuracy = testAccuracy()
        print('For epoch', epoch+1,'the test accuracy over the whole test set is %d %%' % (accuracy))
        
        # we want to save the model if the accuracy is the best
        if accuracy > best_accuracy:
            saveModel()
            best_accuracy = accuracy


# Function to show the images
def imageshow(img):
    img = img / 2 + 0.5     # unnormalize
    npimg = img.numpy()
    plt.imshow(np.transpose(npimg, (1, 2, 0)))
    plt.show()


# Function to test the model with a batch of images and show the labels predictions
def testBatch():
    # get batch of images from the test DataLoader  
    images, labels = next(iter(test_loader))

    # show all images as one image grid
    #imageshow(torchvision.utils.make_grid(images))
   
    # Show the real labels on the screen 
    print('Real labels: ', ' '.join('%5s' % classes[labels[j]] for j in range(batch_size)))
  
    # Let's see what if the model identifiers the  labels of those example
    outputs = model(images)
    
    # We got the probability for every 10 labels. The highest (max) probability should be correct label
    _, predicted = torch.max(outputs, 1)
    
    # Let's show the predicted labels on the screen to compare with the real ones
    print('Predicted: ', ' '.join('%5s' % classes[predicted[j]] 
                              for j in range(batch_size)))

if __name__ == "__main__":
    
    # Let's build our model
    train(30)
    print('Finished Training')

    # Test which classes performed well
    testAccuracy()
    
    # Let's load the model we just created and test the accuracy per label
    model = Network()
    path = "myFirstModel.pth"
    model.load_state_dict(torch.load(path))

    # Test with batch of images
    testBatch()

    torch.onnx.export(model, torch.randn(batch_size, 3, 446, 2048), "saved_model.onnx", export_params=True, opset_version=10, do_constant_folding=True, input_names = ['input'], output_names = ['output'], dynamic_axes={'input' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}})
    onnx_model = onnx.load("saved_model.onnx")
    onnx.checker.check_model(onnx_model)

And here is my code to run the model:

import os
import numpy as np
from PIL import Image
import onnxruntime as ort

# Load the ONNX model
model_path = "C:/Users/username/source/repos/Copper_Plate_Pytorch/Copper_Plate_Pytorch/saved_model.onnx"
sess = ort.InferenceSession(model_path)

# Define the class labels
class_labels = ["Defect", "No_Defect"]

# Specify the folder containing the images
folder_path = "D:/Cropped Copper Plates/Edge Strips/Testing/Testing_D/"

# Loop over each image in the folder
for filename in os.listdir(folder_path):
    # Load the image
    img_path = os.path.join(folder_path, filename)
    img = Image.open(img_path).convert("RGB")

    # Resize the image to the input size expected by the model
    input_shape = (1, 3, 446, 2048)

    # Convert the image to a numpy array
    img = np.array(img).astype(np.float32)
    img /= 255.0
    img = np.transpose(img, (2, 1, 0))
    img = np.expand_dims(img, axis=0)
    # Run the inference
    inputs = {sess.get_inputs()[0].name: img}
    outputs = sess.run(None, inputs)

    # Print the probabilities for each class label
    probs = outputs[0][0]
    for i in range(len(class_labels)):
        print(f"{class_labels[i]}: {probs[i]}")
2

Your code shows a few incompatibilities:

  • nn.BCEWithLogitsLoss(pos_weight=pos_weight) indicates you are working on a binary or multi-label classification. In the latter case each sample belongs to zero, one, or multiple classes.
  • self.fc1 = nn.Linear(7187840, 2) indicates you are working indeed on a multi-label classification where each sample can belong to zero, one, or both classes.
  • _, predicted = torch.max(outputs.data, 1) indicates you are working on a multi-class classification where each sample belongs to one class only, which doesn’t match your criterion.
  • x = F.relu(self.fc1(x)): applying F.relu on the logits is usually not a good idea as it is often causing training issues since you are clipping the logits and are disallowing negative values (you should also check the last dropout layer btw.).

If you are working on a binary classification, you could use nn.BCEWithLogitsLoss and return a single output. To get the predicted class apply a threshold e.g. via:

preds = output > 0.0
# or calculate the probability and apply a prob. threshold
preds = torch.sigmoid(output) > 0.5

or you could return two outputs (as is currently the case) and switch to nn.CrossEntropyLoss. To get the predicted class you could then keep using torch.argmax(output, dim=1).

3

I tried to make my model reflect a true binary classification, but I get CUDA error: device-side assert triggered on loss.backward(). I’m not sure what I’m missing. Here is my updated code:

import os
import torch
import torch.nn as nn
import torchvision
from torchvision import datasets, transforms
from torchvision.transforms import transforms
from torch.utils.data import DataLoader
from torch.optim import Adam
import matplotlib.pyplot as plt
import numpy as np
import torch.nn.functional as F
import onnx

os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:64"
torch.cuda.empty_cache()

# Loading and normalizing the data.
# Define transformations for the training and test sets
transformations = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

batch_size = 10
number_of_labels = 1

data_dir = "D:/Cropped Copper Plates/Edge Strips"

# Load training data
train_dir = os.path.join(data_dir, "Training")
train_set = datasets.ImageFolder(train_dir, transform=transformations)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=0)

# Load testing data
test_dir = os.path.join(data_dir, "Testing")
test_set = datasets.ImageFolder(test_dir, transform=transformations)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=0)

print("The number of images in a training set is: ", len(train_loader)*batch_size)
print("The number of images in a test set is: ", len(test_loader)*batch_size)

print("The number of batches per epoch is: ", len(train_loader))
classes = ('Defect')


class Network(nn.Module):
    def __init__(self):
        super(Network, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3)
        self.bn1 = nn.BatchNorm2d(16)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=3)
        self.drop = nn.Dropout2d()
        self.bn2 = nn.BatchNorm2d(32)
        self.fc1 = nn.Linear(7187840, 1)

    def forward(self, x):

        x = F.relu(self.bn1(self.conv1(x)))
        x = self.pool(x)
        x = F.relu(self.bn2(self.drop(self.conv2(x))))
        x = self.drop(x)
        x = torch.flatten(x, 1)
        #x = x.view(x.size(0), -1)
        #print(x.shape)
        #x = F.relu(self.fc1(x))
        x = self.fc1(x)
        #x = F.dropout(x, training = self.training)
       
        return x

# Instantiate a neural network model 
model = Network()

 
# Define the loss function with Classification Cross-Entropy loss and an optimizer with Adam optimizer
pos_weight = torch.full([10, 1], 8.5238)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
loss_fn = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
loss_fn = loss_fn.to(device)
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, weight_decay=0.001)

from torch.autograd import Variable

# Function to save the model
def saveModel():
    path = "./myFirstModel.pth"
    torch.save(model.state_dict(), path)

# Function to test the model with the test dataset and print the accuracy for the test images
def testAccuracy():
    
    model.eval()
    accuracy = 0.0
    total = 0.0
    
    with torch.no_grad():
        for data in test_loader:
            images, labels = data
            device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
            images, labels = images.to(device), labels.to(device)
            images = images.float()
            labels = labels.float()
            # run the model on the test set to predict labels
            output = model(images)
            # the label with the highest energy will be our prediction
            #_, predicted = torch.max(output.data, 1)
            predicted = output > 0.0
            total += labels.size(0)
            accuracy += (predicted == labels).sum().item()
    
    # compute the accuracy over all test images
    accuracy = (100 * accuracy / total)
    return(accuracy)


# Training function. We simply have to loop over our data iterator and feed the inputs to the network and optimize.
def train(num_epochs):
    
    best_accuracy = 0.0

    # Define your execution device
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    print("The model will be running on", device, "device")
    # Convert model parameters and buffers to CPU or Cuda
    model.to(device)

    for epoch in range(num_epochs):  # loop over the dataset multiple times
        running_loss = 0.0
        running_acc = 0.0
        accum_iter = 4;
        for i, (images, labels) in enumerate(train_loader, 0):
            
            # get the inputs
            images = images.to(device)
            labels = labels.to(device)

            # predict classes using images from the training set
            outputs = model(images)
            labels = F.one_hot(labels, num_classes=1)
            labels = labels.float()
            outputs = outputs.float()
            # compute the loss based on model output and real labels
            print(outputs.shape)
            print(labels.shape)
            loss = loss_fn(outputs, labels)
            loss = loss/accum_iter
            # backpropagate the loss
            loss.backward()
            if((i+1) % accum_iter == 0) or (i+1 == len(train_loader)):
                # adjust parameters based on the calculated gradients
                optimizer.step()
                # zero the parameter gradients
                optimizer.zero_grad()

            running_loss += loss.item()     # extract the loss value
            if i % 10 == 9:    
                # print every 100 (twice per epoch) 
                print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 100))
                # zero the loss
                running_loss = 0.0

        # Compute and print the average accuracy fo this epoch when tested over all test images
        accuracy = testAccuracy()
        print('For epoch', epoch+1,'the test accuracy over the whole test set is %d %%' % (accuracy))
        
        # we want to save the model if the accuracy is the best
        if accuracy > best_accuracy:
            saveModel()
            best_accuracy = accuracy


# Function to show the images
def imageshow(img):
    img = img / 2 + 0.5     # unnormalize
    npimg = img.numpy()
    plt.imshow(np.transpose(npimg, (1, 2, 0)))
    plt.show()


# Function to test the model with a batch of images and show the labels predictions
def testBatch():
    # get batch of images from the test DataLoader  
    images, labels = next(iter(test_loader))

    # show all images as one image grid
    #imageshow(torchvision.utils.make_grid(images))
   
    # Show the real labels on the screen 
    print('Real labels: ', ' '.join('%5s' % classes[labels[j]] for j in range(batch_size)))
  
    # Let's see what if the model identifiers the  labels of those example
    output = model(images)
    
    # We got the probability for every 10 labels. The highest (max) probability should be correct label
    #_, predicted = torch.max(output, 1)
    predicted = output > 0.0
    
    # Let's show the predicted labels on the screen to compare with the real ones
    print('Predicted: ', ' '.join('%5s' % classes[predicted[j]] 
                              for j in range(batch_size)))

if __name__ == "__main__":
    
    # Let's build our model
    train(30)
    print('Finished Training')

    # Test which classes performed well
    testAccuracy()
    
    # Let's load the model we just created and test the accuracy per label
    model = Network()
    path = "myFirstModel.pth"
    model.load_state_dict(torch.load(path))

    # Test with batch of images
    testBatch()

    torch.onnx.export(model, torch.randn(batch_size, 3, 446, 2048), "saved_model.onnx", export_params=True, opset_version=10, do_constant_folding=True, input_names = ['input'], output_names = ['output'], dynamic_axes={'input' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}})
    onnx_model = onnx.load("saved_model.onnx")
    onnx.checker.check_model(onnx_model)
4

Try to rerun your code via export CUDA_LAUNCH_BLOCKING=1 or on the CPU to see which line of code fails.

5

Running on cpu given me an error on “labels = F.one_hot(labels, num_classes=1)” that says “Class values must be smaller than num_classes.”. The shape of labels is the same as my batch size for some reason.

6

Remove this line of code since the target should not be one-hot-enoded for nn.BCEWithLogitsLoss, but should have the same shape as the model output containing values in [0, 1].

7

Removing the line gave me “Target size (torch.Size([11])) must be the same as input size (torch.Size([11, 1]))” on loss = loss_fn, which I solved by calling unsqueeze(1) on labels.
It runs fine until the end of the first epoch, but then loss = loss_fn gives the error “The size of tensor a (11) must match the size of tensor b (2) at non-singleton dimension 0”, it looks like output and labels changed their shapes to [2,1] instead of [11,1]. Why did the shapes suddenly change, and aren’t they still matched if both of them are [2,1]?

Update: This problem was caused by my batch size not evenly dividing the dataset, so my last batch in the epoch was running on the last 2 images. Now the only problem is that my test accuracy is 920%

Current code:

import os
import torch
import torch.nn as nn
import torchvision
from torchvision import datasets, transforms
from torchvision.transforms import transforms
from torch.utils.data import DataLoader
from torch.optim import Adam
import matplotlib.pyplot as plt
import numpy as np
import torch.nn.functional as F
import onnx

os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:64"
torch.cuda.empty_cache()

# Loading and normalizing the data.
# Define transformations for the training and test sets
transformations = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

batch_size = 10
number_of_labels = 1

data_dir = "D:/Cropped Copper Plates/Edge Strips"

# Load training data
train_dir = os.path.join(data_dir, "Training")
train_set = datasets.ImageFolder(train_dir, transform=transformations)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=0)

# Load testing data
test_dir = os.path.join(data_dir, "Testing")
test_set = datasets.ImageFolder(test_dir, transform=transformations)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=0)

print("The number of images in a training set is: ", len(train_loader)*batch_size)
print("The number of images in a test set is: ", len(test_loader)*batch_size)

print("The number of batches per epoch is: ", len(train_loader))
classes = ('Defect')


class Network(nn.Module):
    def __init__(self):
        super(Network, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=3)
        self.bn1 = nn.BatchNorm2d(16)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(16, 32, kernel_size=3)
        self.drop = nn.Dropout2d()
        self.bn2 = nn.BatchNorm2d(32)
        self.fc1 = nn.Linear(7187840, 1)

    def forward(self, x):

        x = F.relu(self.bn1(self.conv1(x)))
        x = self.pool(x)
        x = F.relu(self.bn2(self.drop(self.conv2(x))))
        x = self.drop(x)
        x = torch.flatten(x, 1)
        #x = x.view(x.size(0), -1)
        #print(x.shape)
        #x = F.relu(self.fc1(x))
        x = self.fc1(x)
        #x = F.dropout(x, training = self.training)
       
        return x

# Instantiate a neural network model 
model = Network()

 
# Define the loss function with Classification Cross-Entropy loss and an optimizer with Adam optimizer
pos_weight = torch.full([batch_size, 1], 8.5238)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
#device = torch.device("cpu")
loss_fn = nn.BCEWithLogitsLoss(pos_weight=pos_weight)
loss_fn = loss_fn.to(device)
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, weight_decay=0.001)

from torch.autograd import Variable

# Function to save the model
def saveModel():
    path = "./myFirstModel.pth"
    torch.save(model.state_dict(), path)


# Training function. We simply have to loop over our data iterator and feed the inputs to the network and optimize.
def train(num_epochs):
    
    best_accuracy = 0.0

    # Define your execution device
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    #device = torch.device("cpu")
    print("The model will be running on", device, "device")
    # Convert model parameters and buffers to CPU or Cuda
    model.to(device)

    for epoch in range(num_epochs):  # loop over the dataset multiple times
        running_loss = 0.0
        running_acc = 0.0
        accum_iter = 4;
        for i, (images, labels) in enumerate(train_loader, 0):
            
            # get the inputs
            images = images.to(device)
            labels = labels.to(device)
            images = images.float()
            labels = labels.float().unsqueeze(1)

            # predict classes using images from the training set
            outputs = model(images)
            labels = labels.float()
            outputs = outputs.float()
            # compute the loss based on model output and real labels
            loss = loss_fn(outputs, labels)
            loss = loss/accum_iter
            # backpropagate the loss
            loss.backward()
            if((i+1) % accum_iter == 0) or (i+1 == len(train_loader)):
                # adjust parameters based on the calculated gradients
                optimizer.step()
                # zero the parameter gradients
                optimizer.zero_grad()

            running_loss += loss.item()     # extract the loss value
            if i % 10 == 9:    
                # print every 100 (twice per epoch) 
                print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 100))
                # zero the loss
                running_loss = 0.0

        # Compute and print the average accuracy fo this epoch when tested over all test images
        accuracy = testAccuracy()
        print('For epoch', epoch+1,'the test accuracy over the whole test set is %d %%' % (accuracy))
        
        # we want to save the model if the accuracy is the best
        if accuracy > best_accuracy:
            saveModel()
            best_accuracy = accuracy

# Function to test the model with the test dataset and print the accuracy for the test images
def testAccuracy():
    
    model.eval()
    accuracy = 0.0
    total = 0.0
    
    with torch.no_grad():
        for data in test_loader:
            images, labels = data
            device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
            #device = torch.device("cpu")
            images, labels = images.to(device), labels.to(device)
            images = images.float()
            labels = labels.float()
            # run the model on the test set to predict labels
            output = model(images)
            # the label with the highest energy will be our prediction
            #_, predicted = torch.max(output.data, 1)
            predicted = output > 0.0
            total += labels.size(0)
            accuracy += (predicted == labels).sum().item()
    
    # compute the accuracy over all test images
    accuracy = (100 * accuracy / total)
    return(accuracy)

# Function to show the images
def imageshow(img):
    img = img / 2 + 0.5     # unnormalize
    npimg = img.numpy()
    plt.imshow(np.transpose(npimg, (1, 2, 0)))
    plt.show()


# Function to test the model with a batch of images and show the labels predictions
def testBatch():
    # get batch of images from the test DataLoader  
    images, labels = next(iter(test_loader))

    # show all images as one image grid
    #imageshow(torchvision.utils.make_grid(images))
   
    # Show the real labels on the screen 
    print('Real labels: ', ' '.join('%5s' % classes[labels[j]] for j in range(batch_size)))
  
    # Let's see what if the model identifiers the  labels of those example
    output = model(images)
    
    # We got the probability for every 10 labels. The highest (max) probability should be correct label
    #_, predicted = torch.max(output, 1)
    predicted = output > 0.0
    
    # Let's show the predicted labels on the screen to compare with the real ones
    print('Predicted: ', ' '.join('%5s' % classes[predicted[j]] 
                              for j in range(batch_size)))

if __name__ == "__main__":
    
    # Let's build our model
    train(30)
    print('Finished Training')

    # Test which classes performed well
    testAccuracy()
    
    # Let's load the model we just created and test the accuracy per label
    model = Network()
    path = "myFirstModel.pth"
    model.load_state_dict(torch.load(path))

    # Test with batch of images
    testBatch()

    torch.onnx.export(model, torch.randn(batch_size, 3, 446, 2048), "saved_model.onnx", export_params=True, opset_version=10, do_constant_folding=True, input_names = ['input'], output_names = ['output'], dynamic_axes={'input' : {0 : 'batch_size'}, 'output' : {0 : 'batch_size'}})
    onnx_model = onnx.load("saved_model.onnx")
    onnx.checker.check_model(onnx_model)
8

Your code looks alright and the accuracy calculation should work as seen in this code snippet mimicking your code:

accuracy = 0.0
total = 0.0

for batch in range(100):
    output = torch.randn(16, 1)
    labels = torch.randint(0, 2, (16, 1)).float()
    predicted = output > 0.0
    total += labels.size(0)
    accuracy += (predicted == labels).sum().item()

# compute the accuracy over all test images
accuracy = (100 * accuracy / total)
print(accuracy)
# 50.125

You would have to inspect the predicted, total, as well as intermediate accuracy values to see why accuracy is larger than total.