Получение предвзятого прогноза для моей CNN

Question

В настоящее время я создаю модель CNN, которая может брать набор данных, тренироваться на нем и при тестировании производить высокую скорость классификации. Он работает с набором данных MNIST из pytorch.torchvision, но когда я использую свой собственный набор данных, прогноз обучения очень близок к целевому классу, но прогноз тестирования всегда является одним целевым классом, что означает низкий уровень классификации. Я действительно новичок в этом, и мне нужна помощь, пожалуйста.

Ниже мой учебный и тестовый класс:

Тренировка Def

    # Forward Pass
    Outputs = cnnModel(Input)
    # print ('T: ' + str(Target))
    Loss = criterion(Outputs, Target)
    lossList.append(Loss.item())

    # Backward & Optimise
    optimiser.zero_grad()
    Loss.backward()
    optimiser.step()

    #Track Accuracy
    totalResult += Target.size(0)
    _, predictedResult = torch.max(Outputs.data, 1)
    corretResult += (predictedResult == Target).sum().item()
    wrongResult += (predictedResult != Target).sum().item()
    actualList.append(corretResult / totalResult)

    targetList.append(str(Target))
    predList.append(str(predictedResult))
    targetCounter = Counter(targetList)
    predCounter = Counter(predList)


print ('TL COUNT' + str(targetCounter))
print ('PL COUNT' + str(predCounter))
print ('Average Training Accuracy Loss on this Epoch: ' + str(nP.array(actualList).mean()))
print('Accuracy of the Covnet Neural Network model on the Test Images: {} %'.format(100 * corretResult / totalResult))

Тестирование Def

        Input = Input.to(deviceConfiguration)
        Target = Target.to(deviceConfiguration)

        Outputs = cnnModel(Input)
        temp, predictedResult = torch.max(Outputs.data, 1)
        print(temp)
        print(predictedResult)
        print("\n")
        totalResult += Target.size(0)
        corretResult += (predictedResult == Target).sum().item()
        wrongResult += (predictedResult != Target).sum().item()
        targetList.append(str(Target))
        predList.append(str(predictedResult))
        targetCounter = Counter(targetList)
        predCounter = Counter(predList)
        #cont = input('Continue? ')


    for key, value in targetCounter.items():
        print('Target Class & Occurance: ' + key, value)
    print("\n")

    for key, value in predCounter.items():
        print('Pred Class & Occurance: ' + key, value)
    print("\n")

    print('Accuracy of the Covnet Neural Network model on the Test Images: {} %'.format(100 * corretResult / totalResult))
    print('Misclassification Rate of the Dataset Images : {} %'.format(100 * wrongResult / totalResult))

Мой CNN для вывода

def forward(self, x):
    out = self.layer1(x)
    out = self.layer2(out)
    out = self.layer3(out)
    out = out.reshape(out.size(0), -1)
    out = self.fc1(out)
    #print ('O: ' + str(out.size()))
    return out

Мой CSV Reader

def __getitem__(self, index):
    singleImageLabel = self.dataLabels[index]
    # print ('Path: ' + str(singleImageLabel))

    # Create an Empty Numpy Array to Fill
    imageAsNumpy = nP.ones((32, 32), dtype = 'uint8')

    # Fill the Numpy Array with Data from Pandas DF
    for i in range(1):
        rowPosition = (i-1) // self.imageHeight
        columnPosition = (i-1) % self.imageWidth
        indexFirst = self.dataFromCSV.iloc[index][i].split(";")[3]
        indexLast = self.dataFromCSV.iloc[index][i].split(";")[6]
        imageAsNumpy[rowPosition][columnPosition] = indexFirst + indexLast

    imageAsImage = Image.fromarray(imageAsNumpy)
    imageAsImage = imageAsImage.convert('RGB')

    # Transform Image to Tensor
    if self.transform is not None:
        imageAsTensor = self.transform(imageAsImage)

    # Transform Label to Tensor
    labelAsLabel = int(singleImageLabel.split(";")[7])
    labelAsTensor = torch.from_numpy(nP.array(labelAsLabel))
    # print ('Target: ' + str(labelAsTensor))

    # Return Image & the Label
    return (imageAsTensor, labelAsLabel)

def __len__(self):
    return len(self.dataFromCSV.index)

И это вывод - в терминале

Accuracy of the Covnet Neural Network model on the Test Images: 0.0 %
Misclassification Rate of the Dataset Images : 100.0 %
Rajans-MBP:CNN_Model_GTSRB rajanbindra$ python ./CNN_Model.py
TL COUNTCounter({'tensor([ 26])': 2, 'tensor([ 24])': 1, 'tensor([ 8])': 1, 'tensor([ 9])': 1, 'tensor([ 12])': 1, 'tensor([ 25])': 1, 'tensor([ 2])': 1, 'tensor([ 29])': 1, 'tensor([ 38])': 1})
PL COUNTCounter({'tensor([ 24])': 6, 'tensor([ 26])': 3, 'tensor([ 3])': 1})
Average Training Accuracy Loss on this Epoch: 0.0
Accuracy of the Covnet Neural Network model on the Test Images: 0.0 %
Misclassification Rate of the Dataset Images : 100.0 %
Continue? 

Target Class & Occurance: tensor([ 10]) 1
Target Class & Occurance: tensor([ 35]) 1
Target Class & Occurance: tensor([ 28]) 1
Target Class & Occurance: tensor([ 5]) 1
Target Class & Occurance: tensor([ 12]) 1
Target Class & Occurance: tensor([ 2]) 2
Target Class & Occurance: tensor([ 3]) 1
Target Class & Occurance: tensor([ 1]) 1
Target Class & Occurance: tensor([ 17]) 1


Pred Class & Occurance: tensor([ 26]) 10


Accuracy of the Covnet Neural Network model on the Test Images: 0.0 %
Misclassification Rate of the Dataset Images : 100.0 %