I have trained my CNN model on CIFAR 10 and I got val_accuracy of 87% which is not a low value but when it comes to detection of pictures my model detected most of the pictures wrong. anyone knows why this is happening and how to solve this problem.
4
-
$\begingroup$ I got a similar issue last time I trained a CNN, in my case i don't know exactly what happened, but it seems the classes got mixed up when training and I had to rearrange the order of the outputs. for example network always returned class 1 when it should have returned class 2, etc. So I had to create a dictionnaries to link each class to the correct output. $\endgroup$– UbikuityCommented May 4, 2021 at 9:43
-
$\begingroup$ Thank you for the comment but I am a beginner and this is a university assignment so if you can provide me with the code that you have used it we will be so helpful. $\endgroup$– Osama HamadaCommented May 4, 2021 at 9:45
-
$\begingroup$ I will post the way I solved it in an answer, just remember that this was the solution for me, but might not the solution for your network. Just try sending images from a same class and see it it gives the same class everytime to see your network has the same problem as mine. $\endgroup$– UbikuityCommented May 4, 2021 at 9:50
-
$\begingroup$ I have been given images to detect, I have downloaded an image of a house and send it to my model but my model detect it as a cat. I downloaded 3 more images of houses and my model detect them right this time so I am confused. Now I have 4 house images my model detected one as a cat and the other 3 as houses. $\endgroup$– Osama HamadaCommented May 4, 2021 at 10:03
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Here is how i fixed the classes not corresponding to each other, code is done with Pytorch, most of my function are custom ones but I wrote the equivalent in comments :
def main():
modelPath = '../Results/BestModel/model.ntw'
model = Om.openModel(modelPath) # torch.load(modelPath)
dataPath = 'food-101/images' # DataSet path
listClasses = os.listdir(dataPath) # List of classes, here each folder in dataPath is a class
dicClasses = {}
for i, imClass in enumerate(listClasses):
if i % 5 == 0:
print(str(i) + ' / ' + str(len(listClasses)))
res = []
lIm = os.listdir(os.path.join(dataPath, imClass))
random.shuffle(lIm)
for imName in lIm[:20]: # Take 20 images of each class
im = Om.loadImage(os.path.join(dataPath, imClass, imName)) # Loading image with correct transforms applied
res.append(model(im).cpu().numpy()) # Store the result of the prediction
resmed = np.mean(res, 0)[0] # Sum all predictions
r = indexesMax(resmed, 1) # function is the same as torch.argmax(resmed).
dicClasses[imClass] = r
print(dicClasses)
