Cat Dot Classifier Bad Accuracy

Python
1

Tell us what’s happening:

So after following the steps as closely as I could my model is only giving a 50% accuracy on the validation data after about the 3rd or 4th epoch. The video lectures did not really cover anything about using validation data over just training/testing data and also did not cover ImageDataGenerators or when to use various classifiers and optimizers so I’m not really sure where I might have gone wrong but my model is no better than a coin flip and on a binary classification problem it really can’t get worse than that. I’ve been reading up on some of these concepts and looking at examples but having someone take a look to maybe give me some general direction or suggestions would be helpful. I suspect that despite the fact that the project has me using random transformations to add images to the limited ones provided but I suspect that may also have something to do with it. Also, does anyone have any suggestions about how I can rerun the code faster as I am making alterations to the code an hyper parameters. I try to make a change here or there to see how it will effect the model but rerunning the entire notebook (especially the training part) takes 20-30 minutes which is a long time just to see whether or not something may have made a difference. Here is a link to my notebook. Any suggestions or feedback would be appreciated. Thank you.

Your code so far

https://colab.research.google.com/drive/1qDhEqMqV-rW4gtBbx3jHS2eUziCEJekX?usp=sharing

Challenge: Cat and Dog Image Classifier

Link to the challenge:

2

Go to runtime > change runtime type > select GPU to speed up the training.

I think you might have the Dropout in the wrong place. Here is my model building sequence, for comparison:

model = Sequential()

model.add( Conv2D(32, (3,3), activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH, 3) ) )
model.add( MaxPooling2D( (2, 2) ) )

model.add( Conv2D(64, (3,3), activation='relu', input_shape=(IMG_HEIGHT, IMG_WIDTH, 3) ) )
model.add( MaxPooling2D( (2, 2) ) )

model.add( Dropout(0.5) )

model.add( Flatten() )
model.add( Dense(128, activation='relu') )
# corrected the following line to avoid confusion if anyone else looks at this
model.add( Dense(2, activation='softmax') )

I also used different settings for the compling process, but I don’t think that would be a big issue.

model.compile(optimizer = 'adam',
    loss = tf.keras.losses.sparse_categorical_crossentropy,
    metrics = ['accuracy'])

image
image739×484 35.6 KB

3

Thanks for the feedback. Changing the runtime to GPU really sped things up but unfortunately the accuracy has not improved. I actually began changing my model bit by bit, including the optimizer and loss function to the same ones you had used to see what would have had the biggest impact on the accuracy but nothing worked. In the end I copied and pasted your solution (with regards to the model) and I still can’t get the accuracy above 50%. Guess that tells me that its more likely to do with the arguments I’m using for the ImageDataGenerator but after studying the documentation from Keras I’m still not sure where I might be going wrong. It seems like it should be right.

4

I played around with your code and if you change the dense layer to a 2, it will work.

model.add( Dense(2, activation='softmax') )

Then your probability will be given in a vector format [prob1, prob2].

image
image780×260 15.9 KB

Hope this helps.

5

Ah that was it. Thank you so much. I love this community.

1 Like