Can't reach desired accuraacy for C2W4 task

I know this topic was discussed several times earlier. But I’ve already spent 2 days trying to use hints from other similar questions, and it still didn’t work out for me, so maybe someone can help with a less vague explanation. My current accuracy is around 90 and validation accuracy is 94.

1 Like

What course are you attending?

1 Like

Convolutional Neural Networks in TensorFlow

You posted this in the General Discussions category. Please move it to the relevant course category (TF1 Course 2) as described here. Once you move it to the correct category, mentors of that course will be notify and they will respond you.

Thank you, @IO-321, for moving your query to the right category.

So, you have difficulty improving the accuracy. It is given that you should use no more than 2 Conv2D and 2 MaxPooling2D layers to achieve the desired performance. What is your current architecture and your optimizer & loss?

optimizer = RMSprop with learning rate = 0.001 and loss = sparce_categorical_crossentropy

As for architecture, I have exactly 2 Conv2D layers and 2 MaxPooling2D layers. I also have one Flatten and one Dropout, one Dense layer with 512 neurons, and a final Dense layer with 26 neurons and softmax activation.

Play with different optimizers like Adam or something, and learning_rate.

1 Like

This is a pretty frustrating assignment. I’ve been looking at feedback from different mentors for people saying that they are having issues achieving the accuracy goals. It looks like a lot of people are having trouble with knowing whether they need to one-hot encode the data, but even after this is addressed the accuracy is very low in some cases. I am at 4% accuracy and I can’t seem to figure out what is wrong. I’ve been experimenting with different optimizers but I can’t tell if the inputs to the model have an error that your code-checkers are not catching.

Hi, for me personally the biggest impact was after adding batch normalisation layers to my model, also try to simplify your augmentation and maybe increase amount of dropout. I used in my task architecture shown in the lecture and after implementation of improvements above it worked out.

1 Like

Please see this topic.

Balaji many of the learners are creating multiple threads and multiple questions with the same week. I have already guided Nick for this assignment. He has had errors in his assignment notebook. I guess he is still working on it.

Hello @Nick_VN ,

Do not post your same doubts in multiple threads.
If you see solutions of the threads are not helping you, create a sparate topic. Mentors will follow up with you.

Check this guide : Messages & Notifications

With regards,
Nilosree Sengupta