Hello,

I am struggling with the cifar10 autoencoder model training. The loss goes down, but the accuracy does not go up.

As a starting point, I use the architecture of the CNN autoencoder used for FashionMNIST in Lab 4 of this week. So far I have experimented with the following:

- more conv2d layers
- less conv2d layers
- kernel size tweaking
- pool size tweaking
- number of filters tweaking
- more epochs

As well as several combinations of those. Always keeping in mind the symmetrical shape of the architecture.

Are there any parameters that I have overlooked, which could help in increasing the accuracy?

Thanks and regards,

Prashant

1 Like

If its a matter of model design then you also need Batchnormalizations and and an upsampling layer. I think the structure of the notebook and the model is similar to the labs of that week.

1 Like

The model can be really similar to the one in the LAB 4 of this same week.

Conv2d and MaxPooling in the encoder and Conv2d and upsampling in the decoder will be enough.

How many epochs are you training? And which loss did you get at the final epoch?

Regards!

1 Like

I figured it out, turned out I simply forgot to take into account that these pictures are colour pictures, and thus the number of filters of the last Conv2D layer was incorrect.

After changing that number it worked well.

Thanks

2 Likes

This was a great find. Forgot we were only working with grayscale in the lab. Man, so much of deep learning comes down to counting shapes of tensors its funny.