I try to write the training phase, but there are some problems occurring

[Sweetclover]

[belkakari]

You can ask about the problems here, we will be happy to help

[Sweetclover]

After computing the loss value, the oldest styles are removed from the pool to keep its size at T.

Can I choose 8 as the value of T?

[belkakari]

In the paper, we've used T=100. This was not an issue in terms of memory since we've stored only style vectors of relatively small size (3 for daytime and 12 for wikiart), but you definitely are welcome to experiment with the hyperparameters

[Sweetclover]

initial learning rate 0:0001 for both generators and discriminators, halving the learning rate every 200000 iterations.

But what's the total number of your iterations during training?

[denkorzh]

For the dataset of landscapes:

We trained the translation model for 450 thousand iterations with batch size four on a single NVIDIA Tesla P40.