Preserving identity about stargan-v2 HOT 3 OPEN

is there specific recommendations about lambda_ds in general?
should it be equal to the domain count (minus one?)?
training 7 domains with lambda_ds=2 results in rather subtle variety

from stargan-v2.

@eps696
I have same question. I tested only 3 domain with my own dataset, and it preserves the identity in certain degree.

from stargan-v2.

@iPsych i have trained ~dozen models with 7-8 domains on two configs: img_size 512 batch 1 and img_size 256 batch 3 (these are maximum batchcounts for original code on 11gb 1080ti).
i got satisfactory results (worse than original model though) with lambda_ds, lambda_cyc & lambda_sty all equal 3. increasing lambda_ds indeed leads to losing identity (increasing lambda_cyc seems to get it back).

two more observations:

• lower batch counts definitely lead to worse quality; the experiments with img_size 512 batch 1 were not quite successful. as the authors said, the result is not guaranteed with batch count < 8.
• after adding del x_fake2, y_org, y_trg, s_trg2, s_trg, s_pred, out before this line, i was able to start training with img_size 256 batch 4 on 11gb gpu (no results from that yet).

from stargan-v2.