Shell script:

python taskgen.py \\
-discord_notify_taskgen True \\
-is_force_start False \\
-num_tasks_in_parallel 2 \\
-is_single_cuda_device True \\
-num_cuda_devices_per_task 1 \\
-script main_bicyclegan.py \\
-sequence_name grid_search_alarm_bicyclegan_2 \\
-generator model_unet_generator_bicyclegan \\
-discriminator model_multi_discriminator_bicyclegan \\
-encoder model_encoder_bicyclegan \\
-datasource datasource \\
-data_path /home/asya/documents/alarm_flight_model/ImagesGT_cleaned \\
-in_channels 1 \\
-img_size 256 512 1024 \\
-data_workers 4 \\
-image_size_correction reduced_by_4 \\
-lambda_pixel 5 10 100 \\
-lambda_latent 0.5 0.1 1.0 \\
-lambda_kl 0.01 0.02 0.1 \\
-latent_dim (4 8 32, asya-3), (16 32, asya-4) \\ 
-epochs 10_000 \\
-batch_size 1 \\
-optimizer adam \\
-learning_rate 1e-4 \\
-is_single_cuda_device True

All runs:

blue: asya-3; yellow: asya-4

Screenshot 2022-10-14 at 11.07.37.png

Last few epoch outputs:

12th run (yellow: asya-4):

9900_0.png

9800_0.png

10000_0.png

9th run (blue: asya-3)

9800_0.png

9900_0.png

10000_0.png

11th run (yellow: asya-4):

9800_0.png

9900_0.png

10000_0.png

Inference results:

12th run (yellow: asya-4):

10 000 epoch:

demo_cleaned_1_both.png

master_AdityaRasam_P1_20_D_1_1_1_1_both.png

master_DonMadden_P1_20_D_1_1_1_1_both.png

master_GlennTournier_P2_20_D_1_1_1_1_both.png