Image-to-Image

An Image-to-Image PyTorch repo. Primary done for PhysGen/Urban Noise Dataset.

Data References:

• Dataset
• Main Paper / comparison
• Normalizing Flow Approach

Content:

• Installation
• Process
• Example Commands
• Experiment Commands
• Access Experiment Tracking

For the Physgen benchmark, you have to first train your model, then open the Evaluation Physgen-Benchmark Notebook and run the commands (and adjust the parameters there).

> Documentation <

---

Installation

Clone/Download the repo and use it then:

conda create -n img-to-img python=3.13 pip -y
conda activate img-to-img
pip install --no-cache-dir -r requirements.txt

For a GPU version you can run this afterwards:

pip install --no-cache-dir torch torchvision --index-url https://download.pytorch.org/whl/cu126

Use your CUDA version needed. You may want to check with nvidia-smi.

---

Process

1. Clone/download this repository
2. Run the installation
3. Open for example anaconda prompt / bash and navigate to your project folder

   cd "D:\Studium\Master\Repos\Image-to-Image"
   D:

4. Activate the conda env:

   conda activate img-to-img

5. Maybe check your hardware with:

   nvidia-smi

6. Preparation -> Folder creation

   mkdir logs

7. Run Train script with logging:

   nohup python ./image-to-image/main.py \
   --mode train \
   --epochs 100 \
   --batch_size 16 \
   --lr 0.0001 \
   --loss l1 \
   --optimizer adam \
   --weight_decay \
   --weight_decay_rate 0.0004 \
   --gradient_clipping \
   --gradient_clipping_threshold 0.5 \
   --scheduler step \
   --activate_amp \
   --amp_scaler grad \
   --checkpoint_save_dir ./checkpoints \
   --save_only_best_model \
   --validation_interval 5 \
   --model resfcn \
   --resfcn_in_channels 1 \
   --resfcn_hidden_channels 64 \
   --resfcn_out_channels 1 \
   --resfcn_num_blocks 16 \
   --data_variation sound_reflection \
   --input_type osm \
   --output_type standard \
   --device cuda \
   --experiment_name image-to-image \
   --run_name resfcn_test \
   --tensorboard_path ./tensorboard \
   --save_path ./mlflow_images \
   --cmap gray \
   > ./logs/resfcn_test.log 2>&1 &

   or at windows:

   start /B python ./image-to-image/main.py `
   --mode train `
   --epochs 100 `
   --batch_size 16 `
   --lr 0.0001 `
   --loss l1 `
   --optimizer adam `
   --weight_decay `
   --weight_decay_rate 0.0004 `
   --gradient_clipping `
   --gradient_clipping_threshold 0.5 `
   --scheduler step `
   --activate_amp `
   --amp_scaler grad `
   --checkpoint_save_dir ./checkpoints `
   --save_only_best_model `
   --validation_interval 5 ^
   --model resfcn `
   --resfcn_in_channels 1 `
   --resfcn_hidden_channels 64 `
   --resfcn_out_channels 1 `
   --resfcn_num_blocks 16 `
   --data_variation sound_reflection `
   --input_type osm `
   --output_type standard `
   --device cuda `
   --experiment_name image-to-image `
   --run_name resfcn_test `
   --tensorboard_path ./tensorboard `
   --save_path ./mlflow_images `
   --cmap gray `
   > ./logs/resfcn_test.log 2>&1

8. Now check you log-file and in worse case stop the run:
   Linux:

   ps aux | grep '[p]ython'
   kill -9 -f python

   Windows CMD:

   tasklist /FI "IMAGENAME eq python.exe"
   taskkill /IM python.exe /F

   Windows PowerShell:

   Get-WmiObject Win32_Process | Where-Object { $_.CommandLine -like "*python*" } | Select-Object ProcessId, CommandLine
   Get-WmiObject Win32_Process | Where-Object { $_.CommandLine -like "*python*" } | ForEach-Object { Stop-Process -Id $_.ProcessId -Force }

9. If training is finish, you can open the . > image_to_image > model_interactions > eval_physgen_benchmark.ipynb (or click me), change the variables and run the code blocks to evaluate your model.
10. You can also check your mlrun/tensorboard metrics, to find out how the training was -> see here

When pasting multiline commands on windows, use ` for splitting commands when using Windows PowerShell, for CMD or CMD-kinds use ^ and for bash/linux like command terminales you might want to try \.

See also the Google Coolab notebook.

---

Example Commands

Training run

python ./main.py \
  --mode train \
  --epochs 100 \
  --batch_size 16 \
  --lr 0.0001 \
  --loss l1 \
  --optimizer adam \
  --weight_decay \
  --weight_decay_rate 0.0004 \
  --gradient_clipping \
  --gradient_clipping_threshold 0.5 \
  --scheduler step \
  --activate_amp \
  --amp_scaler grad \
  --checkpoint_save_dir ./checkpoints \
  --save_only_best_model \
  --validation_interval 5 \
  --model resfcn \
  --resfcn_in_channels 1 \
  --resfcn_hidden_channels 64 \
  --resfcn_out_channels 1 \
  --resfcn_num_blocks 16 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda \
  --experiment_name image-to-image \
  --run_name resfcn_test \
  --tensorboard_path ./tensorboard \
  --save_path ./mlflow_images \
  --cmap gray

Testing

python ./main.py \
  --mode test \
  --batch_size 16 \
  --loss l1 \
  --model resfcn \
  --model_params_path ./checkpoints/my_model.pth \
  --resfcn_in_channels 1 \
  --resfcn_hidden_channels 64 \
  --resfcn_out_channels 1 \
  --resfcn_num_blocks 16 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda

Inference

python ./main.py \
  --mode inference \
  --batch_size 16 \
  --model resfcn \
  --model_params_path ./checkpoints/my_model.pth \
  --resfcn_in_channels 1 \
  --resfcn_hidden_channels 64 \
  --resfcn_out_channels 1 \
  --resfcn_num_blocks 16 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda

Or custom data inference:

python ./main.py \
  --mode inference \
  --batch_size 16 \
  --model resfcn \
  --model_params_path ./checkpoints/my_model.pth \
  --resfcn_in_channels 1 \
  --resfcn_hidden_channels 64 \
  --resfcn_out_channels 1 \
  --resfcn_num_blocks 16 \
  --image_dir_path ./data/dataset \
  --device cuda

In windows you have to use ``` or ^ for the line breaks or remove the line breaks completly. Or you use Linux via Docker.

---

Experiment Commands

Pix2Pix Rebuild

Linux:

nohup python ./main.py \
  --mode train \
  --epochs 100 \
  --batch_size 16 \
  --lr 0.0001 \
  --loss weighted_combined \
  --wc_loss_silog_lambda 0.5 \
  --wc_loss_weight_silog 1.0 \
  --wc_loss_weight_grad 50.0 \
  --wc_loss_weight_ssim 100.0 \
  --wc_loss_weight_edge_aware 50.0 \
  --wc_loss_weight_l1 10.0 \
  --wc_loss_weight_var 0.0 \
  --wc_loss_weight_range 0.0 \
  --wc_loss_weight_blur 0.0 \
  --optimizer adam \
  --optimizer_2 adam \
  --weight_decay \
  --weight_decay_rate 0.0004 \
  --gradient_clipping \
  --gradient_clipping_threshold 0.5 \
  --scheduler step \
  --scheduler_2 step \
  --activate_amp \
  --amp_scaler grad \
  --checkpoint_save_dir ./checkpoints \
  --save_only_best_model \
  --validation_interval 5 \
  --model pix2pix \
  --pix2pix_in_channels 1 \
  --pix2pix_hidden_channels 64 \
  --pix2pix_out_channels 1 \
  --pix2pix_second_loss_lambda 100 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda \
  --experiment_name image-to-image \
  --run_name pix2pix_rebuild_test \
  --tensorboard_path ./tensorboard \
  --save_path ./mlflow_images \
  --cmap gray \
  > ./logs/pix2pix_rebuild_test.log 2>&1 &

Windows:

start /B python ./main.py ^
  --mode train ^
  --epochs 100 ^
  --batch_size 16 ^
  --lr 0.0001 ^
  --loss weighted_combined ^
  --wc_loss_silog_lambda 0.5 ^
  --wc_loss_weight_silog 1.0 ^
  --wc_loss_weight_grad 50.0 ^
  --wc_loss_weight_ssim 100.0 ^
  --wc_loss_weight_edge_aware 50.0 ^
  --wc_loss_weight_l1 10.0 ^
  --wc_loss_weight_var 0.0 ^
  --wc_loss_weight_range 0.0 ^
  --wc_loss_weight_blur 0.0 ^
  --optimizer adam ^
  --optimizer_2 adam ^
  --weight_decay ^
  --weight_decay_rate 0.0004 ^
  --gradient_clipping ^
  --gradient_clipping_threshold 0.5 ^
  --scheduler step ^
  --scheduler_2 step ^
  --activate_amp ^
  --amp_scaler grad ^
  --checkpoint_save_dir ./checkpoints ^
  --save_only_best_model ^
  --validation_interval 5 ^
  --model pix2pix ^
  --pix2pix_in_channels 1 ^
  --pix2pix_hidden_channels 64 ^
  --pix2pix_out_channels 1 ^
  --pix2pix_second_loss_lambda 100 ^
  --data_variation sound_reflection ^
  --input_type osm ^
  --output_type standard ^
  --device cuda ^
  --experiment_name image-to-image ^
  --run_name pix2pix_rebuild_test ^
  --tensorboard_path ./tensorboard ^
  --save_path ./mlflow_images ^
  --cmap gray ^
  > ./logs/pix2pix_rebuild_test.log 2>&1

start /B python ./main.py ^
  --mode train ^
  --epochs 50 ^
  --batch_size 16 ^
  --lr 0.0005 ^
  --loss weighted_combined ^
  --wc_loss_silog_lambda 0.5 ^
  --wc_loss_weight_silog 0.1 ^
  --wc_loss_weight_grad 5.0 ^
  --wc_loss_weight_ssim 1.0 ^
  --wc_loss_weight_edge_aware 5.0 ^
  --wc_loss_weight_l1 1.0 ^
  --wc_loss_weight_var 0.0 ^
  --wc_loss_weight_range 0.0 ^
  --wc_loss_weight_blur 0.0 ^
  --optimizer adamw ^
  --optimizer_2 adam ^
  --weight_decay ^
  --weight_decay_rate 0.01 ^
  --gradient_clipping ^
  --gradient_clipping_threshold 2.0 ^
  --scheduler cosine ^
  --scheduler_2 step ^
  --activate_amp ^
  --amp_scaler grad ^
  --checkpoint_save_dir ./checkpoints ^
  --save_only_best_model ^
  --validation_interval 1 ^
  --model pix2pix ^
  --pix2pix_in_channels 1 ^
  --pix2pix_hidden_channels 64 ^
  --pix2pix_out_channels 1 ^
  --pix2pix_second_loss_lambda 500.0 ^
  --data_variation sound_reflection ^
  --input_type osm ^
  --output_type standard ^
  --device cuda ^
  --experiment_name image-to-image ^
  --run_name pix2pix_rebuild_test ^
  --tensorboard_path ./tensorboard ^
  --save_path ./mlflow_images ^
  --cmap gray ^
  > ./logs/pix2pix_rebuild_test.log 2>&1

New Simple Model try out

Linux:

nohub python ./main.py \
  --mode train \
  --epochs 100 \
  --batch_size 16 \
  --lr 0.0001 \
  --loss weighted_combined \
  --wc_loss_silog_lambda 0.5 \
  --wc_loss_weight_silog 1.0 \
  --wc_loss_weight_grad 50.0 \
  --wc_loss_weight_ssim 100.0 \
  --wc_loss_weight_edge_aware 50.0 \
  --wc_loss_weight_l1 10.0 \
  --wc_loss_weight_var 0.0 \
  --wc_loss_weight_range 0.0 \
  --wc_loss_weight_blur 0.0 \
  --optimizer adam \
  --weight_decay \
  --weight_decay_rate 0.0004 \
  --gradient_clipping \
  --gradient_clipping_threshold 0.5 \
  --scheduler step \
  --activate_amp \
  --amp_scaler grad \
  --checkpoint_save_dir ./checkpoints \
  --save_only_best_model \
  --validation_interval 5 \
  --model resfcn \
  --resfcn_in_channels 1 \
  --resfcn_hidden_channels 64 \
  --resfcn_out_channels 1 \
  --resfcn_num_blocks 16 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda \
  --experiment_name image-to-image \
  --run_name resfcn_weighted_combined_test \
  --tensorboard_path ./tensorboard \
  --save_path ./mlflow_images \
  --cmap gray \
  > ./logs/resfcn_weighted_combined_test.log 2>&1 &

Windows:

start /B python ./main.py ^
  --mode train ^
  --epochs 50 ^
  --batch_size 16 ^
  --lr 0.001 ^
  --loss weighted_combined ^
  --wc_loss_silog_lambda 0.5 ^
  --wc_loss_weight_silog 0.1 ^
  --wc_loss_weight_grad 5.0 ^
  --wc_loss_weight_ssim 10.0 ^
  --wc_loss_weight_edge_aware 5.0 ^
  --wc_loss_weight_l1 1.0 ^
  --wc_loss_weight_var 0.0 ^
  --wc_loss_weight_range 0.0 ^
  --wc_loss_weight_blur 0.0 ^
  --optimizer adamw ^
  --weight_decay ^
  --weight_decay_rate 0.05 ^
  --gradient_clipping ^
  --gradient_clipping_threshold 2.0 ^
  --scheduler step ^
  --activate_amp ^
  --amp_scaler grad ^
  --checkpoint_save_dir ./checkpoints ^
  --save_only_best_model ^
  --validation_interval 2 ^
  --model resfcn ^
  --resfcn_in_channels 1 ^
  --resfcn_hidden_channels 64 ^
  --resfcn_out_channels 1 ^
  --resfcn_num_blocks 16 ^
  --data_variation sound_reflection ^
  --input_type osm ^
  --output_type standard ^
  --device cuda ^
  --experiment_name image-to-image ^
  --run_name resfcn_weighted_combined_test ^
  --tensorboard_path ./tensorboard ^
  --save_path ./mlflow_images ^
  --cmap gray ^
  > ./logs/resfcn_weighted_combined_test.log 2>&1

Residual Design Model Test

Linux:

nohup python ./main.py \
  --mode train \
  --epochs 120 \
  --batch_size 12 \
  --lr 0.0001 \
  --loss weighted_combined \
  --wc_loss_silog_lambda 0.5 \
  --wc_loss_weight_silog 1.0 \
  --wc_loss_weight_grad 40.0 \
  --wc_loss_weight_ssim 80.0 \
  --wc_loss_weight_edge_aware 40.0 \
  --wc_loss_weight_l1 8.0 \
  --wc_loss_weight_var 0.0 \
  --wc_loss_weight_range 0.0 \
  --wc_loss_weight_blur 0.0 \
  --optimizer adam \
  --optimizer_2 adamw \
  --weight_decay \
  --weight_decay_rate 0.0004 \
  --gradient_clipping \
  --gradient_clipping_threshold 0.5 \
  --scheduler step \
  --scheduler_2 step \
  --activate_amp \
  --amp_scaler grad \
  --checkpoint_save_dir ./checkpoints \
  --save_only_best_model \
  --validation_interval 5 \
  --model residual_design_model \
  --base_model pix2pix \
  --complex_model resfcn \
  --combine_mode nn \
  --loss_2 weighted_combined \
  --wc_loss_silog_lambda_2 0.5 \
  --wc_loss_weight_silog_2 1.0 \
  --wc_loss_weight_grad_2 60.0 \
  --wc_loss_weight_ssim_2 30.0 \
  --wc_loss_weight_edge_aware_2 60.0 \
  --wc_loss_weight_l1_2 6.0 \
  --resfcn_2_in_channels 1 \
  --resfcn_2_hidden_channels 64 \
  --resfcn_2_out_channels 1 \
  --resfcn_2_num_blocks 12 \
  --pix2pix_in_channels 1 \
  --pix2pix_hidden_channels 64 \
  --pix2pix_out_channels 1 \
  --pix2pix_second_loss_lambda 100 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda \
  --experiment_name image-to-image \
  --run_name residual_design_nn_test \
  --tensorboard_path ./tensorboard \
  --save_path ./mlflow_images \
  --cmap gray \
  > ./logs/residual_design_nn_test.log 2>&1 &

Windows:

start /B python ./main.py ^
  --mode train ^
  --epochs 120 ^
  --batch_size 12 ^
  --lr 0.0001 ^
  --loss weighted_combined ^
  --wc_loss_silog_lambda 0.5 ^
  --wc_loss_weight_silog 1.0 ^
  --wc_loss_weight_grad 40.0 ^
  --wc_loss_weight_ssim 80.0 ^
  --wc_loss_weight_edge_aware 40.0 ^
  --wc_loss_weight_l1 8.0 ^
  --wc_loss_weight_var 0.0 ^
  --wc_loss_weight_range 0.0 ^
  --wc_loss_weight_blur 0.0 ^
  --optimizer adam ^
  --optimizer_2 adamw ^
  --weight_decay ^
  --weight_decay_rate 0.001 ^
  --gradient_clipping ^
  --gradient_clipping_threshold 0.5 ^
  --scheduler step ^
  --scheduler_2 step ^
  --activate_amp ^
  --amp_scaler grad ^
  --checkpoint_save_dir ./checkpoints ^
  --save_only_best_model ^
  --validation_interval 5 ^
  --model residual_design_model ^
  --base_model pix2pix ^
  --complex_model resfcn ^
  --combine_mode nn ^
  --loss_2 weighted_combined ^
  --wc_loss_silog_lambda_2 0.5 ^
  --wc_loss_weight_silog_2 1.0 ^
  --wc_loss_weight_grad_2 60.0 ^
  --wc_loss_weight_ssim_2 30.0 ^
  --wc_loss_weight_edge_aware_2 60.0 ^
  --wc_loss_weight_l1_2 6.0 ^
  --resfcn_2_in_channels 1 ^
  --resfcn_2_hidden_channels 64 ^
  --resfcn_2_out_channels 1 ^
  --resfcn_2_num_blocks 12 ^
  --pix2pix_in_channels 1 ^
  --pix2pix_hidden_channels 64 ^
  --pix2pix_out_channels 1 ^
  --pix2pix_second_loss_lambda 100 ^
  --data_variation sound_reflection ^
  --input_type osm ^
  --output_type standard ^
  --device cuda ^
  --experiment_name image-to-image ^
  --run_name residual_design_nn_test ^
  --tensorboard_path ./tensorboard ^
  --save_path ./mlflow_images ^
  --cmap gray ^
  > ./logs/residual_design_nn_test.log 2>&1

PhysFormer tryout

Linux:

nohub python ./main.py \
  --mode train \
  --epochs 50 \
  --batch_size 16 \
  --lr 0.001 \
  --loss weighted_combined \
  --wc_loss_silog_lambda 0.5 \
  --wc_loss_weight_silog 0.1 \
  --wc_loss_weight_grad 5.0 \
  --wc_loss_weight_ssim 10.0 \
  --wc_loss_weight_edge_aware 5.0 \
  --wc_loss_weight_l1 1.0 \
  --wc_loss_weight_var 0.0 \
  --wc_loss_weight_range 0.0 \
  --wc_loss_weight_blur 0.0 \
  --optimizer adamw \
  --weight_decay \
  --weight_decay_rate 0.05 \
  --gradient_clipping \
  --gradient_clipping_threshold 2.0 \
  --scheduler cosine \
  --use_warm_up \
  --warm_up_start_lr 0.000005 \
  --warm_up_step_duration 2000 \
  --activate_amp \
  --amp_scaler grad \
  --checkpoint_save_dir ./checkpoints \
  --save_only_best_model \
  --validation_interval 2 \
  --model physicsformer \
  --physicsformer_in_channels 1 \
  --physicsformer_out_channels 1 \
  --physicsformer_img_size 256 \
  --physicsformer_patch_size 4 \
  --physicsformer_embedded_dim 1024 \
  --physicsformer_num_blocks 8 \
  --physicsformer_heads 16 \
  --physicsformer_mlp_dim 2048 \
  --physicsformer_dropout 0.1 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --device cuda \
  --experiment_name image-to-image \
  --run_name physformer_test \
  --tensorboard_path ./tensorboard \
  --save_path ./mlflow_images \
  --cmap gray \
  > ./logs/physformer_test.log 2>&1

Windows:

start /B python ./main.py ^
  --mode train ^
  --epochs 50 ^
  --batch_size 16 ^
  --lr 0.001 ^
  --loss weighted_combined ^
  --wc_loss_silog_lambda 0.5 ^
  --wc_loss_weight_silog 0.1 ^
  --wc_loss_weight_grad 5.0 ^
  --wc_loss_weight_ssim 10.0 ^
  --wc_loss_weight_edge_aware 5.0 ^
  --wc_loss_weight_l1 1.0 ^
  --wc_loss_weight_var 0.0 ^
  --wc_loss_weight_range 0.0 ^
  --wc_loss_weight_blur 0.0 ^
  --optimizer adamw ^
  --weight_decay ^
  --weight_decay_rate 0.05 ^
  --gradient_clipping ^
  --gradient_clipping_threshold 2.0 ^
  --scheduler cosine ^
  --use_warm_up ^
  --warm_up_start_lr 0.000005 ^
  --warm_up_step_duration 2000 ^
  --activate_amp ^
  --amp_scaler grad ^
  --checkpoint_save_dir ./checkpoints ^
  --save_only_best_model ^
  --validation_interval 2 ^
  --model physicsformer ^
  --physicsformer_in_channels 1 ^
  --physicsformer_out_channels 1 ^
  --physicsformer_img_size 256 ^
  --physicsformer_patch_size 8 ^
  --physicsformer_embedded_dim 1024 ^
  --physicsformer_num_blocks 8 ^
  --physicsformer_heads 16 ^
  --physicsformer_mlp_dim 2048 ^
  --physicsformer_dropout 0.1 ^
  --data_variation sound_reflection ^
  --input_type osm ^
  --output_type standard ^
  --device cuda ^
  --experiment_name image-to-image ^
  --run_name physicsformer_test ^
  --tensorboard_path ./tensorboard ^
  --save_path ./mlflow_images ^
  --cmap gray ^
  > ./logs/physicsformer_test.log 2>&1

Ray-Tracing Test

Linux:

nohub python ./main.py \
  --mode train \
  --epochs 120 \
  --batch_size 12 \
  --lr 0.0001 \
  --loss weighted_combined \
  --wc_loss_silog_lambda 0.5 \
  --wc_loss_weight_silog 1.0 \
  --wc_loss_weight_grad 40.0 \
  --wc_loss_weight_ssim 80.0 \
  --wc_loss_weight_edge_aware 40.0 \
  --wc_loss_weight_l1 8.0 \
  --wc_loss_weight_var 0.0 \
  --wc_loss_weight_range 0.0 \
  --wc_loss_weight_blur 0.0 \
  --optimizer adam \
  --optimizer_2 adamw \
  --weight_decay \
  --weight_decay_rate 0.0004 \
  --gradient_clipping \
  --gradient_clipping_threshold 0.5 \
  --scheduler step \
  --scheduler_2 step \
  --activate_amp \
  --amp_scaler grad \
  --checkpoint_save_dir ./checkpoints \
  --save_only_best_model \
  --validation_interval 5 \
  --model residual_design_model \
  --base_model pix2pix \
  --complex_model pix2pix \
  --combine_mode nn \
  --loss_2 weighted_combined \
  --wc_loss_silog_lambda_2 0.5 \
  --wc_loss_weight_silog_2 1.0 \
  --wc_loss_weight_grad_2 60.0 \
  --wc_loss_weight_ssim_2 30.0 \
  --wc_loss_weight_edge_aware_2 60.0 \
  --wc_loss_weight_l1_2 6.0 \
  --pix2pix_in_channels 1 \
  --pix2pix_hidden_channels 64 \
  --pix2pix_out_channels 1 \
  --pix2pix_second_loss_lambda 100 \
  --pix2pix_2_in_channels 2 \
  --pix2pix_2_hidden_channels 64 \
  --pix2pix_2_out_channels 1 \
  --pix2pix_2_second_loss_lambda 100 \
  --data_variation sound_reflection \
  --input_type osm \
  --output_type standard \
  --reflexion_channels \
  --reflexion_steps 36 \
  --device cuda \
  --experiment_name image-to-image \
  --run_name raytracing_test \
  --tensorboard_path ./tensorboard \
  --save_path ./mlflow_images \
  --cmap gray \
  > ./logs/raytracing_test.log 2>&1 &

Windows:

start /B python ./main.py ^
  --mode train ^
  --epochs 120 ^
  --batch_size 12 ^
  --lr 0.0001 ^
  --loss weighted_combined ^
  --wc_loss_silog_lambda 0.5 ^
  --wc_loss_weight_silog 1.0 ^
  --wc_loss_weight_grad 40.0 ^
  --wc_loss_weight_ssim 80.0 ^
  --wc_loss_weight_edge_aware 40.0 ^
  --wc_loss_weight_l1 8.0 ^
  --wc_loss_weight_var 0.0 ^
  --wc_loss_weight_range 0.0 ^
  --wc_loss_weight_blur 0.0 ^
  --optimizer adam ^
  --optimizer_2 adamw ^
  --weight_decay ^
  --weight_decay_rate 0.0004 ^
  --gradient_clipping ^
  --gradient_clipping_threshold 0.5 ^
  --scheduler step ^
  --scheduler_2 step ^
  --activate_amp ^
  --amp_scaler grad ^
  --checkpoint_save_dir ./checkpoints ^
  --save_only_best_model ^
  --validation_interval 5 ^
  --model residual_design_model ^
  --base_model pix2pix ^
  --complex_model pix2pix ^
  --combine_mode nn ^
  --loss_2 weighted_combined ^
  --wc_loss_silog_lambda_2 0.5 ^
  --wc_loss_weight_silog_2 1.0 ^
  --wc_loss_weight_grad_2 60.0 ^
  --wc_loss_weight_ssim_2 30.0 ^
  --wc_loss_weight_edge_aware_2 60.0 ^
  --wc_loss_weight_l1_2 6.0 ^
  --pix2pix_in_channels 1 ^
  --pix2pix_hidden_channels 64 ^
  --pix2pix_out_channels 1 ^
  --pix2pix_second_loss_lambda 100 ^
  --pix2pix_2_in_channels 2 ^
  --pix2pix_2_hidden_channels 64 ^
  --pix2pix_2_out_channels 1 ^
  --pix2pix_2_second_loss_lambda 100 ^
  --data_variation sound_reflection ^
  --input_type osm ^
  --output_type standard ^
  --reflexion_channels ^
  --reflexion_steps 36 ^
  --device cuda ^
  --experiment_name image-to-image ^
  --run_name raytracing_test ^
  --tensorboard_path ./tensorboard ^
  --save_path ./mlflow_images ^
  --cmap gray ^
  > ./logs/raytracing_test.log 2>&1 &

---

Access Experiment Tracking

In order to see the history of previous trainings you can start tensorboard of mlflow. Both start a local server which can be acces over an address which is very handy because they often are stored on another device and so can be viewed over SSH.
While tensorboard shows curves for the loss (there are some other features, but this is the most important), does mlflow shows different metrices, parameters and model-artifacts (and also optionally some inferences).

You have to start one of them:

1. Go with a terminal to the top folder:

mlflow ui

Or with tensorboard:

tensorboard --logdir your_save_dir/tensorboard

Both starts a local website:

• mlflow most likely: http://localhost:5000
• tensorboard most likely: http://localhost:6006