tracker_run_name and tracker_project_name not working for wandb

[zhuliyi0]

wandb shows tracker_run_name as project name, and random run names.

I ran this command directly in console:

accelerate launch --mixed_precision="bf16"
--num_processes="1"
--num_machines="1"
--dynamo_backend="no"
train_sdxl.py
--pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0"
--num_train_epochs="10"
--learning_rate="8e-7"
--lr_scheduler="constant"
--seed=1234
--random_flip
--instance_data_dir="/root/autodl-tmp/Dataset/lvchengguiyuxi"
--seen_state_path="/root/autodl-tmp/simpletunerOutput/lvchengguiyuxi/training_images_seen.json"
--state_path="/root/autodl-tmp/simpletunerOutput/lvchengguiyuxi/training_state.json"
--output_dir="/root/autodl-tmp/simpletunerOutput/lvchengguiyuxi"
--train_batch="1"
--gradient_accumulation_steps="1"
--caption_dropout_probability=0.05
--validation_prompt="a photo of lower part of a lvchenguiyuxi style high-rise building, walkway, trees, bush, background building, stone sculpture on the ground"
--num_validation_images=1
--resolution="768"
--validation_resolution="1024"
--checkpointing_steps="800"
--checkpoints_total_limit="1"
--validation_steps="50"
--tracker_run_name="test"
--tracker_project_name="sptnr-sdxl-lvchengguiyuxi"
--validation_guidance="7"
--validation_guidance_rescale="0"
--use_8bit_adam
--use_ema
--enable_xformers_memory_efficient_attention
--set_grads_to_none
--freeze_encoder
--gradient_checkpointing

[bghira]

I've taken a whack at it, if you can figure out a solution, I'll include it.

[zhuliyi0]

this works:

            args.tracker_project_name,
            config=vars(public_args),
            init_kwargs={
                "wandb":
                {
                    "name": args.tracker_run_name,
                    "id": f"{args.tracker_project_name},{args.tracker_run_name}",
                    "resume": "allow",
                }
            },        

[bghira]

thanks