AttributeError: 'NoneType' object has no attribute 'view' when training flux-lora with loss_type = "smooth_l1"

[SunWintor]

Description:

I'm experiencing an issue when training a flux model using the lora-scripts project (version 1.9.0). The training process crashes with an AttributeError stating that 'NoneType' object has no attribute 'view' when using loss_type = "smooth_l1".

Error Message:

Traceback (most recent call last): File "A:\funny\flux_train\Lora训练-测试版\lora-scripts-v1.9.0\scripts\dev\flux_train_network.py", line 565, in trainer.train(args) File "A:\funny\flux_train\Lora训练-测试版\lora-scripts-v1.9.0\scripts\dev\train_network.py", line 1192, in train loss = train_util.conditional_loss( File "A:\funny\flux_train\Lora训练-测试版\lora-scripts-v1.9.0\scripts\dev\library\train_util.py", line 5815, in conditional_loss huber_c = huber_c.view(-1, 1, 1, 1) AttributeError: 'NoneType' object has no attribute 'view'

Training Parameters:

Here are the training parameters I'm using:

model_train_type = "flux-lora" pretrained_model_name_or_path = "A:/funny/ConfyUI-aki/ComfyUI-aki-v1.3/models/unet/flux1-dev.sft" ae = "A:/funny/ConfyUI-aki/ComfyUI-aki-v1.3/models/vae/ae.safetensors" clip_l = "A:/funny/ConfyUI-aki/ComfyUI-aki-v1.3/models/clip/clip_l.safetensors" t5xxl = "A:/funny/ConfyUI-aki/ComfyUI-aki-v1.3/models/clip/t5xxl_fp16.safetensors" clip_g = "A:/funny/ConfyUI-aki/ComfyUI-aki-v1.3/models/clip/clip_g.safetensors" timestep_sampling = "shift" sigmoid_scale = 1 model_prediction_type = "raw" discrete_flow_shift = 3.158 loss_type = "smooth_l1" guidance_scale = 1 train_data_dir = "A:/funny/flux_train/Lora训练-测试版/lora-scripts-v1.9.0/train/model_v5" prior_loss_weight = 1 resolution = "1024,1024" enable_bucket = true min_bucket_reso = 256 max_bucket_reso = 2048 bucket_reso_steps = 64 bucket_no_upscale = true output_name = "model_v5" output_dir = "./output" save_model_as = "safetensors" save_precision = "bf16" save_every_n_epochs = 2 max_train_epochs = 30 train_batch_size = 1 gradient_checkpointing = true gradient_accumulation_steps = 1 network_train_unet_only = true network_train_text_encoder_only = false learning_rate = 0.0001 unet_lr = 1 text_encoder_lr = 1 lr_scheduler = "cosine" lr_warmup_steps = 0 lr_scheduler_num_cycles = 0 optimizer_type = "Prodigy" optimizer_args = [ "decouple=True", "weight_decay=0.01", "use_bias_correction=True", "d_coef=1" ] network_module = "networks.lora_flux" network_dim = 8 network_alpha = 8 log_with = "tensorboard" logging_dir = "./logs" caption_extension = ".txt" shuffle_caption = false weighted_captions = false keep_tokens = 0 seed = 21337 clip_skip = 1 mixed_precision = "bf16" fp8_base = true sdpa = true lowram = false cache_latents = true cache_latents_to_disk = true cache_text_encoder_outputs = true cache_text_encoder_outputs_to_disk = true persistent_data_loader_workers = true

Steps to Reproduce:

1. Set up the training environment using the parameters specified above.
2. Begin training .
3. The error occurs during the training process.

Investigation and Findings:

• The error indicates that huber_c is None, and the code attempts to call huber_c.view(-1, 1, 1, 1), which leads to an AttributeError.
• In scripts/dev/train_network.py at line 1187, the get_noise_pred_and_target function from scripts/dev/flux_train_network.py is called.
• The get_noise_pred_and_target function returns huber_c as None.
• At line 1191 in scripts/dev/train_network.py, the conditional_loss function from scripts/dev/library/train_util.py is called, which attempts to execute huber_c.view(-1, 1, 1, 1).
• Since huber_c is None, this results in the AttributeError.

Additional Information:

• I observed that this issue seems to occur when training flux models using the huber or smooth_l1 loss types.
• Other users in the community have not reported this issue, suggesting it might be specific to my environment or configuration.
• I have tried reinstalling the latest version of lora-scripts (v1.10.0), but the problem persists.
• I'm not deeply familiar with Python, so I'm uncertain whether the code is intended to call get_noise_pred_and_target from scripts/dev/train_network.py or if conditional_loss should be using a different module (e.g., sd-scripts/library/train_util.py or scripts/stable/library/train_util.py).
• It appears that there might be an incorrect reference or import in the code that's causing the wrong function to be called.

Environment:

• Operating System: Windows 10 (64-bit)
• Python Version: 3.12.2
• lora-scripts Version: v1.10.0
• Relevant Hardware: NVIDIA RTX 4090
• Dependencies: Up-to-date versions as per the lora-scripts installation instructions

Expected Behavior:

• The training process should proceed without errors when using loss_type = "smooth_l1" with flux models.
• huber_c should be properly initialized and not be None when required.

Actual Behavior:

• The training process fails with an AttributeError because huber_c is None and does not have a view method.

Request:

• Could you please help identify the cause of this issue?
• Is there a possible fix or workaround that I can apply?
• Do I need to adjust my training parameters or update any code references?

Thank you for your assistance!

If you need any additional information or logs, please let me know, and I'll be happy to provide them.

[bobo3313]

me too