unexpected keyword argument 'grad_wrt' when using TRAKer
Opened this issue · 1 comments
HiFei4869 commented
I'm following the MSCOCO demo script in the folder examples
. I got an error like this:
traker = TRAKer(
File "/home/user_name/.conda/envs/sd38/lib/python3.8/site-packages/trak/traker.py", line 177, in __init__
self.gradient_computer = gradient_computer(
TypeError: __init__() got an unexpected keyword argument 'grad_wrt'
I didn't change the initialization:
traker = TRAKer(
model=model,
task=task,
gradient_computer=DiffusionGradientComputer,
proj_dim=2048,
train_set_size=len(loader_train.dataset),
save_dir='./MSCOCO_trak_results',
device='cuda'
)
I'm using traker 0.3.2
. Can you help me locate the bug?
###############################################################
I also got this warning, and I put it here though it seems to be unrelated to the error.
The configuration file of this scheduler: DDPMScheduler {
"_class_name": "DDPMScheduler",
"_diffusers_version": "0.15.1",
"beta_end": 0.02,
"beta_schedule": "linear",
"beta_start": 0.0001,
"clip_sample": true,
"clip_sample_range": 1.0,
"dynamic_thresholding_ratio": 0.995,
"num_train_timesteps": 1000,
"prediction_type": "epsilon",
"sample_max_value": 1.0,
"thresholding": false,
"trained_betas": null,
"variance_type": "fixed_small"
}
has not set the configuration `clip_sample`. `clip_sample` should be set to False in the configuration file. Please make sure to update the config accordingly as not setting `clip_sample` in the config might lead to incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json` file
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
Thanks!
HiFei4869 commented
Looking back after several months, using traker==0.2.2
diffusers==0.15.1
makes it work, although according to the setup.py
traker 0.3.2
should also work.