Edit model card

STFPM model from Anomalib fine-tuned for capsule category of the MVTec dataset. Checkpoint trained using the following notebook.

──────────────────────────────────────────────────
       Test metric             DataLoader 0
──────────────────────────────────────────────────
       image_AUROC          0.9541285037994385
      image_F1Score         0.9680365324020386
       pixel_AUROC          0.9857622385025024
      pixel_F1Score         0.4696350395679474
──────────────────────────────────────────────────

The main intent is to use it in samples and demos for model optimization. Here is the advantages:

  • MVTec dataset can automatically downloaded and is quite small.
  • The model from the anomaly detection domain such as STFPM is sensitive to the optimization methods to allows demonstrate methods with accuracy controll.

Here is the code to test the checkpoint:

from pytorch_lightning import Trainer
from anomalib.config import get_configurable_parameters
from anomalib.data import get_datamodule
from anomalib.models import get_model
from anomalib.utils.callbacks import LoadModelCallback, get_callbacks

CHECKPOINT_URL = 'https://huggingface.co/alexsu52/stfpm_mvtec_capsule/resolve/main/pytorch_model.bin'
CHECKPOINT_PATH = '~/pytorch_model.bin'

#Download CHECKPOINT_URL to CHECKPOINT_PATH

config = get_configurable_parameters(config_path="./anomalib/models/stfpm/config.yaml")
config["dataset"]["path"] = <path_to_dataset>
config['dataset']['category'] = 'capsule'

datamodule = get_datamodule(config)
datamodule.setup()  # Downloads the dataset if it's not in the specified `root` directory
datamodule.prepare_data()  # Create train/val/test/prediction sets.

model = get_model(config)

callbacks = get_callbacks(config)
load_model_callback = LoadModelCallback(weights_path=CHECKPOINT_PATH)
callbacks.insert(0, load_model_callback)

trainer = Trainer(**config.trainer, callbacks=callbacks)
trainer.test(model=model, datamodule=datamodule)
Downloads last month
0
Unable to determine this model's library. Check the docs .