HuggingFace Integration Issues¶
Problems using HuggingFace Hub models.
HuggingFace Hub Issues¶
Model Not Found on Hub¶
Solution: Verify model name exists:
import autotimm as at # recommended alias
from autotimm import list_hf_hub_backbones
# Search for model
models = list_hf_hub_backbones(model_name="resnet", limit=10)
print(models)
Model Download is Slow¶
Explanation: Models are cached after first download. Subsequent runs are fast.
Location: Models cached in ~/.cache/huggingface/hub/
Checkpoint Loading Fails with HF Models¶
Solution: Must pass the same backbone argument:
# Save
model = ImageClassifier(backbone="hf-hub:timm/resnet50.a1_in1k", ...)
# Load - must match backbone; re-supply ignored params
loaded = ImageClassifier.load_from_checkpoint(
path,
backbone="hf-hub:timm/resnet50.a1_in1k", # Must match original
metrics=metrics, # not saved in checkpoint
)
HuggingFace Transformers Issues¶
Model Expects 'pixel_values' Keyword Argument¶
Problem: HF Transformers models need specific input format
Solution:
# :material-close: Wrong
output = model(x)
# :material-check: Correct
output = model(pixel_values=x)
RuntimeError about Trainer Attachment¶
Problem: Calling configure_optimizers() without a trainer
Solutions:
# Option 1: Attach model to trainer first
trainer.fit(model, datamodule=data)
# Option 2: Use scheduler=None for inference
model = ImageClassifier(
backbone="hf-hub:timm/resnet50.a1_in1k",
scheduler=None, # No scheduler
)
HuggingFace Hub Push Issues¶
from huggingface_hub import login
# Login to HuggingFace
login(token="your_token")
# Push model with retry
model.push_to_hub(
repo_id="username/model-name",
commit_message="Initial commit",
private=True, # Make repository private
)
# If push fails, check permissions
# https://huggingface.co/settings/tokens
Related Issues¶
- Model Loading - Checkpoint issues
- Installation - HuggingFace dependencies