Missing `serving_default` Signature in Google CXR Foundation ELIXR-C-v2 TensorFlow Model

🤗Hub
1

Hi Hugging Face team,

I encountered an issue while trying to use the ELIXR-C-v2-pooled model from the Google CXR Foundation hosted on Hugging Face (google/cxr-foundation).

Problem Description

According to the documentation and example notebooks (from Hugging Face and Google), the TensorFlow SavedModel should include a serving_default signature for inference. However, after downloading the model using huggingface_hub.snapshot_download, the TensorFlow SavedModel does not contain any predefined signatures.

–_

Steps to Reproduce

  1. Download the model using the provided repository and the huggingface_hub library:
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="google/cxr-foundation",
    local_dir="./saved_model",
    allow_patterns=["elixr-c-v2-pooled/**"]
)
  1. Load the model in TensorFlow and check its signatures:
import tensorflow as tf

model_path = "./saved_model/elixr-c-v2-pooled"
model = tf.saved_model.load(model_path)
print("Available model signatures:", model.signatures)
print("Keys:", model.signatures.keys())
  1. Output shows no available signatures:
Available model signatures: _SignatureMap({})
KeysView(_SignatureMap({}))

Expected Behavior

The model should include the serving_default signature, allowing inference as shown in the official example notebooks.

Actual Behavior

The model does not have any predefined signatures, making it impossible to invoke for inference using the example code.

System Details

  • Environment: Lightning.ai with NVIDIA L4 GPU
  • TensorFlow Version: 2.18.0
  • CUDA Version: 12.2
  • Python Version: 3.10
  • Hugging Face Library Version: Latest (pip install huggingface_hub)

Proposed Solution

If the missing signature is due to an incomplete export or versioning issue:

  1. Re-export the model with proper signature definitions (serving_default).
  2. Update the hosted model to include this signature.

Alternatively, if the serving_default signature is deprecated for this model version:

  1. Provide documentation on how to use the model without predefined signatures.
  2. Include examples showing the correct methods or attributes to use for inference.

Additional Notes

  • I followed the official instructions in the Hugging Face documentation for Google CXR Foundation and tested with both manual downloads and automated pipelines.
  • The missing signature prevents inference as described in the example notebooks.
  • I am happy to test any suggestions or updated versions of the model.

Thank you for your help! Please let me know if additional details are needed.

1 Like