Spaces:
Runtime error
Runtime error
Commit History
-
Add HF Spaces YAML frontmatter to README.md
Fix 'Missing configuration in README' error by adding required sdk: docker frontmatter block. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>12a8324 -
update: experimentV2
e205f22
-
Remove MobileNetV3 monkey-patch, proceed with 5/7 models
MobileNetV3 has Keras 3.x hard_swish serialization bug that cannot be fixed without matching exact Kaggle keras version. The monkey-patch approach causes 30min timeout on cpu-basic. Experiment proceeds with EfficientNetB0, EfficientNetV2B0, ResNet50, InceptionV3, Xception — still valid for paper comparison.2764868 -
Fix: scope Layer.__call__ patch to only activate during load_model
Previous version patched all layer calls permanently, causing 30min timeout during model warmup. Now _patch_active flag gates the patch to only during load_model, restored after.ca341b2 -
Fix: smarter Layer.__call__ patch — convert floats + fallback list wrap
ddb0cd7 -
Fix: use enable_unsafe_deserialization for MobileNetV3 loading
9cb67cf -
Fix: use tf.constant instead of keras.backend.convert_to_tensor
7e70b8e -
Fix: monkey-patch Layer.__call__ for MobileNetV3 hard_swish compat
Keras 3.x rejects float positional args in Layer.__call__ during model deserialization. MobileNetV3 hard_swish passes 3.0 as positional arg. Patch auto-converts float/int to tensor constants. Back to keras==3.8.0 which works for 5/7 models via custom_objects.95e9d6b -
Fix: use TF 2.16.2 + Keras 3.3.3 — pre-hard_swish regression
Keras 3.8+ has hard_swish serialization bug breaking MobileNetV3. Keras 3.3.3 predates the regression. custom_objects still handles quantization_config from newer-saved models.33478ca -
Fix: load MobileNetV3 weights via h5py to handle layer count mismatch
Full model .h5 has 4 top-level layers, rebuilt architecture has 88/110. Extract weights from HDF5 groups and set them by layer name.bfde56b -
Fix: rebuild MobileNetV3 architecture + load weights to bypass hard_swish bug
Keras 3.x has serialization bug where hard_swish passes 3.0 as positional arg during model reconstruction. Workaround: rebuild the MobileNetV3 architecture in code and load only the weights.7576838 -
Fix: pin keras==3.8.0 to avoid MobileNetV3 hard_swish regression in 3.12
custom_objects handles quantization_config, so keras 3.8 is sufficient. Keras 3.12 has positional arg regression breaking MobileNetV3 loading.fab976b -
Fix: add custom_objects compat layer to handle quantization_config from newer Keras
Models saved on Kaggle with Keras 3.9+ include quantization_config in Dense/Conv2D/BatchNormalization configs. Custom wrapper classes strip these unknown kwargs so load_model works on older Keras.782b71a -
Fix: bump keras>=3.8.0 to match model quantization_config format
6e5be74 -
Fix: bump tensorflow to 2.18.0 for Keras 3 model compatibility
Models were saved with Keras 3 (TF 2.16+) which uses batch_shape/optional in InputLayer config. TF 2.15 (Keras 2) does not recognize these args, causing all 7 models to fail loading at startup.98fcac5 -
Fix: add compile=False to load_model, capture load errors in /debug
6c48d7c -
Add /debug endpoint to diagnose model loading issue
279d845 -
Fix TensorFlow API usage: replace tf.h5 with tf.keras
a074855 -
Merge remote-tracking branch 'origin/main'
94265dd -
Add Hugging Face Spaces YAML metadata to README
14c68b2 -
Delete measure_latency.py
Verified signature1bd5c76 -
Merge branch 'main' of https://huggingface.co/spaces/wisamWR/latency-experiment
6a0d8bb -
Initial commit for Hugging Face latency experiment
389057d -
initial commit
Verified signatured10a223