hacnho/keras-compile-config-callback-poc

Keras compile_config learning_rate callback PoC

This repository contains bounded proof-of-concept .keras archives for the already-open Keras Native compile_config learning_rate callbacks root-cause family.

These are not new-report artifacts by themselves. They are intended to strengthen the already-open bounty:

• https://huntr.com/bounties/b6ae18b3-51bf-4607-8366-58f3448f5911

Files

• MobileNetV3Large.patched.keras
• fashion_mnist_load_data.patched.keras
• reuters_get_word_index.patched.keras
• reproduce_compile_callback_variants.py
• verify_remote_poc.py
• summary.json
• app-probe-summary.json
• SHA256SUMS.txt
• requirements.txt

What the files do

The patched .keras archives keep the same general root cause:

• safe_mode=True load succeeds
• a later benign fit(...) path reaches a deserialized compile_config.optimizer.config.learning_rate callable

Variant 1: keras.datasets.fashion_mnist.load_data

Observed side effect during later benign fit(...):

• writes four files under datasets/fashion-mnist/

Variant 2: keras.datasets.reuters.get_word_index

Observed side effect during later benign fit(...):

• writes datasets/reuters_word_index.json

Strongest current application-constructor variant

Local strengthening evidence also includes:

• MobileNetV3Large.patched.keras
• benign fit(...) triggers: keras.applications.MobileNetV3Large
• observed side effect: models/weights_mobilenet_v3_large_224_1.0_float.h5

See:

• app-probe-summary.json

Reproduce

python3 -m venv /tmp/keras-compile-callback-poc
. /tmp/keras-compile-callback-poc/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
python reproduce_compile_callback_variants.py

Expected:

• both .keras files load with safe_mode=True
• later benign fit(...) triggers the configured callable
• fashion_mnist writes four gzip members
• reuters_get_word_index writes one JSON file

Verify the public HF artifacts

You can also verify the exact public artifacts after unauthenticated download:

python verify_remote_poc.py

Expected:

• all three patched .keras files load with safe_mode=True
• later benign fit(...) reaches the configured callable
• MobileNetV3Large writes: models/weights_mobilenet_v3_large_224_1.0_float.h5
• fashion_mnist writes four gzip members
• reuters_get_word_index writes one JSON file
• all three variants fail later benign fit(...) as non-scalar learning-rate callbacks

Safety note

This is a bounded research PoC:

• no ACE claim
• no arbitrary file path control
• only Keras dataset-helper writes inside an isolated KERAS_HOME
• intended to strengthen an existing Keras MFV bounty, not to create a fresh duplicate report