patches.py

import os
import sys
import logging
import tensorflow as tf

def apply_fixes():
    # 1. Mute TF Logs
    os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
    tf.get_logger().setLevel(logging.ERROR)
    tf.autograph.set_verbosity(0)

    # 2. Force TF 1.15 behavior
    if hasattr(tf, 'compat') and hasattr(tf.compat, 'v1'):
        tf.compat.v1.disable_v2_behavior()

    # 3. Keras Bridge for Mask R-CNN
    try:
        from tensorflow.python.keras import engine as KE
    except ImportError:
        from tensorflow.python.keras.api._v1.keras import engine as KE

    import tensorflow.python.keras as keras
    sys.modules['keras'] = keras
    sys.modules['keras.engine'] = KE
    sys.modules['keras.layers'] = keras.layers
    sys.modules['keras.models'] = keras.models

def patch_model_file(model_path="mrcnn/model.py"):
    if not os.path.exists(model_path):
        return
        
    with open(model_path, 'r') as f:
        content = f.read()

    # Apply all dynamic shape fixes
    replacements = {
        'KL.Reshape((s[1], num_classes, 4), name="mrcnn_bbox")(x)': 'KL.Reshape((-1, num_classes, 4), name="mrcnn_bbox")(x)',
        'KL.Reshape((s[1], s[2], s[3], num_classes), name="mrcnn_mask")(x)': 'KL.Reshape((-1, s[2], s[3], num_classes), name="mrcnn_mask")(x)',
        'tf.range(probs.shape[0])': 'tf.range(tf.shape(probs)[0])'
    }

    for old, new in replacements.items():
        content = content.replace(old, new)

    with open(model_path, 'w') as f:
        f.write(content)
    print("TF 1.15 patches applied to mrcnn/model.py")