Limits of Keras load_weights skip_mismatch for rollback after a schema change

Direct answer: skip_mismatch=True has no documented way to enumerate skipped layers. Keras silently ignores any layer whose name or shape doesn't match the checkpoint, leaving it in its freshly initialized (usually random) state. Manual comparison of layer names and shapes before and after the load is the only reliable audit.

Auditing which layers were skipped

Since the load returns nothing, snapshot the model's weights first and diff afterward:

initial = {layer.name: layer.get_weights() for layer in model.layers}

model.load_weights('old_weights.h5', by_name=True, skip_mismatch=True)

for layer in model.layers:
    before = initial.get(layer.name)
    after = layer.get_weights()
    if before is None:
        print(f"{layer.name}: NEW layer, untouched (random init)")
    elif all((b == a).all() for b, a in zip(before, after)):
        print(f"{layer.name}: SKIPPED or unchanged")
    else:
        print(f"{layer.name}: weights loaded")

Note the caveat: "unchanged" can also mean the checkpoint genuinely contained identical values, so treat the audit as a name/shape cross-check, not proof of a skip.

Does by_name=True match correctly?

Matching is strictly by the layer's name attribute:

• Reordering is safe — a layer keeps its weights even if its position in the model changes.
• Renaming breaks the match — a renamed layer is skipped even with identical shapes.
• Name collisions (e.g., relying on auto-incremented names like dense_1 across different schema versions) can cause weights to land on the wrong layer. Assign explicit, stable names to every layer you intend to reuse.

Rollback: config + weights vs full model save

model.get_config() captures architecture metadata only — no weights. It is a legitimate schema-versioning tool, but a rollback-safe workflow needs both pieces:

1. Save the architecture (model.to_json() or the config) to record the schema version.
2. Save weights separately with model.save_weights().
3. To roll back, rebuild the model from the saved config, then call load_weights() — with matching names and shapes, no skip_mismatch is needed.

Full model saving (model.save(), SavedModel format) bundles architecture and weights and is the most reliable rollback path, but it prevents partial reuse into a modified architecture. A practical pattern: keep full saves per version for rollback, and use by_name=True, skip_mismatch=True only for forward migration — always followed by the audit above.

Assumption: TensorFlow 2.x / tf.keras; skip_mismatch is unavailable in older standalone Keras. Verify exact behavior against your installed version, as weight-file formats (H5 vs .weights.h5) differ across releases.