Add input layer to profiling plots (fixes #404)

[siddiquifaras]

Description

Fixes #404 by including input layer distribution and precision overlay in profiling plots.

Problem: Profiling plots showed all layers except the input. The activations_keras, activations_torch, and activations_hlsmodel functions all skipped input data, and activation_types_hlsmodel had no input tick label to anchor the type overlay to.

Solution: Added two private helper functions (_normalize_input_data and _add_input_distributions) and updated four existing functions to prepend input distributions before processing layer activations. The type overlay now explicitly includes input variable precisions from model.get_input_variables().

Dependencies: None - uses existing numpy operations and follows the same vectorized pattern as existing layer processing.

Type of change

• Bug fix (non-breaking change that fixes an issue)

Implementation Details

New helpers

def _normalize_input_data(X, input_names):
    """Return list of (name, ndarray) pairs from X, respecting input_names order."""
    if isinstance(X, dict):
        return [(n, np.asarray(X[n])) for n in input_names if n in X]
    if isinstance(X, (list, tuple)):
        return [(n, np.asarray(a)) for n, a in zip(input_names, X)]
    return [(input_names[0], np.asarray(X))]


def _add_input_distributions(data, X, input_names, fmt='longform', plot='boxplot'):
    """Prepend input-data distribution entries to *data* (modified in place)."""
    for name, arr in _normalize_input_data(X, input_names):
        print(f'   {name}')
        y = arr.flatten()
        y = abs(y[y != 0])  # Same pattern as existing layer processing
        if len(y) == 0:
            print(f'Input data for {name} contains only zeros, ignoring.')
            continue
        if fmt == 'longform':
            data['x'].extend(y.tolist())
            data['weight'].extend([name] * len(y))
        elif fmt == 'summary':
            data.append(array_to_summary(y, fmt=plot))
            data[-1]['weight'] = name

Updated functions

activations_keras: Prepends input distribution by collecting InputLayer names

input_names = [layer.name for layer in model.layers if isinstance(layer, keras.layers.InputLayer)]
if input_names:
    _add_input_distributions(data, X, input_names, fmt=fmt, plot=plot)

activations_torch: Prepends input distribution before converting X to tensor

activations_hlsmodel: Prepends input distribution and handles multi-input X formats (ndarray/list/dict) for trace

activation_types_hlsmodel: Explicitly adds input variable precisions with O(1) duplicate checking

input_var_names = set()
for input_var in model.get_input_variables():
    T = input_var.type.precision
    W, I, F, S = ap_fixed_WIFS(T)
    data['layer'].append(input_var.name)
    data['low'].append(-F)
    data['high'].append(I - 1 if S else I)
    input_var_names.add(input_var.name)

for layer in model.get_layers():
    if layer.name in input_var_names:
        continue  # Skip duplicates
    # ... rest unchanged

Why the ordering works

boxplot(fmt='summary') calls data.reverse() before plotting. Input entries are prepended first, so after reversal they appear at the top (highest y-position). Type overlay matches by layer name, so both the distribution tick and precision rectangle align automatically.

Tests

New tests (test/pytest/test_profiling_input_layer.py):

• Unit tests for _normalize_input_data helper (single input, multi-input list, multi-input dict)
• Integration tests for activations_keras (summary and longform formats)
• Integration tests for activation_types_hlsmodel (presence and valid precision bounds)

Test execution:

# New tests (8 tests)
pytest test/pytest/test_profiling_input_layer.py -v
# Result: 8 passed in 2.38s

# Existing profiling tests (32 tests)
pytest test/pytest/test_keras_v3_profiling.py test/pytest/test_pytorch_profiler.py -v
# Result: 32 passed, 1 skipped (pre-existing), 0 regressions

Test Configuration:

• Python 3.12.12
• pytest 9.0.2
• keras 3.10+
• torch 2.10.0
• tensorflow 2.19.1
• All tests run with -p no:randomly to avoid unrelated qonnx seed issues

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.

Additional Notes

• No breaking changes: All public function signatures unchanged
• No new dependencies: Uses existing numpy, pandas, matplotlib
• Performance: O(n) vectorized operations, O(1) duplicate checking with set
• Compatibility: Works for Keras, PyTorch, and HLS ModelGraph; handles single-input and multi-input models
• Pre-commit: All 17 hooks passed (ruff, ruff-format, flake8, pyupgrade, trailing whitespace, etc.)