Support for parsing ONNX Pad node (#1352)

vloncar · JanFSchulte · web-flow · commit b493ea7fd282 · 2025-10-31T20:19:48.000-04:00
* Support for parsing ONNX Pad node

* Use dynamo onnx export

* Parse Pad node with ONNX opset &gt;= 11

* Make ONNX models manually instead of using torch's export

---------

Co-authored-by: Jan-Frederik Schulte &lt;jschulte@cern.ch&gt;
diff --git a/hls4ml/converters/onnx/reshape.py b/hls4ml/converters/onnx/reshape.py
@@ -1,4 +1,4 @@
-from hls4ml.converters.onnx_to_hls import get_onnx_attribute, onnx_handler
+from hls4ml.converters.onnx_to_hls import get_constant_value, get_onnx_attribute, onnx_handler
 
 
 @onnx_handler('Transpose')
@@ -58,3 +58,75 @@ def parse_resize_layer(node, input_names, input_shapes, graph):
     )
 
     return layer
+
+
+@onnx_handler('Pad')
+def parse_pad_layer(node, input_names, input_shapes, graph):
+    layer = {}
+    layer['name'] = node.name
+    layer['class_name'] = 'ZeroPadding'
+    layer['inputs'] = input_names
+    layer['outputs'] = list(node.output)
+    layer['data_format'] = (
+        'channels_last' if any(node.domain == 'qonnx.custom_op.channels_last' for node in graph.node) else 'channels_first'
+    )
+
+    mode = get_onnx_attribute(node, 'mode')
+    if mode is not None and mode != 'constant':
+        raise RuntimeError(f'Unsupported padding mode: {mode} in node {node.name}')
+
+    pads = get_constant_value(graph, node.input[1])
+    if len(input_names) > 2:
+        const_val = get_constant_value(graph, node.input[2])
+        if const_val != 0:
+            raise RuntimeError(f'Only constant value of 0 supported for Pad node {node.name}, got {const_val}')
+
+    if len(input_names) > 3:
+        raise RuntimeError(f'Parsing axes input of Pad node {node.name} is not supported.')
+
+    dim = 0
+    if len(input_shapes[0]) == 3:
+        dim = 1  # 2D input (batch, channels, width), will use ZeroPadding1D
+        if layer['data_format'] == 'channels_first':
+            _, channels, width = input_shapes[0]
+            pad_left, pad_right = pads[2], pads[5]
+        else:
+            _, width, channels = input_shapes[0]
+            pad_left, pad_right = pads[1], pads[4]
+        out_width = width + pad_left + pad_right
+
+        layer['n_chan'] = channels
+        layer['in_width'] = width
+        layer['out_width'] = out_width
+
+        layer['pad_left'] = pad_left
+        layer['pad_right'] = pad_right
+    elif len(input_shapes[0]) == 4:
+        dim = 2  # 3D input (batch, channels, height, width), will use ZeroPadding2D
+        if layer['data_format'] == 'channels_first':
+            _, channels, height, width = input_shapes[0]
+            pad_top, pad_bottom = pads[2], pads[6]
+            pad_left, pad_right = pads[3], pads[7]
+        else:
+            _, height, width, channels = input_shapes[0]
+            pad_top, pad_bottom = pads[1], pads[5]
+            pad_left, pad_right = pads[2], pads[6]
+        out_height = height + pad_top + pad_bottom
+        out_width = width + pad_left + pad_right
+
+        layer['n_chan'] = channels
+        layer['in_height'] = height
+        layer['in_width'] = width
+        layer['out_height'] = out_height
+        layer['out_width'] = out_width
+
+        layer['pad_top'] = pad_top
+        layer['pad_bottom'] = pad_bottom
+        layer['pad_left'] = pad_left
+        layer['pad_right'] = pad_right
+    else:
+        raise RuntimeError(f'Unsupported input shape: {input_shapes[0]} for Pad node {node.name}')
+
+    layer['class_name'] += str(dim) + 'D'
+
+    return layer
diff --git a/hls4ml/converters/pytorch/reshape.py b/hls4ml/converters/pytorch/reshape.py
@@ -207,6 +207,8 @@ def parse_constantpad2d_layer(operation, layer_name, input_names, input_shapes,
     layer['out_height'] = out_height
     layer['out_width'] = out_width
 
+    layer['data_format'] = 'channels_first'  # Default data format in PyTorch
+
     return layer, output_shape
 
 
@@ -246,4 +248,6 @@ def parse_constantpad1d_layer(operation, layer_name, input_names, input_shapes,
     layer['in_width'] = width
     layer['out_width'] = out_width
 
+    layer['data_format'] = 'channels_first'  # Default data format in PyTorch
+
     return layer, output_shape
diff --git a/hls4ml/model/optimizer/__init__.py b/hls4ml/model/optimizer/__init__.py
@@ -34,6 +34,7 @@
     'parse_qonnx',
     [
         'reshape_constant',
+        'padding_constant',
         'resize_remove_constants',
         'quant_constant_parameters',
         'bipolar_quant_constant_parameters',
diff --git a/hls4ml/model/optimizer/passes/pad_const.py b/hls4ml/model/optimizer/passes/pad_const.py
@@ -0,0 +1,37 @@
+from hls4ml.model.layers import Constant, ZeroPadding1D, ZeroPadding2D
+from hls4ml.model.optimizer import OptimizerPass
+
+
+class PaddingConstant(OptimizerPass):
+    """
+    ONNX has the padding come as an input, not a parameter. This removes the Constant node from the input.
+    The constant value was already used; this is just a cleanup uptimization.
+    """
+
+    def match(self, node):
+        is_match = (
+            isinstance(node, (ZeroPadding1D, ZeroPadding2D))
+            and len(node.inputs) > 1
+            and isinstance(node.get_input_node(node.inputs[1]), Constant)
+        )
+
+        return is_match
+
+    def transform(self, model, node):
+        """
+        Remove Constant node(s) from the graph. Note, padding is already present in ZeroPadding node.
+        """
+        if len(node.inputs) > 2:
+            const_val_node = node.get_input_node(node.inputs[2])
+            if not isinstance(const_val_node, Constant):
+                raise RuntimeError(f'Non-constant padding inputs are not currently supported ({node.name})')
+            model.remove_node(const_val_node)
+            node.inputs.pop(2)
+
+        pad_node = node.get_input_node(node.inputs[1])
+        if not isinstance(pad_node, Constant):
+            raise RuntimeError(f'Non-constant padding inputs are not currently supported ({node.name})')
+        model.remove_node(pad_node)
+        node.inputs.pop(1)
+
+        return True
diff --git a/test/pytest/test_pytorch_constpadmapping.py b/test/pytest/test_pytorch_constpadmapping.py
diff --git a/test/pytest/test_zeropadding_pytorch_onnx.py b/test/pytest/test_zeropadding_pytorch_onnx.py
@@ -0,0 +1,138 @@
+from pathlib import Path
+
+import numpy as np
+import torch.nn as nn
+from onnx import TensorProto, helper
+
+from hls4ml.converters import convert_from_onnx_model, convert_from_pytorch_model
+from hls4ml.utils.config import config_from_onnx_model, config_from_pytorch_model
+
+test_root_path = Path(__file__).parent
+
+
+def _make_constantpad_onnx_1d():
+    input_tensor = helper.make_tensor_value_info('global_in', TensorProto.FLOAT, [1, 2, 4])
+    output_tensor = helper.make_tensor_value_info('global_out', TensorProto.FLOAT, [1, 2, 9])
+    pads_tensor = helper.make_tensor_value_info('pads', TensorProto.INT64, [6])
+    value_tensor = helper.make_tensor_value_info('value', TensorProto.FLOAT, [])
+
+    # Pads = [N_before, C_before, W_before, N_after, C_after, W_after]
+    pads = [0, 0, 2, 0, 0, 3]
+
+    pads_initializer = helper.make_tensor(name='pads', data_type=TensorProto.INT64, dims=[6], vals=pads)
+    value_initializer = helper.make_tensor(name='value', data_type=TensorProto.FLOAT, dims=[], vals=[0.0])
+
+    pad_node = helper.make_node(
+        'Pad', name='const_pad', inputs=['global_in', 'pads', 'value'], outputs=['global_out'], mode='constant'
+    )
+
+    graph = helper.make_graph(
+        nodes=[pad_node],
+        name='Pad1DGraph',
+        inputs=[input_tensor],
+        outputs=[output_tensor],
+        initializer=[pads_initializer, value_initializer],
+        value_info=[pads_tensor, value_tensor],
+    )
+
+    model = helper.make_model(graph)
+
+    return model
+
+
+def test_constantpad_1d():
+    class Pad1DModel(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.pad = nn.ConstantPad1d((2, 3), 0)  # pad 2 left, 3 right
+
+        def forward(self, x):
+            return self.pad(x)
+
+    model = Pad1DModel()
+    model.eval()
+    config_pytorch = config_from_pytorch_model(model, (2, 4), channels_last_conversion='off')
+    hls_model_pytorch = convert_from_pytorch_model(
+        model, output_dir=str(test_root_path / 'hls4mlprj_constpad_1d/pytorch'), hls_config=config_pytorch
+    )
+
+    hls_model_pytorch.compile()
+
+    pad1d_onnx = _make_constantpad_onnx_1d()
+
+    config_onnx = config_from_onnx_model(pad1d_onnx)
+    hls_model_onnx = convert_from_onnx_model(
+        pad1d_onnx, output_dir=str(test_root_path / 'hls4mlprj_constpad_1d/onnx'), hls_config=config_onnx
+    )
+
+    hls_model_onnx.compile()
+
+    input_data = np.random.randn(10, 2, 4)
+    pred_pytorch = hls_model_pytorch.predict(input_data)
+    pred_onnx = hls_model_onnx.predict(input_data)
+
+    np.testing.assert_allclose(pred_pytorch, pred_onnx, rtol=0, atol=1e-5)
+
+
+def _make_constantpad_onnx_2d():
+    input_tensor = helper.make_tensor_value_info('global_in', TensorProto.FLOAT, [1, 2, 3, 4])
+    output_tensor = helper.make_tensor_value_info('global_out', TensorProto.FLOAT, [1, 2, 10, 7])
+    pads_tensor = helper.make_tensor_value_info('pads', TensorProto.INT64, [8])
+    value_tensor = helper.make_tensor_value_info('value', TensorProto.FLOAT, [])
+
+    # Pads = [N_before, C_before, H_before, W_before, N_after, C_after, H_after, W_after]
+    pads = [0, 0, 3, 1, 0, 0, 4, 2]
+
+    pads_initializer = helper.make_tensor(name='pads', data_type=TensorProto.INT64, dims=[8], vals=pads)
+    value_initializer = helper.make_tensor(name='value', data_type=TensorProto.FLOAT, dims=[], vals=[0.0])
+
+    pad_node = helper.make_node(
+        'Pad', name='const_pad', inputs=['global_in', 'pads', 'value'], outputs=['global_out'], mode='constant'
+    )
+
+    graph = helper.make_graph(
+        nodes=[pad_node],
+        name='Pad2DGraph',
+        inputs=[input_tensor],
+        outputs=[output_tensor],
+        initializer=[pads_initializer, value_initializer],
+        value_info=[pads_tensor, value_tensor],
+    )
+
+    model = helper.make_model(graph)
+
+    return model
+
+
+def test_constantpad_2d():
+    class Pad2DModel(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.pad = nn.ConstantPad2d((1, 2, 3, 4), 0)  # left, right, top, bottom
+
+        def forward(self, x):
+            return self.pad(x)
+
+    model = Pad2DModel()
+    model.eval()
+    config_pytorch = config_from_pytorch_model(model, (2, 3, 4), channels_last_conversion='off')
+    hls_model_pytorch = convert_from_pytorch_model(
+        model, output_dir=str(test_root_path / 'hls4mlprj_constpad_2d/pytorch'), hls_config=config_pytorch
+    )
+
+    hls_model_pytorch.compile()
+
+    pad2d_onnx = _make_constantpad_onnx_2d()
+
+    config_onnx = config_from_onnx_model(pad2d_onnx)
+    hls_model_onnx = convert_from_onnx_model(
+        pad2d_onnx, output_dir=str(test_root_path / 'hls4mlprj_constpad_2d/onnx'), hls_config=config_onnx
+    )
+
+    hls_model_onnx.compile()
+
+    input_data = np.random.randn(10, 2, 3, 4)
+    pred_pytorch = hls_model_pytorch.predict(input_data)
+    pred_onnx = hls_model_onnx.predict(input_data)
+
+    np.testing.assert_allclose(pred_pytorch, pred_onnx, rtol=0, atol=1e-5)

Original file line number	Diff line number	Diff line change
`@@ -34,6 +34,7 @@`
`34`	`34`	`'parse_qonnx',`
`35`	`35`	`[`
`36`	`36`	`'reshape_constant',`
	`37`	`+ 'padding_constant',`
`37`	`38`	`'resize_remove_constants',`
`38`	`39`	`'quant_constant_parameters',`
`39`	`40`	`'bipolar_quant_constant_parameters',`