Fixed keyword argument issue with Pad

onnx2pytorch/convert/attribute.py

@@ -168,7 +168,10 @@ def extract_attributes(node):
         elif attr.name == "transA":
             kwargs["transpose_activation"] = bool(extract_attr_values(attr))
         elif attr.name == "value":
-            kwargs["constant"] = extract_attr_values(attr)
+            if node.op_type == "Pad":
+                kwargs["value"] = extract_attr_values(attr)
+            else:
+                kwargs["constant"] = extract_attr_values(attr)
         elif attr.name == "value_float":
             kwargs["constant"] = extract_attr_values(attr)
         elif attr.name == "value_floats":

onnx2pytorch/convert/model.py

@@ -10,6 +10,7 @@
 from torch import nn
 from torch.jit import TracerWarning
 from torch.nn.modules.linear import Identity
+from torch.nn import Dropout
 
 from onnx2pytorch.constants import (
     COMPOSITE_LAYERS,
@@ -221,6 +222,9 @@ def forward(self, *input_list, **input_dict):
                 for out_op_id, output in zip(node.output, op(*in_activations)):
                     activations[out_op_id] = output
             else:
+                if 'dropout' in out_op_id:
+                    op = Dropout()
+
                 activations[out_op_id] = op(*in_activations)
 
             # Remove activations that are no longer needed

onnx2pytorch/operations/pad.py

@@ -4,12 +4,15 @@
 
 
 class Pad(Operator):
-    def __init__(self, mode="constant", padding=None):
+    def __init__(self, mode="constant", padding=None, value=0):
         self.mode = mode
         self.padding = padding
+        self.value = value
         super().__init__()
 
-    def forward(self, input, pads=None, value=0):
+    def forward(self, input, pads=None, value=None):
+        if value is None:
+            value = self.value
         if self.padding is not None:
             pads = self.padding
         elif pads is None:

requirements.txt

@@ -5,3 +5,4 @@ torchvision>=0.9.0
 onnx>=1.6.0
 onnxruntime>=1.5.0
 numpy>=1.18.1
+protobuf==3.20.0