Fix inconsistent padding scheme between onnx and pytorch

onnx2pytorch/convert/model.py

@@ -202,7 +202,10 @@ def forward(self, *input_list, **input_dict):
                 ]
 
             in_activations = [in_act for in_act in in_activations if in_act is not None]
-
+            if node.op_type == "Pad":
+                # preprocess pad in case it is 8-d array or 6-d array
+                from onnx2pytorch.operations.pad import preprocess_pads
+                in_activations = preprocess_pads(in_activations)
             # store activations for next layer
             if isinstance(op, Loop):
                 outputs = op((self,), activations, *in_activations)

onnx2pytorch/operations/pad.py

@@ -3,6 +3,26 @@
 from onnx2pytorch.operations.base import Operator
 
 
+def preprocess_pads(in_activations):
+    """
+    If pads is 8-d array for 4d input or pads is 6-d array for 3d input.
+    Convert pads from [b1,b2,...,e1,e2,...] to [b1,e1,b2,e2,...]
+
+    """
+    input = in_activations[0]
+    pads = list(in_activations[1])
+    if len(pads)//2 == len(input.size()):
+        import torch
+        new_pads = []
+        mid_idx = len(pads)//2
+        pads.reverse()
+        for i in range(mid_idx, len(pads)):
+            new_pads.append(pads[i])
+            new_pads.append(pads[i-mid_idx])
+        in_activations[1] = torch.tensor(new_pads)
+    return in_activations
+
+
 class Pad(Operator):
     def __init__(self, mode="constant", padding=None):
         self.mode = mode