Question Barracuda has different output, is there any operation unsupported?

this is the pytorch model and I upload the onnx model. Is there any operation unsupported?

Code (Python):

1. import torch
2. import torch.nn as nn
3.  
4. # Custom BI-LSTM, because unity barracuda doesn't support "bidirectional = True"
5. class BILSTM(nn.Module):
6.     def __init__(self, inputSize, hiddenSize, numLayers, dropOut):
7.         super(BILSTM, self).__init__()
8.         self.biLayer1 = nn.LSTM(input_size=inputSize,hidden_size=hiddenSize, num_layers=numLayers, batch_first=True, dropout=dropOut).cuda()
9.         self.biLayer2 = nn.LSTM(input_size=inputSize,hidden_size=hiddenSize, num_layers=numLayers, batch_first=True, dropout=dropOut).cuda()
10.     def forward(self, x):
11.         out1, (hidden1, _) = self.biLayer1(x)
12.         out2, (hidden2, _) = self.biLayer2(torch.flip(x, dims=[1]))
13.         out2 = torch.flip(out2,dims=[1])
14.         hidden = torch.cat([hidden1, hidden2], dim = 0)
15.         return (out1,out2), (hidden, 0)
16.  
17. class SimNN(nn.Module):
18.     def __init__(self, inputSize, hiddenSize, numLayers):
19.         super(SimNN, self).__init__()
20.         self.BILSTM = BILSTM(inputSize*2, hiddenSize, numLayers, 0.5)
21.         self.classifyLayer = nn.Linear(hiddenSize, 2)
22.         self.dropOut = nn.Dropout(p=0.2)
23.  
24.     def forward(self, x):
25.         xL = x[:,0,:,:]
26.         xR = x[:,1,:,:]
27.         x = torch.cat([xL, xR], dim = 2)
28.         _, (h_n, c_n) = self.BILSTM.forward(x)
29.         out = h_n[3]
30.         out = self.classifyLayer(out)
31.         out = self.dropOut(out)
32.         return out

 

I switch to use CNN, it seems LSTM is not supported in this situation