I have a traced model that works correctly on Python and C++. However, when i try loading it in Android, I’m getting inconsistent output tensors on every run. I did call model.eval() before tracing the model, plus the same .pt file is working as intended on Python and C++.
I tried loading these pretrained models in Android and they are returning consistent output tensors on each run though. I would appreciate if anyone could provide me some insights on why my traced model is having such inconsistent output on every run. 
Considering @David_Reiss’s reply in this issue, could this behavior be a bug?
Here’s the layers to my traced model
Backbone(
original_name=Backbone
(input_layer): Sequential(
original_name=Sequential
(0): Conv2d(original_name=Conv2d)
(1): BatchNorm2d(original_name=BatchNorm2d)
(2): PReLU(original_name=PReLU)
)
(output_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Dropout(original_name=Dropout)
(2): Flatten(original_name=Flatten)
(3): Linear(original_name=Linear)
(4): BatchNorm1d(original_name=BatchNorm1d)
)
(body): Sequential(
original_name=Sequential
(0): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(1): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(2): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(3): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): Sequential(
original_name=Sequential
(0): Conv2d(original_name=Conv2d)
(1): BatchNorm2d(original_name=BatchNorm2d)
)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(4): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(5): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(6): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(7): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): Sequential(
original_name=Sequential
(0): Conv2d(original_name=Conv2d)
(1): BatchNorm2d(original_name=BatchNorm2d)
)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(8): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(9): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(10): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(11): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(12): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(13): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(14): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(15): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(16): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(17): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(18): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(19): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(20): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(21): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): Sequential(
original_name=Sequential
(0): Conv2d(original_name=Conv2d)
(1): BatchNorm2d(original_name=BatchNorm2d)
)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(22): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
(23): bottleneck_IR_SE(
original_name=bottleneck_IR_SE
(shortcut_layer): MaxPool2d(original_name=MaxPool2d)
(res_layer): Sequential(
original_name=Sequential
(0): BatchNorm2d(original_name=BatchNorm2d)
(1): Conv2d(original_name=Conv2d)
(2): PReLU(original_name=PReLU)
(3): Conv2d(original_name=Conv2d)
(4): BatchNorm2d(original_name=BatchNorm2d)
(5): SEModule(
original_name=SEModule
(avg_pool): AdaptiveAvgPool2d(original_name=AdaptiveAvgPool2d)
(fc1): Conv2d(original_name=Conv2d)
(relu): ReLU(original_name=ReLU)
(fc2): Conv2d(original_name=Conv2d)
(sigmoid): Sigmoid(original_name=Sigmoid)
)
)
)
)