I converted the pytorch model to ONNX in this way , I dont have much idea how performance would differ in different formats(any suggestion on the same is welcome.)

import torch
from model_core import Two_Stream_Net 

# Load your model
model = Two_Stream_Net()
checkpoint = torch.load('/media/node9/hdd/face-fery-detection/src/weights_finetune_clean/checkpoint_epoch_140.pth', 
                       map_location='cpu')

# Load only the model state dict
model.load_state_dict(checkpoint['model_state_dict'])

# Set model to evaluation mode
model.eval()

# Create dummy input with 4 channels instead of 3
dummy_input = torch.randn(1, 4, 256, 256)  # Changed from 3 to 4 channels

# For ONNX export:
torch.onnx.export(
    model,
    dummy_input,
    "two_stream_net.onnx",
    export_params=True,
    opset_version=12,
    do_constant_folding=True,
    input_names=['input'],
    output_names=['output', 'features', 'attention_map'],
    dynamic_axes={'input': {0: 'batch_size'},
                  'output': {0: 'batch_size'},
                  'features': {0: 'batch_size'},
                  'attention_map': {0: 'batch_size'}}
)

Now , How can I convert it into TensorRT, TorchScript? And what is difference between trace and script in torchscript.