TypeError: forward() missing 1 required positional argument: 'img_metas'

Sorry for this basic doubt. I am trying to worl with the SWIN transformer backbone and need to do a forward pass for the model in order to achieve quantization with the help of the Tensor-RT library .I am trying to do so using: model('img_path'). However, I get the following error : TypeError: forward() missing 1 required positional argument: 'img_metas'

I am not really sure what exactly does the method exactly expects as img_metas. If I am wrong in trying to do a forward pass with this, Is there a better or a different way to do this?

[psychomugs]

Assuming all your input data is uniformly formatted, you can manually create the dict and construct your own img_metas list. I set a breakpoint in a function containing img_metas and copied one of the results.

_img_metas = {
    'ori_shape': (480, 640, 3),
    'img_shape': (480, 640, 3),
    'pad_shape': (480, 640, 3),
    'scale_factor': [1., 1., 1., 1.],
    'flip': True,
    'flip_direction': 'horizontal',
    'img_norm_cfg': {
        'mean': [123.675, 116.28 , 103.53 ],
        'std': [58.395, 57.12 , 57.375],
        'to_rgb': True
    },
    'batch_input_shape': (480, 640)
    }

img_metas = [_img_metas] * number_of_inputs

I haven’t tested if the augmentation parameters (e.g. flip, flip_direction) affect the results.

[SourabhRanade]

@dnth I am trying to use a MASK RCNN model with a SWIN transformer backbone.