swin2_tiny failed to run forward(): RuntimeError: unflatten: Provided sizes [64, 64] don't multiply up to the size of dim 2 (64) in the input tensor.

[yqsony]

I tried with

model = DPTDepthModel(
            path=None,
            backbone="swin2t16_256",
            non_negative=True,
        )

During inference at

MiDaS/midas/backbones/utils.py

Line 72 in bdc4ed6

layer_1 = pretrained.act_postprocess1(layer_1)

it gave the error

RuntimeError: unflatten: Provided sizes [64, 64] don't multiply up to the size of dim 2 (64) in the input tensor

The input at this layer is of a shape (b, 64, 64, 96) where b is the batch size. The next operator pretrained.act_postprocess1 is a

Sequential(
  (0): Transpose()
  (1): Unflatten(dim=2, unflattened_size=torch.Size([64, 64]))
)

I don't think Unflatten(dim=2, unflattened_size=torch.Size([64, 64])) work on any of the dimensions (b, 64, 64, 96). On the other hand it seems (b, 64, 64, 96) has already been unflattened.

Did anyone tried training or inference with the swin backbones?

[NielsRogge]

Hi,

See #259 for easy inference with DPT + Swin backbone