[Bug] A list of missing op conversion (Swin Transformer)

[Kyrie-Zhao]

I am currently accelerating swin_transformer (ICCV2021 best paper)
https://github.com/Kyrie-Zhao/Swin-Transformer/blob/main/models/swin_transformer.py

which shows:

NotImplementedError: The following operators are not implemented: ['aten::rand', 'aten::adaptive_avg_pool1d', 'aten::roll', 'aten::floor_']

BTW, anyone can provide some guidance regarding the procedure of implementing & registering an op by yourself? Appreciate it.

[masahi]

Can you show your script to trace the model? I got an error during tracing.

BTW, anyone can provide some guidance regarding the procedure of implementing & registering an op by yourself?

By this, do you mean you want to work on this yourself, instead of asking us to do it?

[Kyrie-Zhao]

Can you show your script to trace the model? I got an error during tracing.

BTW, anyone can provide some guidance regarding the procedure of implementing & registering an op by yourself?

By this, do you mean you want to work on this yourself, instead of asking us to do it?

Here are the codes:
...
from swin_transformer import SwinTransformer

net = SwinTransformer()
net = net.cuda()

img = torch.randn(1, 3, 224, 224)
device = torch.device("cuda")
img = img.to(device)

scripted_model = torch.jit.trace(net, img).eval()
input_name = "swin_input"
shape_list = [(input_name, img.shape)]
swin_mod, swin_params = relay.frontend.from_pytorch(scripted_model, shape_list)

---

Nah, I am just curious about the procedure of this implementation (to help me learn more about TVM). But the answer can be yes if it is not that complicated :D