I found a very simple operation like:
scores, labels = torch.max(prob, dim=-1)
scores = scores.unsqueeze(axis=-1)
labels = labels.unsqueeze(axis=-1) * 1.0
print(scores.shape)
print(conf.shape)
print(labels.shape)
outputs = torch.cat((xy, wh, conf, scores, labels), dim=2)
Can not converted to tvm, this operation is apply a tensor.max and get scores, labels of the model.
But tvm will fail.