SA
메인 내용으로 이동

Project Pontassieve

Federation 1, Epoch 1
torch.Size([128, 3072])
tensor([ 0.0134, -0.0129, -0.0139, ..., -0.0115, -0.0179, -0.0077],
grad_fn=<SelectBackward0>)
torch.Size([128, 3072])
tensor([ 0.0136, -0.0128, -0.0141, ..., -0.0117, -0.0180, -0.0077],
grad_fn=<SelectBackward0>)
torch.Size([128, 3072])
tensor([ 0.0139, -0.0135, -0.0144, ..., -0.0114, -0.0183, -0.0081],
grad_fn=<SelectBackward0>)
torch.Size([128, 3072])
tensor([ 0.0135, -0.0128, -0.0143, ..., -0.0120, -0.0188, -0.0079],
grad_fn=<SelectBackward0>)
torch.Size([128, 3072])
tensor([ 0.0136, -0.0130, -0.0142, ..., -0.0117, -0.0182, -0.0079],
grad_fn=<SelectBackward0>)

Is it possible to have the first element updated to something 0.0136 when there are transparent pixels? - Yes. At first it randomly initializes the pixels.

→ Seems like it's the filter being not uniform throughout the dataset. Pegged it to be equal.

이 문서를 언급한 문서들