| Data type | dtype | CPU tensor | GPU tensor | Size/bytes |
|---|---|---|---|---|
| 32-bit floating | torch.float32 or torch.float | torch.FloatTensor | torch.cuda.FloatTensor | 4 |
| 64-bit floating | torch.float64 or torch.double | torch.DoubleTensor | torch.cuda.DoubleTensor | 8 |
| 16-bit floating | torch.float16or torch.half | torch.HalfTensor | torch.cuda.HalfTensor | - |
| 8-bit integer (unsigned) | torch.uint8 | torch.ByteTensor | torch.cuda.ByteTensor | 1 |
| 8-bit integer (signed) | torch.int8 | torch.CharTensor | torch.cuda.CharTensor | - |
| 16-bit integer (signed) | torch.int16or torch.short | torch.ShortTensor | torch.cuda.ShortTensor | 2 |
| 32-bit integer (signed) | torch.int32 or torch.int | torch.IntTensor | torch.cuda.IntTensor | 4 |
| 64-bit integer (signed) | torch.int64 or torch.long | torch.LongTensor | torch.cuda.LongTensor | 8 |
以上PyTorch中的數(shù)據(jù)類型和numpy中的相對應(yīng)，占用字節(jié)大小也是一樣的
補充：pytorch tensor比較大小 數(shù)據(jù)類型要注意
a = torch.tensor([[0, 0], [0, 0]]) print(a>=0.5)
輸出
tensor([[1, 1],
[1, 1]], dtype=torch.uint8)
結(jié)果明顯不對, 分析原因是因為, a是long類型, 而0.5是float. 0.5會被轉(zhuǎn)化為 long, 變?yōu)?. 因此結(jié)果會出錯, 做出如下修改就可以得到正確答案
a = torch.tensor([[0, 0], [0, 0]]).float() print(a>=0.5)
以上為個人經(jīng)驗，希望能給大家一個參考，也希望大家多多支持腳本之家。
