Pytorch中torch.flatten()和torch.nn.Flatten()實例詳解
torch.flatten(x)等於torch.flatten(x,0)默認將張量拉成一維的向量,也就是說從第一維開始平坦化,torch.flatten(x,1)代表從第二維開始平坦化。
import torch x=torch.randn(2,4,2) print(x) z=torch.flatten(x) print(z) w=torch.flatten(x,1) print(w) 輸出為: tensor([[[-0.9814, 0.8251], [ 0.8197, -1.0426], [-0.8185, -1.3367], [-0.6293, 0.6714]], [[-0.5973, -0.0944], [ 0.3720, 0.0672], [ 0.2681, 1.8025], [-0.0606, 0.4855]]]) tensor([-0.9814, 0.8251, 0.8197, -1.0426, -0.8185, -1.3367, -0.6293, 0.6714, -0.5973, -0.0944, 0.3720, 0.0672, 0.2681, 1.8025, -0.0606, 0.4855]) tensor([[-0.9814, 0.8251, 0.8197, -1.0426, -0.8185, -1.3367, -0.6293, 0.6714] , [-0.5973, -0.0944, 0.3720, 0.0672, 0.2681, 1.8025, -0.0606, 0.4855] ])
torch.flatten(x,0,1)代表在第一維和第二維之間平坦化
import torch x=torch.randn(2,4,2) print(x) w=torch.flatten(x,0,1) #第一維長度2,第二維長度為4,平坦化後長度為2*4 print(w.shape) print(w) 輸出為: tensor([[[-0.5523, -0.1132], [-2.2659, -0.0316], [ 0.1372, -0.8486], [-0.3593, -0.2622]], [[-0.9130, 1.0038], [-0.3996, 0.4934], [ 1.7269, 0.8215], [ 0.1207, -0.9590]]]) torch.Size([8, 2]) tensor([[-0.5523, -0.1132], [-2.2659, -0.0316], [ 0.1372, -0.8486], [-0.3593, -0.2622], [-0.9130, 1.0038], [-0.3996, 0.4934], [ 1.7269, 0.8215], [ 0.1207, -0.9590]])
對於torch.nn.Flatten(),因為其被用在神經網絡中,輸入為一批數據,第一維為batch,通常要把一個數據拉成一維,而不是將一批數據拉為一維。所以torch.nn.Flatten()默認從第二維開始平坦化。
import torch #隨機32個通道為1的5*5的圖 x=torch.randn(32,1,5,5) model=torch.nn.Sequential( #輸入通道為1,輸出通道為6,3*3的卷積核,步長為1,padding=1 torch.nn.Conv2d(1,6,3,1,1), torch.nn.Flatten() ) output=model(x) print(output.shape) # 6*(7-3+1)*(7-3+1) 輸出為: torch.Size([32, 150])
總結
到此這篇關於Pytorch中torch.flatten()和torch.nn.Flatten()的文章就介紹到這瞭,更多相關Pytorch torch.flatten()和torch.nn.Flatten()內容請搜索WalkonNet以前的文章或繼續瀏覽下面的相關文章希望大傢以後多多支持WalkonNet!
推薦閱讀:
- pytorch中nn.Flatten()函數詳解及示例
- Python深度學習pytorch卷積神經網絡LeNet
- 支持PyTorch的einops張量操作神器用法示例詳解
- Python深度學習pytorch神經網絡塊的網絡之VGG
- Python編程pytorch深度卷積神經網絡AlexNet詳解