What are the common pits in pytorch?
This article mainly shows you "what are the common pits in pytorch". The content is simple and clear. I hope it can help you solve your doubts. Let the editor lead you to study and learn this article "what are the common pits in pytorch?"
1.pytorch transforms tensorx=np.random.randint (1010 100, (1010 10) 10) x=TF.to_tensor (x) print (x)
This function automatically normalizes the input data. For example, sometimes if we need to convert 0-255images to numpy-type data, it will automatically change to between 0-1.
The difference between 2.stack and cat
Stack
X=torch.randn (1) z.shape 2)) y=torch.randn (1) 2)) z=torch.stack (x) # default dim=0print (2) # torch.Size ([2, 1)
So the data after stack is easy to understand, z [0J.] The data is xQuery z [1BI..] The data is y.
Cat
Z=torch.cat ((xQuery y)) print (z.size ()) # torch.Size ([2,2,3])
After cat, the data z [0jcjjjg:] is the value of x, and z [1je jade:] is the value of y.
The most important thing is that the size of the data after stack will have an extra dimension, while cat will not. There is a very simple example to illustrate, for example, to train a detection model, label is some marked points, eg: [x1memy1memx2jiny2]
If you add batchsize to the network, then Size: [batchsize,4], if I already have two piles of data, data1:Size [128jue 4] and data2:Size [128jue 4], if you need to put the two data together, then the target data:Size [256jin4].
Obviously what we need to do is: torch.cat ((data1,data2))
If our data is like this: there are 100 label, each label is put into a list (data), [X1Magi y1rect x2Magi y2], [x1Magi y1memx2Ji y2],.] Where data is a list with a length of 100, and each element in list is a tag for a picture, and size is [4]. We need to put them together into a Size: [100jue 4] data.
Obviously what we are going to do is torch.stack (data). And the input parameter of torch.stack is of type list!
Add: cat, stack, tranpose, permute, unsqeeze in pytorch
Transformation operations commonly used for tensor are provided in pytorch.
Cat connection
Splice the data along a certain dimension. The total dimension of the data remains unchanged after cat.
For example, the following code splices two two-dimensional tensor (2-dimensional 3-camera-1-dimensional 3), and after splicing, it becomes 3-dimensional 3-dimensional tensor or 2-dimensional tensor.
The code is as follows:
Import torchtorch.manual_seed (1) x = torch.randn (2jue 3) y = torch.randn (1pr 3) print (x Magi y)
Results:
0.6614 0.2669 0.0617
0.6213-0.4519-0.1661
[torch.FloatTensor of size 2x3]
-1.5228 0.3817-1.0276
[torch.FloatTensor of size 1x3]
Put the two tensor together:
Torch.cat ((XBI y), 0)
Results:
0.6614 0.2669 0.0617
0.6213-0.4519-0.1661
-1.5228 0.3817-1.0276
[torch.FloatTensor of size 3x3]
More flexible spelling:
Torch.manual_seed (1) x = torch.randn (2Magne3) print (x) print (torch.cat ((xmemx), 0)) print (torch.cat ((xmemx), 1))
Result
/ / x
0.6614 0.2669 0.0617
0.6213-0.4519-0.1661
[torch.FloatTensor of size 2x3]
/ / torch.cat ((xPowerx), 0)
0.6614 0.2669 0.0617
0.6213-0.4519-0.1661
0.6614 0.2669 0.0617
0.6213-0.4519-0.1661
[torch.FloatTensor of size 4x3]
/ / torch.cat ((xPowerx), 1)
0.6614 0.2669 0.0617 0.6614 0.2669 0.0617
0.6213-0.4519-0.1661 0.6213-0.4519-0.1661
[torch.FloatTensor of size 2x6]
Stack, adding new dimensions for stacking
On the other hand, stack adds new dimensions.
If two 1D tensor are stack on the 0th dimension, then the tensor; of 2D2will become stack on the first dimension, and it will become tensor of 2D2V2.
See the code:
A = torch.ones ([1mai 2]) b = torch.ones ([1jue 2]) c = torch.stack ([arecine b], 0) / / the 0th dimension stack
Output:
(0,.) =
1 1
(1.) =
1 1
[torch.FloatTensor of size 2x1x2]
C = torch.stack ([aforme b], 1) / / 1st dimension stack
Output:
(0,.) =
1 1
1 1
[torch.FloatTensor of size 1x2x2]
Transpose, the two dimensions are interchangeable
The code is as follows:
Torch.manual_seed (1) x = torch.randn (2pm 3) print (x)
The original result of x:
0.6614 0.2669 0.0617
0.6213-0.4519-0.1661
[torch.FloatTensor of size 2x3]
Interchange the dimensions of x
X.transpose (0Phone1)
Result
0.6614 0.6213
0.2669-0.4519
0.0617-0.1661
[torch.FloatTensor of size 3x2]
Permute, multi-dimension interchange, more flexible transpose
Permute is a more flexible transpose, which can flexibly change the dimensions of the original data, while the data itself remains unchanged.
The code is as follows:
X = torch.randn (2 → 3 2print 4) print (x.size ()) x quop = x.permute (1 → 0) # change the original first dimension to 0 dimension, similarly, 0 meme 1 prime2 → (x_p.size ())
Results:
Torch.Size ([2,3,4])
Torch.Size ([3,2,4])
Squeeze and unsqueeze
It is often used to increase or decrease dimensions. For example, if there is no batch dimension, increase the batch dimension to 1.
Squeeze (dim_n) compression reduces the dim_n dimension, that is, removes the dim_n dimension with a number of elements of 1.
Unsqueeze (dim_n), increase the dim_n dimension, the number of elements is 1.
The above code:
# define the tensor import torchb = torch.Tensor (2Magne1) b.shapeOut [28]: torch.Size ([2Magne1]) # without parameters, remove all dimensions with one number of elements, b _ = b.squeeze () b_.shapeOut [30]: torch.Size ([2]) # plus parameters, remove the element of the first dimension as 1, and it will not work Because there are two elements in the first dimension, b _ = b.squeeze (0) b_.shapeOut [32]: torch.Size ([2,1]) # so it's OK, b _ = b.squeeze (1) b_.shapeOut [34]: torch.Size ([2]) # add a dimension b _ = b.unsqueeze (2) b_.shapeOut [36]: torch.Size ([2,1,1]) above is all the content of the article "what are the common pits in pytorch?" Thank you for reading! I believe we all have a certain understanding, hope to share the content to help you, if you want to learn more knowledge, welcome to follow the industry information channel!
