Pytorch usage and tricks

#Import the package firstimport torchimport numpy as np

1. Creation of tensor

1.1 Method of initializing tensor

data = [[1, 2][3, 4]] # Initialize tensor with listtensor_data = torch.tensor(data)# create tensor from numpy arraydata = np.random.normal((2, 3))tensor_data = torch.tensor(data)# Create an all-one tensor of the same size as another tensordata_ones = torch.ones_like(data)# Create a random tensor of the same size as another tensordata_rand = torch.rand_like(data)# create random tensorrand_data = torch.rand([2, 3]) # where size can be an array, tuple, not a dictionary

1.2 View tensor properties

data = torch.randn((2, 3))# Determine if it is a tensordata.is_tensor() # return True or Falsedata.is_complex(). # Returns True if the data is complex, otherwise returns Falsedata.is_floating_point() # Returns True if the data is a floating point type, otherwise Falsetorch.is_nonzero(data) # Does it contain a single element tensor, and returns False if it is 0, otherwise Truetorch.numel(data) # Determine the number of all elements in the tensor

1.3 Tensor Operations

torch.zeros([H, W]) # Returns an all-zero tensortorch.range(start=0, end, step=1) # Generate continuous tensors, passing in a parameter defaults to endtorch.arange(start, end, step=1) # also generates continuous tensors# Note that the contiguous array generated by torch.range has one more element than torch.arangetorch.eye(n, m=None) # Create an n * m identity matrix, without m will create an n * n square matrixtorch.full([H, W], m) , create a H * W matrix with all m elementstorch.cat()# Concatenate tensors# E.g

I didn't think about this when I used tensorflow. I read the code today and found that torch has such an important statement, record it and learn it:

1. model.train()
Enable BatchNormalization and Dropout
2. model.eval()
Disable BatchNormalization and Dropout
After training the train samples, the generated model will beto test the sample.Before model(test), model.eval() needs to be added, otherwise as long as there is input data, even if there is no training, the model will change the weights.This is the property brought by the batch normalization layer contained in the model.
Reprinted from: https://zhuanlan.zhihu.com/p/208233193