Solution: various error reports and pit stepping and pit avoidance records encountered in the alchemist cultivation plan pytoch+deeplearning (II)

problem 1：pytorch Next , To multidimensional tensor Zoom

Question why ： Sometimes you want to make the size of the image too large or too small , Want to generate tensor The tensor is scaled , At the same time, keep the same proportion of the original image . If there is one now tensor, His size is ：[64,3,256,256],64 representative batch size,3 Indicates the number of channels ,256 It stands for length and width , Want to narrow it down to 32*32 The size of the .

Solution ： We can use transform.resize(), But after trying , Discover the use of torch.nn.functional.interpolate It will work better

First, explain its parameters ：torch.nn.functional.interpolate(input, size=None, scale_factor=None, mode='nearest', align_corners=None)

1. input Is our input tensor (Tensor)
2. size It's the output size
3. scale_factor (float or Tuple[float]) Specify how many times the output is the input
4. mode It is the choice of up sampling algorithm ： Yes ’nearest’, ‘linear’, ‘bilinear’, ‘bicubic’ , ‘trilinear’ and ’area’, The default is ’nearest’
5. align_corners (bool, optional) If set to True, Then the input and output tensors are aligned by the center point of their corner pixels , This preserves the value at the corner pixels . If set to False, Then the input and output tensors are aligned by the corners of their corner pixels , Interpolation uses the boundary value of the value outside the boundary to fill

Use samples ：

print(train_mask.size())
# train_mask The size was originally [64,3,256,256]
train_mask = torch.nn.functional.interpolate(train_mask, scale_factor=1 / 4, mode='bilinear',align_corners=False)
# After the change ,train_mask The size changes to [64,3,64,64]
print(train_mask.size())

problem 2：‘PIL.Image’ has no attribute ‘****’

Question why ： This is because the system has been installed PIL It's too new , Some properties have been deprecated in the new version
Solution ： Just change back to the original version ：pip install pillow==4.1.1, Install the old version pillow, Finally, I will tell you that this version is out of date , Never mind .

problem 3：ImportError: cannot import name ‘imread’

Question why ：“scipy” The version of the library is too high
Solution ： Change back to the original version ：pip install scipy==1.2.1

problem 4：models/bird/netG.pth is a zip archive (did you mean to use torch.jit.load()?)

Question why ： Where the parameters were saved after training pytorch Environment and now loading parameters pytorch The environment may be different , Version inconsistency , As a result, the parameters saved at that time cannot be read now .

Solution ： First in 1.x Loading models under version , Then save as non zip Format , Setting the use_new_zipfile_serialization=False That's it .

#torch_version==1.x
import torch
from models import net
checkpoint = 'xxx.pth'

model = net()
model.load_state_dict(torch.load(checkpoint))
model.eval()

torch.save(model.state_dict(), model_path, use_new_zipfile_serialization=False)

problem 5： Model saving and loading

Problem description ： In actual operation , Pay attention to every 100 round epoch Or every time 50 round epoch Save the trained parameters , provide against any misfortune （ power failure 、 Disconnection 、 Hardware failure 、 Earthquake, fire, etc ）, In this way, this round can be loaded directly next time epoch The parameters of are then trained , You don't have to start over .

Solution ：
Save parameters ：

torch.save(model.state_dict(), path)

Loading of parameters

model.load_state_dict(torch.load(path))

problem 6：UnicodeDecodeError: ‘gbk’ codec can’t decode byte 0x80 in position 226: illegal multibyte sequence

Question why ： A file is gbk code , It cannot be decoded .

Solution ： See if you are yml Chinese is written in the document or other documents （ As in the figure below I accidentally wrote Chinese in the notes , Cause decoding failure ）, Just delete it