Pytorch programming knowledge (2)

1. num_worker（ finger dataloader Number of worker processes created ）
2. GE2E：Speaker Verification（ Verify whether the input speech belongs to a specific speaker The task of ） Loss function , End to end loss function , Fast convergence 、 Implement a simple

3. torch.gather(): From batch tensor Extract the data under the specified out of order index

4. array To tensor:torch.from_numpy()

5. A two-dimensional tensor Transposition ：tensor = torch.transpose(tensor, dim0=0, dim1=1)

6. pytorch After middle convolution, the size of the image before and after input and output is guaranteed to remain unchanged , I.e. implementation TensorFlow in “same” The effect of ：

   1. If convolution kernel size is 3x3, Then it should be set padding=1, I.e. filling 1 Layer edge pixels ;

   2. If convolution kernel size is 7x7, Then it should be set padding=3, fill 3 Layer edge pixels ;

   3. That is to say padding The size is usually set to half the size of the core （ Rounding down ）.

7. When debugging the network , Look at the tensor tensor shape shape：

   1. emb.detach().cpu().numpy()

   2. explain ：detach(): take tensor Break away from the calculation diagram ,cpu() Place in cuda Upper tensor The switch to cpu On （ Because only cpu Upper tensor Can be called numpy）, take tensor To array