Conceptual understanding of deep learning (notes)

tensor： tensor , It can be understood as a matrix

torch.rand(batsh_size,chanel,w,h)

batsh_size: The quantity that can be handled at one time

channel： The channel number , For example, the general image is RGB Three channels , Grayscale images have only grayscale values , It's a single channel 1

w,h: Length and width of tensor

nn.moudule: neural network

contains forward Class function ,

Contains a convolution layer , Pool layer, etc .

Convolution layer ：

Explanation from Baidu Encyclopedia ：

Each convolution layer in convolution neural network （Convolutional layer） It consists of several convolution units , Of each convolution unit Parameters It's all through Back propagation algorithm The result of optimization . The purpose of convolution is to extract different features of input , The first convolution layer may only extract some low-level features such as edges 、 Lines and corners, etc , More layers of networks can iteratively extract more complex features from low-level features .

effect : Similar to a filter （ Frequency selection , Noise reduction ）, Extraction of image features .

Convolution kernel ： The initialization value automatically given by the system , The size of convolution kernel is set by oneself , Accept tuples or integer input , such as 3x3 Convolution kernel , The operation of convolution is to multiply the values of the convolution kernel one by one with the input values and then accumulate , To get the output , The length and width of the output depend not only on the input and convolution kernel, but also on step（ step ） influence , Step size refers to the number of steps that the convolution kernel moves down the right phase when performing convolution operation .

Pooling layer ：

effect ： abstract , Dimension reduction , It's like putting 1080p Video becomes 720p, Reduce training time .

Linear layer ：

effect ： Transform feature dimensions

Nonlinear transformation ： Introduce nonlinear characteristics into the network , More nonlinearity , To train a model for fitting various curves .

according to loss The gradient is obtained for back propagation optimization