During the outbreak, I participated in the deep learning course of Baidu Flying Paddle. The earliest open source domestic deep learning framework, paddlepaddle has been updated to version 1.7, but currently it does not support CUDA 10.2, but some free deep learning courses are really good, easy to understand, full of dry goods, and GPU computing cards on ai studio Tesla V100).
Image recognition basic framework
Measurement space ---------> feature space ---------> category space,
whether it is scene recognition, target recognition or face recognition through deep learning, leave the learning of features, start with The measurement space gets the feature representation, and then the category feature matching is performed to get the result.
The specific operations are basically the following steps:
- Model structures , the number of layers of neural network design, and the number of each element, bearing in mind the need to activate a function to a linear function -> nonlinear function, are common
sigmoid
,relu
the output layer is usedsoftmax
; - The loss function is generally cross entropy;
- Parameter learning generally chooses gradient descent or back propagation;
- Brand new KaTeX mathematical formula grammar;
- Added support for the Mermaid Syntax 1 function of Gantt charts ;
- Increased multi-screen editing Markdown article features;
- Increased focus writing mode, a preview mode, simple writing mode, the left and right area of the synchronous roller set functions, function buttons in the middle of the editing area and a preview area;
- Increase the checklist function.
PaddleHub
PaddleHub is a pre-training model management and migration learning tool for flying paddles.
Installation:
pip install paddlehub
Installation model:
hub install [modelname]
View installed models:
hub list
View the properties of the locally installed model, including its name, version, and description:
hub show [modelname]
Retrieve matching models on the server side by keywords:
hub search ssd
The most powerful feature is the one-click model transfer service:
only a simple command can complete the HTTP service deployment of the model
hub serving start -m resnet_v2_50_imagenet -p 8867
The interface url format is: http://0.0.0.0:8866/predict// Among them, it is text or image, which corresponds to the model type and is the model name.