[Audio recognizer based on GD32F350RB] 4. Use Caffe2 to build a training platform

北方

[Audio recognizer based on GD32F350RB] 4. Use Caffe2 to build a training platform [Copy link]

1. As mentioned in the previous article, if this NN model is to run, the preliminary preparation work is indeed quite large. The example selected handwritten digit recognition, which is a summary based on the work of many people. Moreover, this model is a standard example implemented in multiple deep learning engines. So, if you need to build your own model, you must first build this calculation graph, then input data for training, and then output the training set. The problem is that the progress and update of this deep computing model is much faster than learning. Before you can figure out the version, the new version has a big update. This caffe model cited in the example can be seen. The updated caffe2 is the closest to caffe, so the input data is already very different. In the specific caffe, layer and weight are independent, but in caffe2 they are combined into a new data format bolb. So after training, they must be separated before they can be used in this original data. The most popular tensorflow actually opens a fool window for users. You don't even need to define the calculation graph yourself. It directly gives a default one, which makes it too difficult to import it into low-level hardware. However, caffe2 will be integrated into pytorch1.0, so it's best to start with the use of pytorch. Which deep learning framework is better is a bit too deep a knowledge pit, so it's best to skip it if you are engaged in hardware. 2. PytorchInstall 2.1 The best and least problematic way is to install it in the anaconda environment, which will check the compatibility in advance.

1. conda install pytorch-cpu -c pytorch pip3 install torchvision

复制代码

This installs the non-GPU version of pytorch and the graphical torchvision. Entering the following feedback means that there is no problem. It can be installed successfully. The version currently displayed is still 0.4.1. Pytorch1.0 will have to wait before it is released. 2.2 Test it

1. from __future__ import print_function import torch x = torch.rand(5, 3) print(x)

复制代码

英尚微电子