1.SwinTransformer overview#Since Transformer has made breakthrough progress in NLP tasks, the industry has been trying to use Transformer in the CV field. Several previous attempts have used Transformer in the field of image classification, but these methods all face two very severe challenges, one ...

Overview#The BM1684X processor has successfully deployed the C++ code of ChatGLM2-6B, the code implementation link is: https://github.com/sophgo/ChatGLM2-TPU. This article summarizes some technical points in the process of deploying this model. First, it covers the overall operation process of ChatG ...

Overview#processing flow# The wav voice undergoes pre-processing before being sent to the encoder. The output of the encoder is then passed to both the CTC decoder and the attention decoder. The CTC decoder performs a depth-first search algorithm to find a set of candidate prediction sequences. The ...

Normally, the input of a model is the preprocessed tensor with NCHW format, but sometimes we may want to directly feed the original image to the model. In this case, we need to incorporate preprocessing operations as part of the model. We currently support these types of image input, which correspon ...

As we all know, a complete model is actually composed of a series of operators, so if we wanna make a compiler more general, supporting as many operators as possible is a must-do. So that no matter if the op is from onnx, caffe, pytorch and whatnot, we can find a corresponding op in TPU-MLIR to exp ...