Acuity support

[hawk081]

Hi,

This patch enables the support of Acuity models, please help to review the patch, any suggestions are welcome.

Acuity is an AI framework that can convert the models of other framework(such as tensorflow, tflite, caffe, onnx, etc) to Acuity models, then generate high performance code for embedded platforms.

.json file is the Acuity model file which contains the model structure, and .data file will be loaded for weights and biases, .quantize file is for quanzation parameters. the output tensors' shape of each layer will be inferred if the input shape(s) are provied in model files.

Some popular models converted by Acuity can be found here: https://verisilicon.github.io/acuity-models/

Thanks.

#657

[hawk081]

Hi Lutz,

Is there anything need to be modified ?

Thanks.

[lutzroeder]

1. There are multiple similar formats for deep learning models already. Can Acuity use an existing format instead of creating another duplicate format?
2. jsyaml should not be checked in as obfuscated code. For JSON, Protocol Buffers and FlatBuffers there are custom implementation using base.TextDecoder. A similar approach should be followed for YAML.
3. Using JSON and YAML and Pickle (3 different standards) for a single model is a fragile design. Ideally there should be no YAML dependency.

[lutzroeder]

The file is sorted alphabetically by type. This is only testing the JSON code, not the YAML and binary readers. Please add a more comprehensive sample.

[hawk081]

Thanks!

1. Current format of .json, .data and .quantize are used in our production, It is difficult to migrate to other formats.
2. We will follow the JSON to implement a YAML parser.
3. For some historical reasons, we have to use these 3 files to describe the model for now.

[lutzroeder]

Current format of .json, .data and .quantize are used in our production, It is difficult to migrate to other formats.

It will only get more difficult over time. This is an overcomplicated design with lots of moving parts. Better to fix it early on while you still can. Ideally you should consider what value the framework actually provides and if it does need yet another new file format or can just use an existing one.

[hawk081]

Thanks for your suggestions.

I will remove the code of loading .quantize and .data, let netron just handle the model structure.

[lutzroeder]

Why is this format needed? There are a dozen other deep learning formats that are more or less identical.

[hawk081]

Thanks! When we started the project at 2016, there weren't suitable formats for us. so ...