daquexian

random_tensor() returns [-1, 1) for float and [-10000, 10001) for int

17

修复 random_tensor() 默认范围没有负数的问题，修复 random_tensor(dtype=int64) 只返回 0 的问题

DTR

1

基于 add_dtr_feature 分支，仅供 comment 使用

Could you please share your training configuration file?

1

Thanks for your sharing your model. I also trained mobilenetv2 by myself. But unfortunately I can only get 70.9% top-1 on imagenet 2012. I will be very appreciated if you...

fix https://github.com/onnx/optimizer/issues/59

expand and unsqueeze op are sometimes used to do broadcasting, and they can be fused if they are before an op with built-in broadcast feature (like add/mul op)

[Feature] Separate graph rewriting and constant folding

2

For op fusion (like the fusion of conv and bn), we have implemented a "small onnxruntime" in [tensor.h](https://github.com/onnx/onnx/blob/master/onnx/common/tensor.h#L174-L368). It increases the workload (the more fusion we want to do, the...

Operator requests

4

Currently dabnn only support primitive operators, such as Conv, FullyConnected, Pool, Add, Concat and ReLU. While they are already enough for many networks (e.g., ResNet, SqueezeNet), some more operators may...

Use ncnn low level api instead of eigen for float convolutions

1

Eigen is not the optimal choice for deep learning on mobile. It lacks optimization for convolutions (like winograd) and also not focus on mobile devices.

ONNX operators needed to be supported

11

This issue is to summarize all operators which needed to be supported. If you want an operator to be supported, please feel free to post it here. - [x] LRN(#16,...