ArtyZe

MultiHeadedAttention与onmt结果不对应

1

您好，非常感谢您这边的分享，我在测试turbo_transoformers的时候发现一个问题，如果Q矩阵有负数，或者mask矩阵有True的时候，turbo_transformers和onmt的输出结果就无法对应上了，请问是否我测试方法有问题？ 

bn融合训练时就没有bn了吗相当于？

1

您好，我看您代码，理解着是只要开启了bn融合后，训练时就相当于没有bn参数的更新了，就用之前的rolling mean 和variance了吗？

onnx model has no weights and bias

6

Hello, thanks for your great work! Now I can export onnx model with dcn op successfully. But my onnx model is not correct:  As my weights and bias have...

Anyone has trained voc dataset succussfully?

2

Hello, I have trained VOC dataset for more than 100 epochs, but got poor mAP like follow picture, could you give me some advice? ``` 61 0.385491111925769 +-------+-------------+---------+ | Index...

Hi， @AlexeyAB ，thanks for your great job. I can understand the weights quantization scale parameter with entropy_calibration( ) method, But in the new method, why float multiplier = 1 /...

Hello, I'm focusing on convert se-resnet model from caffe to tensorflow? But I met some issues, could you share your "post-processing"? thanks :)