Yin Guobing

我看了下官方文档：https://coral.ai/docs/edgetpu/models-intro/#model-requirements 使用EdgeTPU需要模型满足如下条件： - Tensor parameters are quantized (8-bit fixed-point numbers; int8 or uint8). - Tensor sizes are constant at compile-time (no dynamic sizes). - Model parameters (such as bias tensors)...

我记得HRNet构建时全部基于Functional API，没有自定义layer，也没有自定义量化的过程。如果有float的类型的确是一件奇怪的事情。也许是TFLite转换时出了差错？

我发现官方示例中使用的是uint8，我使用的是int8。你可以再试一下吗？我的开发PC已经打包了，暂时没有办法调试。

量化模型在测试时没有报错，但是现在回忆我当时并没有检查输入的类型，有可能当时是float32输入所以没有遇到你说的错误。 我觉得应该是模型转换出了问题，但是具体问题在哪里暂时不清楚。

我认为与数据集关系不大。训练后的模型已经定型，不受mark点个数的影响。量化时仅需要图像，与mark点也没有关系。 这个问题现在比较复杂。从截图来看转换后的模型参数包含浮点数，这个不应该。也许我们应该关注下TFLite转换时的log，看能否发现有用的信息。

I guess the increasing val loss is a sign of over fitting. HRNet is actually a large model and it may easily get over-fitted on this task. I would use...

Please avoid using that ResNet, which is a experimental project. Meanwhile TensorFlow has a mutual implementation here: https://www.tensorflow.org/api_docs/python/tf/keras/applications/resnet?hl=en

Hi, @luoww1992 The current model's input size is [batch_size, 256, 256, 3], and the output is a group of heatmaps, not images, which means they do not need to be...

Recommended steps if you are trying to transfer this project into another usage: 1. Do they have similar problems to solve? If you want to do classifications like separating cats...

Then `dataset.py` is a good start point.