Steff Lin

label smoothing loss问题

2

Hi： 在使用您代码的过程中，我对比了nn.CrossEntropyLoss()和LabelSmoothingCrossEntropy()，并在每个step打印loss值，得到以下，左边为smoothing loss，右边为nn.CrossEntropyLoss，似乎smoothing的loss稳定且只能以微小变化（0.5~0.6），而右边则还能降得更低。 我在自己写的代码上同样遇到过，请问您有没留意到这个问题？期待您的回复 ``` 0.6104838848114014 0.13801012933254242 0.609204888343811 0.14818525314331055 0.6149349212646484 0.1491081565618515 0.6321774125099182 0.17487832903862 0.6126755475997925 0.1317736655473709 0.6100113987922668 0.1340378224849701 0.6236982345581055 0.15228210389614105 0.5979918837547302 0.11493214219808578 0.6044023036956787 0.12213070690631866 0.5954407453536987 0.1088051125407219 Epoch:3 || epochiter: 32/139|| Totel...

problem of crowhuman datase?

7

The crowdhuman dataset doesn't have any ID tag, so how it's training successfully?

low map, cant train a good result in my datasets

16

I trained my datasets with 2 classes. First, I trained it withhead only, lr=1e-3, and then I trained the whole model, lr=1e-5. Val's class_loss is about 0.2, and reg_loss is...

hi, I know how to generate tfrecord , also I wanna know api how to read tfrecord. It seems to be use slim,but the official source code is too abstract...

own model input without internsity,how to modify this code

1

Hi, my model input only has x, y, z, andwithou intensity.Feature shape is 3. By modifying the relevant parameters of exporter.py, onnx was successfully converted and got params.h. **But don't...

Is there any reference command?

关于ncnn

3

您好： 请问有没测试过retinaface（ncnn的官方例子）在树莓派4b的速度如何，期待您的回复