ssd_mobilenet_edgetpu_coco always reports 50% score for positive classifications on training (Object detection)

[NicholaiStaalung]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Debian 9.9 (Google Cloud Platform)
• Mobile device (e.g., Pixel 4, Samsung Galaxy 10) if the issue happens on mobile device: no
• TensorFlow installed from (source or binary): Binary
• TensorFlow version (use command below): 1.14
• Python version: 3.5.3
• Bazel version (if compiling from source): 0.28.0
• GCC/Compiler version (if compiling from source):
• CUDA/cuDNN version: 10
• GPU model and memory: P100 (15gb)

Please provide the entire URL of the model you are using?

t https://storage.cloud.google.com/mobilenet_edgetpu/checkpoints/ssdlite_mobilenet_edgetpu_coco_quant.tar.gz

Describe the current behavior The model always provides 50 % as the output score for positive classfications in training. In inference on Coral EdgeTPU it is providing a larger range of scores.

Describe the expected behavior

Code to reproduce the issue

Other info / logs

[pkulzc]

I didn't understand your question. Are you saying "when finetuning from that checkpoint the mAP for your training job look weird?"

[NicholaiStaalung]

Im saying that when i walk through individual predictions they are always scored as 50%. Which is odd in two ways

1. 50% should result in negative 2 Model doesn't differentiate between individual classes or detection's

Maybe it is something with the activation function for the output. I have played with both SIGMOID and SOFTMAX. However it is kind of hard for me to A/B test.

I have recently trained new models where predictions are returned higher than 50%. But they are still for most returned as the same value across observations. I think its weird and not plausible.

[fossil-ai]

I've been dealing with the same issue - what is the design choice here?