ResNet-50 version of the Faster RCNN (object detection) does not work in the browser

[mmmaks2004 mmmaks2004]

Hi,

1.I chose the ResNet-50 version of the Faster RCNN pre-trained model (object detection)

2.Trained this model on my data (tensorflow==1.5.0)

3. Convert Frozen Model (

1. tensorflowjs_converter --input_format=tf_frozen_model 
2. --output_node_names='detection_boxes,detection_scores,detection_classes,num_detections'  
3. --saved_model_tags=serve frozen_inference_graph.pb ./web_model

)

4. Run in browser (index.htm attached). The model works for 3-5 minutes and does not give any results (bad or good)

What am I doing wrong?

My Web Model:https://github.com/mmmaks2004/recognigetensor

5.Who has such a converted web model (object detectio)? I want to try a working model.Please, help me!

[Ping Yu]

Hi

The resnet model is quite big (over 100MB), The behavior you observed sounds like related to a GPU memory paging issue we fixed in the latest version.

Please make sure you are running the latest version of tfjs (v0.10.5) and try again.

We do have a more light-weight object detection model (coco-ssd) using mobilenet kernel, the accuracy might not be as good as resnet, but it depends on

your use case. You can check it out, the inference time is around 100ms on my desktop.

Ping

--
You received this message because you are subscribed to the Google Groups "TensorFlow.js Discussion" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tfjs+uns...@tensorflow.org.
Visit this group at https://groups.google.com/a/tensorflow.org/group/tfjs/.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/tfjs/f870712e-61a7-4566-9826-0d87e7cbc3b7%40tensorflow.org.

[mmmaks2004 mmmaks2004]

Hi
 No results in Firefox (Chrome - slowly but the result is)