Open sodagreat opened 5 years ago
Odd! The JS model is generated FROM the Keras model. Would you like to share your Keras comparison code?
Odd! The JS model is generated FROM the Keras model. Would you like to share your Keras comparison code?
I just use the predict interface of nsfw_detecor as the example. Found two factors that affect the results: image resize params and model quantization. I have checked the nsfwjs source code, which use bilinear interpolation. And i also compared results between normal model and quantized model, quantized model gives more reasonable predictions.
This is news to me! I wonder why quantization improves accuracy.
I have compared the results between keras/pb and tensorflowjs , the js model is always good, but the tensorflow pb model and keras h5 model produce much worse result. Every model I have all tried as i can. Is there something wrong with the model?