nbortolotti / tensorflow-experiences

TensorFlow Experiences

Overview

Within this repository you will find several sections representing the development experiences in TensorFlow that I have experienced.

Sections

• Colaboratory: here you can find some experiences represented in Colaboratory, the simplicity and flexibility of the tool make the development of examples and tests directly in the cloud proposed by Google very attractive.

• Jupyter: here you can find some examples in a pure Jupyter format, these examples often need some libraries and elements that are not totally compatible with Colaboratory and that require a traditional local environment.

• experiences: these experiences are directly related in an ebook where I’m also including academic content on many of these concepts related to TensorFlow and Machine Learning.

• Python scripts and their tests: there are some python scripts that represent simple practices on TensorFlow. These will be migrated promptly to order them in a specific section.

• Experiences explained in Medium articles:

  • "custom_model_object_detection" where it is proposed, the generation of a personalized model for the detection of soccer players, specifically the article shows the experience with Lionel Messi.
  • "tie_dominant_color" This experience uses an object detection model and trims elements to then analyze its color and deliver options to the developer.

Translations

• English
• Español

tensorflow-experiences's Issues

synthesize_speech on fixed input text

Hi,

I come from a research team which works on bugs in real-world applications.

I want to make a report about a few potential performance improvement in your project.

In file experiences/assistant_sdk_action/hotword_inception_serving.py, synthesize_speech is called to generate audio files in each run. However, its input is a (or from a set of) fixed string. This behavior increases execution time and brings unnecessary service cost. A caching mechanism could help such situation, e.g. save audio file on disk at first API call, and re-use that file later.

Perhaps someone might be able to confirm if these seem reasonable. Also, if you need any help, I'm happy to provide some code snippet for fixing.

If you are not offended by it, we may include your response as a percentage number (e.g. xx% of applications faces xx problem, and xx% of them confirms it) in our research paper.

Thanks.

Could the custom dataset of Messi be open source?

Hi, @nbortolotti
I wonder if the custom dataset of Messi could also be open source?