Tfhub object detection improvements

[cfascina]

Improvements at tfhub computer vision models (base64 and URL versions) after reviews.

Risks and Area of Effect

Testing

• Unit test
• Deployed to dev env
• Other (explain)

Reverting

• Contains Migration - Do Not Revert

[liuverta]

@cfascina In addition to my feedback above, I have an additional proposal. Let me know what you think:

The client's ModelAPI() current has a requirement (Slack thread)

arguments to ModelAPI() must be lists of data

Could we change these models' predict() to take a list and return a list? Something like

    def predict(self, data):
        result = []
        for image in data:
            file_name, img = image
            result.append(self.detect_objects(file_name, img))

        return result

Then we should be able to correctly match model API to this new format

    model_api = ModelAPI(
        [
            ['file_name', 'img']
        ],
        [{
            'file_name': '',
            'has_car': False,
            'score': 0,
            'bboxes': {
                'ymin': 0,
                'xmin': 0,
                'ymax': 0,
                'xmax': 0
            }
        }]
    ),

[cfascina]

@cfascina In addition to my feedback above, I have an additional proposal. Let me know what you think:

The client's ModelAPI() current has a requirement (Slack thread)

arguments to ModelAPI() must be lists of data

Could we change these models' predict() to take a list and return a list? Something like

    def predict(self, data):
        result = []
        for image in data:
            file_name, img = image
            result.append(self.detect_objects(file_name, img))

        return result

Then we should be able to correctly match model API to this new format

    model_api = ModelAPI(
        [
            ['file_name', 'img']
        ],
        [{
            'file_name': '',
            'has_car': False,
            'score': 0,
            'bboxes': {
                'ymin': 0,
                'xmin': 0,
                'ymax': 0,
                'xmax': 0
            }
        }]
    ),

@liuverta, the way you suggested is clearer to me. I made some changes.
Note that ModelAPI input argument had to be a dict: [{'file_name': '', 'url': ''}], with ['file_name', 'img'] the API shows 0 and 1 for these values.
And a few more adjstuments were necessary too.

[liuverta]

Great! Thank you so much for your work on this