This second example covers the more advanced use case of visual classification. The final retail model should recognize an uploaded photo of a dress and display two attributes: length and color.
Use case B uses a pre-trained general-purpose model and a large data set of women’s dresses to train two neural network models: recognizing color and the other to identify length. The TensorFlow framework and Intel Xeon powered machines help it achieve acceptable accuracy. Serving capability is part of Kubeflow and will be used to deploy this model as an API service. For this example, the web-based UI allows the user to upload a picture of a women’s dress and receive appropriate descriptions in a user-friendly manner.
Figure 4. Container Stack: TensorFlow, Keras, Kubeflow Python SDK, Kubeflow pipeline, retail, search engines, recommendation systems
While this is an example of the implementation there are innumerable ways that this Machine Learning framework can be implemented for unique business or industries. Three models that Grid Dynamics have deployed for real-life customers include:
The model can pick up the sample dress's main characteristics and find similar-looking short striped dresses from the catalog. This feature can also be termed “visual recommendation” or “show me more like this.”
Finding a matching plate or mug is a challenging process for the customer. Often, they can’t remember the name of their pattern or the company that produced it. Grid Dynamics engaged with Replacements.com to design and train a set of deep learning models and develop a set of visual search APIs capable of matching patterns based on a customer's photo uploaded.
Screws are visually very similar to each other, and therefore hard to distinguish. Traditional approaches to matching are frustrating and slow, and usually involve visual comparison to a huge catalog of choices. Wouldn’t it be nice if you could just point your smartphone to the screw that you need, and have it easily show up online?
Figure 5. Applications for Machine Learning framework