You can find the project at MemeflyML.
Nick Burkhalter | Harsh Desai | Han Lee |
---|---|---|
Automatic meme generation model using Tensorflow Keras. Model is Dockerized and served as a REST API with FastAPI/uvicorn ASGI endpoint. A separate serving model serving is done with a combination of FastAPI/uvicorn ASGI endpoint with models served using Tensorflow Serving on Sagemaker.
- Numpy
- Pandas
- Tensorflow
- FastAPI
- Selenium
- Tensorflow Serving
- Docker
- MySQL
- MongoDB
- AWS ECR
- AWS Elastic Beanstalk
- AWS S3
- AWS Sagemaker
We used an encoder-decoder architecture for the meme generation task. Pre-trained Inception V3 architecture and weights are used as the encoder to extract embeddings from an input image. At the same time, we encode the texts into text embeddings and concat them together with image embeddings. For the decoder, we used GRU to to map the image and text embeddings to predict the next word in the text string.
At training time, we repeat the same image embeddings as input and send in text sequences in order, e.g., 0. this
, 1. this is
, 2. this is a
, 3. this is a sequence
. The model will try to predict the next word in the sequence given the input image embedding and text embeddings. We denote the beginning and the end of a text sequence with startseq
and endseq
.
At inferencing time, we send in image embeddings and the seed token startseq
to the model, and then repeatly send in the image embeddings and the prediction output of the previous timestep, until either we see endseq
or reach maximum sentence length. To improve the quality of the output, we used beam search to greedily select the best N sentences. But it has to be noted that beam search is neither optimal nor complete algorithm.
To increase varieties, we tried 1) adding Guassian noise to the input image and 2) choosing top N sentence scores using beam search.
The architecture is summarized here:
- Image
- Text
Please see Data Engineering for details.
Please see Machine Learning Engineering - Deployment for details.
Please see Data Engineering for details.
When contributing to this repository, please first discuss the change you wish to make via issue, email, or any other method with the owners of this repository before making a change.
Please note we have a code of conduct. Please follow it in all your interactions with the project.
If you are having an issue with the existing project code, please submit a bug report under the following guidelines:
- Check first to see if your issue has already been reported.
- Check to see if the issue has recently been fixed by attempting to reproduce the issue using the latest master branch in the repository.
- Create a live example of the problem.
- Submit a detailed bug report including your environment & browser, steps to reproduce the issue, actual and expected outcomes, where you believe the issue is originating from, and any potential solutions you have considered.
We would love to hear from you about new features which would improve this app and further the aims of our project. Please provide as much detail and information as possible to show us why you think your new feature should be implemented.