nutrify

nutrify copied to clipboard

Minor changes on requirements.txt

1

Added some packages as i tried to install the packages to use the project but hit the dependencies issue

jovyinny

Create a way to generate images to boost inaccurate classes

1

Image generation is getting very good. For samples where gathering images is hard, could just generate them on the fly. This blog post showed good results with generated images: https://blog.ml6.eu/using-synthetic-data-to-boost-the-performance-of-your-object-detection-model-351a7f2171e2...

mrdbourke

A list of potential public food image datasets to explore & integrate into Nutrify. See another list of datasets/other resources here: https://github.com/mrdbourke/nutrify/discussions/23 ## Datasets * [Food101](https://paperswithcode.com/dataset/food-101) - classic food dataset,...

mrdbourke

Dashboard creation for data exploration

7

Create a dashboard to view and interact with different statistics about the data. This could be built using Streamlit. For example, input the annotations + results + more, then: *...

mrdbourke

Start breaking down `class_names` into more easily understandable foods

2

There can be a few kinds of models here, one for higher-level foods and one for lower-level foods. For example: * Higher-level model = predicts high-level classes, e.g. onion, pasta,...

mrdbourke

Go through `all_images` on Google Storage and resize them to take up less space

1

JPEG images don't have to be 100% quality in storage. The models only train on images of sizes 224, 299, 384 etc... So could potentially resize all images to only...

mrdbourke

OpenCLIP seems to work a bit better than generic CLIP. This could be set up in a way to perform autolabelling with several different models simultaneously. E.g. have a script...

mrdbourke

Resources for automatically creating labels for object detection and semantic segmentation: * Facebook's CutLER for unsupervised object detection/semantic segmentation - https://github.com/facebookresearch/CutLER

mrdbourke

Autolabelling: Setup a data collection pipeline (e.g. what happens when new data comes in?)

4

Data collection pipeline should be reactive to data coming into a bucket, for example: Images get added to bucket -> autolabelling pipeline happens for unlabelled images -> labelling cleaning happens...

mrdbourke

Create `export.py` (or something similar) for exporting models from PyTorch -> CoreML

1

Could have multiple export types as well: * PyTorch -> CoreML (for iOS) * PyTorch -> ONNX Web Runtime (for web) More: - [ ] Should be able to run...

mrdbourke