Implement YoloX-tiny and nano for inference

blake-espeland commented 2 years ago

Based on this repository.

This offers an insane improvement in terms of performance over the current anchor-based YOLOv5. This feature is mainly the implementation of a new detection head.

blake-espeland commented 2 years ago

[x] Decoupled head: "Concretely, it contains a 1 × 1 conv layer to reduce the channel dimension, followed by two parallel branches with two 3 × 3 conv layers respectively."
[x] SPP: Spatial Pyramid Pooling. Selects the best features from the feature map to improve extraction.
[x] EMA weight updates: Exponential Moving Average. Places much higher weight on more recent data points. Here is an example implementation in PyTorch.
[x] SimOTA: Simulated Optimal Transport Assignment. Globally selects the best k bounding boxes. A simplified version of the Sinkhorn-Knopp algorithm. This paper describes it.
[x] IoU Loss: "We use BCE Loss for training cls and obj branch, and IoU Loss for training reg branch"

blake-espeland commented 2 years ago

Here is the structure of the decoupled head

blake-espeland commented 2 years ago

@jared3831 is having issues with the datatypes/shapes of input tensors. The repo we're using expects a json file that lists the image file and the annotations. Either there is a conversion between these two shapes, or it may be easier to just start over from scratch.

blake-espeland commented 2 years ago

Merged #10, however, this doesn't include @jared3831's changes.

blake-espeland commented 2 years ago

Finally included @jared3831's changes in #12

blake-espeland commented 2 years ago

Implementation complete, ironing things out now

Sprayer-Mods / ModelForge

Implement YoloX-tiny and nano for inference #5

Here is the structure of the decoupled head