Neural Network Pre-trained Models
Here, we’re releasing some pre-trained neural network models. These models are trained on different subsets of our single mouse tracking training data found here.
The code used to train the models as well as to run inferences is located on our github here.
The pre-trained models can be found at this Download Link.
Additionally, we share the same models in Zenodo with this Alternative Download Link.
This is the network we use for the strain survey. It has been trained on single mice images of varying coat colors inside an open field arena with a white background and grey walls.
This network we applied transfer learning of the first model (AllCoatColor_RegularOFA) to fine tune the model onto a new arena. The new arena contains bedding and a food cup.
This network was trained from scratch on a much more difficult arena. This environment contains reflections, non-uniform lighting, and high variability.
This network is trained on all of the above datasets.
While training our initial models and inferring on big datasets, we found difficult frames and incorporated them into our annotated dataset. To compare the influence of these frames, we trained a separate model without them in the training set.