## Data Meeting Update (Feb. 17, 2023)
### Cell Type Controls Clean-up
* Icon buttons with tooltips to save space
* Channel chips colored with channel color if that channel is opened
### Imported Embedding Visualization and Training
* Embeddings can be loaded, currently from embeddings.json in a zip file which is parsed into an array
* Sample image and embedding data
* File-name: `/nb/data-registry/data/labels/static/2d/Tissue-Lymph_Node/Liu_Validation_MIBI/Slide27Stain5_Point11_R1C10.npz:`
* Distribution: `total: 2354
cell_types:
BCELL: 1472
ENDOTHELIAL: 78
TCELL: 746
MACROPHAGE: 58`
* Julie's embeddings from: `np.savez_compressed("/workspace/output/embeddings.npz", embedding=embedding, y_truth=y_truth, file_names=file_names, cell_indices=cell_indices)`

Cell Types | UMAP | Raw Image
:---:|:---:|:---:
 |  | 

* Embeddings can also be trained on with hand-labeled data

Parameters | Loss Curve (confusion matrix has issues right now)
:---:|:---:
 | 

"Hand-labeled" | Prediction From Embedding | Ground Truth
:---:|:---:|:---:
 |  | 

* As we can see, cells with thrown out cell types need to either be removed from segmentation or labeled OTHER
### Other concerns, some more things to tackle
* Tried weighting classes proportionally by number of samples (tensorflow.js does not have built in imbalanced loss metrics) and results were much worse
* Streamlined data loading (including npz support) and export