Tracking using Neural Networks

[SinRas]

• [x] loading data with augmentations
• [x] more annotated data with different magnifications/lightings
• [ ] easily train and compare models
• [ ] inference using trained model, e.g. measure latency, memory, GPU/CPU requirements

@mtorkashvand missing anything?

[mtorkashvand]

@SinRas Where can I find the notebook for this?

[SinRas]

@mtorkashvand here is the link to the file https://github.com/mtorkashvand/compact-flourescent-microscope/blob/main/network/20230613_annotated_data_loader.ipynb

[mtorkashvand]

I added a function that reads annotations from 'annotations.h5' and frames from 'data.h5' (compatible with annotator), and updates a file named 'training_data.h5' located in 'V:\Mahdi\OpenAutoScope2.0\data\training_data'

'training_data.h5' has two groups: 'data' and 'annotations'. the 'annotations' group for each frame is an array of [y, x] coordinates.

[mtorkashvand]

training data:

image worms in following situations:

• [ ] 10x objective, L1 (focused and out of focus)
• [ ] 10x objective, L2 (focused and out of focus)
• [x] 10x objective, L3 (focused and out of focus)
• [x] 10x objective, L4 (focused and out of focus)
• [x] 10x objective, YA (focused and out of focus)
• [ ] 4x objective, L4 (focused and out of focus)
• [ ] 4x objective, YA (focused and out of focus)

and the same when the worms are between the glass and and the agar pad:

• [ ] 10x objective, L1 (focused and out of focus)
• [ ] 10x objective, L2 (focused and out of focus)
• [x] 10x objective, L3 (focused and out of focus)
• [x] 10x objective, L4 (focused and out of focus)
• [x] 10x objective, YA (focused and out of focus)
• [ ] 4x objective, L4 (focused and out of focus)
• [ ] 4x objective, YA (focused and out of focus)

[mtorkashvand]

I assume we won't use 4x for now because even young adults if tracked properly fit nicely in FoV.

[SinRas]

Just fixed a problem with the "no worm images" data that was used for training. the point of interest was jumping around, causing problem with the learning and also evaluation metrics.

In this commit: https://github.com/mtorkashvand/compact-flourescent-microscope/commit/cd7776c39ae01a6c2c428f6e0afbc557e47f4279

[SinRas]

Reduced Gamma variations since some images were too dark: 902d98d

[SinRas]

MODEL: ResNet18

We I confirm that ONNX running on CPU for Single/Small Batch Size, is much faster than PyTorch!
Roughlt ONNX 1/4 Batch takes 21/86 ms. PyTorch for 1/4 Batch takes 32/147 ms.

Here is the runtime of running model using ONNX and PyTorch on CPU of my laptop:

[SinRas]

MODEL: MNASNet0_75

We can see that the run-time roughly proportional to the GFLOP column in PyTorch PreTrained Model Specs.