use MC simulation to get more data

1. The first bin always gets 0 event.
2. It's not creating integer samples.
3. The zero time bin issue.

need simulate data with random fraction of light sources instead of fixed ones
#30

The quantization aware training fails because tf lite does not support some layers, such as conv1d, maxpool1d...

# annotate that only the Dense layers should be quantized.
def apply_quantization(layer):
    if isinstance(layer, tf.keras.layers.Dense):
        return tfmot.quantization.keras.quantize_annotate_layer(layer)
    elif isinstance(layer, tf.keras.layers.Conv1D):
        return tfmot.quantization.keras.quantize_annotate_layer(layer)
    elif isinstance(layer, tf.keras.layers.MaxPool1D):
        return tfmot.quantization.keras.quantize_annotate_layer(layer)
    return layer

annotated_model = tf.keras.models.clone_model(
    model = base_model,
    clone_function= apply_quantization )
# make the quantization aware
qaware_model = tfmot.quantization.keras.quantize_apply(base_model)

Error message:

RuntimeError: Layer conv1d_4:<class 'tensorflow.python.keras.layers.convolutional.Conv1D'> is not supported. You can quantize this layer by passing a `tfmot.quantization.keras.QuantizeConfig` instance to the `quantize_annotate_layer` API.

For now everything is working in my local computer. The final goal is to implement them in the hardware.

Should add a load data from files function

there's no linting for the python programme !!

Some features such as ground truth, threshold need to be seperated from the class for easily tuning.

a sketch to this is : commit 3ceb71b

It's necessary to apply the Bayesian Model into the algorithm since we need a prediction for few-shot data to get a feedback in real-time measurements!

One possible reason is the strides of the optimazer is too large, which couldn't cover the dip at zero time delay.
After set it to 1 the accuracy improves.
It's also because the data generated has a g2(0) close to the ground truth 0.5, which makes it harder for the classifier

related to #8
Currently it could only classify really 'dog-like dogs' and 'cat-like cats'.
If the source is mixed or close to the border between two classes the accuracy is pretty bad

g2(0) from analyzed data as ground truth
the Purdue paper

the probability function is different for n0, indicated in the paper

But it doesn't work in my implementation commit e83178b

Can tune the g2(0) by adding another source (sps, thermal ....) to check how the model changes #11

there is part of g2 values larger than 1 near the zero point, which is not useful for a simulated data #11

The simulation works only for one sps but once we add another sps the result does not seem correct. This needs to be solved

Since pip install does not work so need to build it from source.

I need more g2 data for training the model.

missing a .gitignore file

So far the simulation only have sps, coherent laser and non-detected options. The thermal source case needs to be added.

design a CNN model

I need to do a Levenberg-Marquardt fitting in one of the data to see the ML speedup.

I need to figure out all references:

• where to add: after each "claim" sentence
• find the related paper/ source
• also add them in the video

The developed model needs to be quantizated before compling on the edge TPU or coral