devolearn/lineage_population_model: add postprocessing algorithm

[Mainakdeb]

• The new post processing algorithm modifies outputs from the lineage population model.
• The attached plots showcase postprocessed and non-postprocessed outputs (input video from EPIC dataset, id: 20080709_ceh-27_1_L1 )
• The loss is approximately 30% lower compared to non post processed outputs (MSELoss over 36 videos)
• The dotted lines represent model outputs, and the non dotted lines represent true values.

[ABD-01]

Hi! I think the function postprocess_preds function can be improved further by using boolean indices like this :

def postprocess_preds(self, preds):
    """
    Args:
        preds (numpy.ndarray): array that needs to be postprocessed.
    """
    for i in range(1, len(preds)):
        mask = preds[i] <= preds[i-1]
        preds[i][mask] = preds[i-1][mask] 

    return(preds)

Analysis (on 1000x1000 np.ndarray):

[Mayukhdeb]

Interesting, can you provide us with the benchmarks which shows the speedup in devolearn ? You can make the modifications in your own fork and test it out on colab.

cc: @Mainakdeb

[Mainakdeb]

Interesting, can you provide us with the benchmarks which shows the speedup in devolearn ? You can make the modifications in your own fork and test it out on colab.

cc: @Mainakdeb

Hi, I tried out the new post-processing function suggested by @ABD-01 on a sample video. Seems like this technique works well for higher dimensional numpy ndarays, but the difference is barely noticeable in the arrays we are working with.

cc: @Mayukhdeb