weighting_class_distributor.py

'''
After noticing such a substantial improvment 2.6% after my own manual changes,
validating my expectations, i've decided to automate this process.
'''
import keras
import os
import pickle
import numpy as np

validation_dir = "prepared_validation_data"

USE_WEIGHTS = True


DECAY = 0.9
#          w0     s1     a2    d3    wa4     wd5   sa6     sd7     nk8
WEIGHTS = [1.0,  1.0,   1.0,  1.0,   1.0,    1.0,  1.0,    1.0,    1.0]

#WEIGHTS =[0.030903154382632643, 1000.0, 0.020275559590445278, 0.013302794647291147, 0.0225283995449392, 0.025031555049932444, 1000.0, 1000.0, 0.016423203268260675]


mapping_dict = {0: "W",
                1: "S",
                2: "A",
                3: "D",
                4: "WA",
                5: "WD",
                6: "SA",
                7: "SD",
                8: "NK",}


close_dict = {
              0: {4: 0.3, 5: 0.3, 8: 0.05},  # Should be W, but said WA or WD NK nbd
              1: {6: 0.3, 7: 0.3, 8: 0.05},  # Should be S, but said SA OR SD, NK nbd
              2: {4: 0.3, 6: 0.3},           # Should be A, but SA or WA
              3: {5: 0.3, 7: 0.3},           # Shoudl be D, but SD or WD
              4: {2: 0.5},                   # Should be WA, but A
              5: {3: 0.5},                   # Should be WD, but D
              6: {1: 0.3, 2: 0.3},           # Should be SA, but S or A
              7: {1: 0.3, 3: 0.3},           # Should be SD, but S or D
              8: {0: 0.05, 1: 0.05, 2: 0.05, 3: 0.05, 4: 0.05, 5: 0.05, 6: 0.05, 7: 0.05, },  # should be NK... but whatever.
              }


model = keras.models.load_model(f"trained_models/1536269759_xception-v1_p-01.0456_ta-0.73_va-0.38.model")


while True:

    dist_dict = {0: 0,
                 1: 0,
                 2: 0,
                 3: 0,
                 4: 0,
                 5: 0,
                 6: 0,
                 7: 0,
                 8: 0}

    total = 0
    correct = 0
    closeness = 0
    # step 1
    for f in os.listdir(validation_dir):
        if ".pickle" in f:
            chunk = pickle.load(open(os.path.join(validation_dir, f), "rb"))

            for data in chunk:
                total += 1
                X = data[1]
                X = X/255.0
                y = data[0]

                prediction = model.predict([X.reshape(-1, X.shape[0], X.shape[1], X.shape[2])])[0]
                if USE_WEIGHTS:
                    prediction = np.array(prediction) * np.array(WEIGHTS)

                dist_dict[np.argmax(prediction)] += 1

                if np.argmax(prediction) == np.argmax(y):
                    correct += 1
                    closeness += 1
                else:
                    if np.argmax(prediction) in close_dict[np.argmax(y)]:
                        closeness += close_dict[np.argmax(y)][np.argmax(prediction)]
    print(30*"_")
    print("Weights:", WEIGHTS)
    print(f"Real mobile2-32-batch-0001.hdf5 accuracy: {round(correct/total, 3)}. Accuracy considering 'closeness': {round(closeness/total, 3)}")
    print(dist_dict)
    largest_key = max(dist_dict, key=dist_dict.get)

    print()

    with open('log.txt', "a") as f:
        f.write("Weights: "+str(WEIGHTS))
        f.write('\n')
        f.write(f"Real mobile2-32-batch-0001.hdf5 accuracy: {round(correct/total, 3)}. Accuracy considering 'closeness': {round(closeness/total, 3)}\n")
        f.write("Distribution: "+str(dist_dict))
        f.write("\n")
        f.write("\n")
        f.write("\n")

    WEIGHTS[largest_key] *= DECAY