k_umr_mri/trainHistogram.py

import pydicom
import numpy as np
import matplotlib.pyplot as plt
import os
import csv
import math

import wavelet
import network
import keras
import tensorflow as tf
import random



actions = [ lambda img: wavelet.rotate(img, 1)
          , lambda img: wavelet.rotate(img, 2)
          , lambda img: wavelet.rotate(img, 3)
          , lambda img: wavelet.rotate(img, 4)
          , lambda img: wavelet.rotate(img, 5)
          , lambda img: wavelet.rotate(img, 6)
          , lambda img: wavelet.rotate(img, 7)
          ]

#while len(actions) > 4:
#  p = random.randint(0, len(actions) - 1)
#  del actions[p]

genData = []
labels = []
data = []

model = network.createModelHistogram()
pos = 0
first = True
with open("mimx.csv") as f:
  for row in csv.reader(f, delimiter=","):
    if first or len(row) < 9:
      first = False
      continue

    n = f"{row[2]}"
    while len(n) < 4:
      n = f"0{n}"

    fileName = f"Proband {row[0]}/SE00000{row[1]}/{row[0]}_{n}.dcm_histogram.npy"


    print(pos, end="\r")
    pos += 1

    x = np.load(fileName, allow_pickle=False)
    y = network.toOneHot([1.0,2.0,3.0,4.0,5.0], float(row[7]))
    for _ in range(10):
      noise = [float(random.randint(0,4)) for _ in range(4096)]
      labels.append(y)
      data.append(np.log(1 + x + noise))


print()
print("Start Train")
data = np.array(data)
labels = np.array(labels)


bestLoss = None
for epoch in range(128):
  print(f"Epoch {epoch + 1}")
  h = model.fit(data, labels, epochs=1)
  h = h.history
  loss = h['loss']
  if bestLoss is None or loss < bestLoss:
    bestLoss = loss
    print(f"Update best loss to {bestLoss} and save model")
    network.save(model, "modelHistogram.keras")
print("done")