k_umr_mri/train.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 = []

model = network.createModel()
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"

    y = np.array([float(row[7]), float(row[8]), float(row[9]), float(row[10])])
    genData.append( (fileName, y) )


def genDs():
  pos = 0
  for xy in genData:
    fileName = xy[0]
    #print(f"load '{fileName}' -> {pos}", end="\r")
    pos += 1

    images = np.load(f"{fileName}_images.npy", allow_pickle=False)
    #img = pydicom.dcmread(fileName).pixel_array
    #for action in actions:
    #  w = wavelet.refine(action(img))
    #  yield (1.0 * w.reshape((512 * 512,)) , xy[1])
    for img in images:
      yield (img, xy[1])

ds = tf.data.Dataset.from_generator(genDs, output_signature=(tf.TensorSpec(shape=(512*512,), dtype=float), tf.TensorSpec(shape=(4,), dtype=float)))
ds = ds.batch(32)

print()
print("Start Train")

checkpoint_filepath = '/home/kristian/Dokumente/SBI/2025/UmrMri/checkpoint.weights.h5'
model_checkpoint_callback = keras.callbacks.ModelCheckpoint(
    filepath=checkpoint_filepath,
    save_weights_only=True,
    monitor='val_accuracy',
    mode='max',
    save_best_only=True)


bestLoss = None
for epoch in range(128):
  print(f"Epoch {epoch + 1}")
  h = model.fit(ds, epochs=1, callbacks=[model_checkpoint_callback])
  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)
print("done")