python - k-fold accuracy does not return to 0

Question

I have a problem where the accuracy of my k-fold does not return to 0, but continues the accuracy of the last epoch on the previous fold, please find a solution

Below is the code for the image generator

IMAGE_SHAPE = (224, 224)
TRAINING_DATA_DIR = str(directory_root)
  datagen_kwargs = dict(
  rescale = 1. / 255,
  rotation_range=25,
  shear_range=0.2,
  horizontal_flip=True,
  validation_split=.20)

# Make Validation Datagen
valid_datagen = tf.keras.preprocessing.image.ImageDataGenerator(**datagen_kwargs)
  valid_generator = valid_datagen.flow_from_directory(
  TRAINING_DATA_DIR,
  subset='validation',
  shuffle=True,
  target_size=IMAGE_SHAPE,
  batch_size=100)

# Make Train Datagen
 train_datagen = tf.keras.preprocessing.image.ImageDataGenerator(**datagen_kwargs)
 train_generator = train_datagen.flow_from_directory(
   TRAINING_DATA_DIR,
   subset='training',
   shuffle=True,
   target_size=IMAGE_SHAPE,
   batch_size=100)

Below is the code for k-fold

for train, test in kfold.split(inputs, targets):
# Define callbacks
checkpoint_path = f'/content/drive/MyDrive/Colab Notebooks/saveModel/Model 1/{fold_no}'
os.mkdir(checkpoint_path)
keras_callbacks = [
ModelCheckpoint(checkpoint_path, monitor='val_loss', save_best_only=True, mode='min')
]

x_t, x_ts  = inputs[train] , targets[test]
y_t, y_ts  = inputs[train] , targets[test]

model_history = model.fit(
  train_generator, 
  epochs=EPOCHS,
  verbose=1,
  steps_per_epoch=steps_per_epoch,
  validation_data=valid_generator,
  validation_steps=val_steps_per_epoch,
  callbacks=keras_callbacks).history

question from:https://stackoverflow.com/questions/65843778/k-fold-accuracy-does-not-return-to-0

Answer 1

Since I don't see where some variables are declared and assigned in the code, I'm not sure that this solves the problem. Anyway, if you call "model.fit()" without instantiating a new model for each fold, the result will be that the same model, with the already trained weights, will fit the new data increasing its previous accuracy.