kaggle image dataset_tensorflow

(0) go to account profile and create new api token and save the json file for later use
(1) !pip install -q kaggle
(2) !mkdir ~/.kaggle
    ! cp kaggle.json ~/.kaggle/
(3) !chmod 600 /root/.kaggle/kaggle.json
(4) go to dataset page and copy 'copy API command'
(5) !<paste copy api command here>
(6) !unzip "path to downloaded zip file" -d "path to target folder"

###############################################################################################################################################################################
# create tensorflow train_dataset

train_directory = 'path to downloaded train dataset'
validation_directory = 'path to downloaded validation dataset'

train_dataset = tf.keras.utils.image_dataset_from_directory(
  train_directory,
  labels='inferred', # folder structure with folder name as the class name
  label_mode='int', # class names map to integers, for example, happy=0, angry=1,sad=2 -> change 'int' to 'categorical', becomes onehot: (1,0,0), (0,1,0), (0,0,1) 
  class_name=None, # you can define variable CLASS_NAME=['angry','happy','sad'], the order(?) and name(!) must match with the folder structure
  color_mode='rgb',
  batch_size=32,
  image_size=(256,256),
  shuffle=True,
  seed=None,
  validation_split=None, # split train_dataset to train and validation, 0.2
  subset=None, # either 'training' or 'validation'. only used if validation_split is set
  interpolation='bilinear',
  follow_links=False,
  crop_to_aspect_ratio=False,
  **kwargs
  )
# same to create tensorflow validation dataset

###############################################################################################################################################################################
# check data
for i in train_dataset.take(1): # 1 batch size
  print(i)

plt.figure(figsize=(12,12))
for images, labels in train_datast.take(1):
  for i in range(16):
    ax = plt.subplot(4,4,i+1)
    plt.imshow(images[i]/255.)
    plt.title(tf.argmax(labels[i],axis=0).numpy()) # convert integers to class name: plt.title(CLASS_NAME[tf.argmax(labels[i],axis=0).numpy()])
    plt.axis("off")

###############################################################################################################################################################################
# efficient usage of data
training_dataset = (
    train_dataset,
    prefetch(tf.data.AUTOTUNE) # no need to define batch size here as it was defined above already
  )
# same to the validation dataset

###############################################################################################################################################################################
# need resize and scaling on the images
# build as a layer in neural network, so when deployed, it has built-in preprocessing functions
resize_rescale_layer=tf.keras.Sequential([
  Resizing(image_size,image_size),
  Rescaling(1./255),
  ])

###############################################################################################################################################################################  
# test single image
test_image=cv2.imread('path to image')
im=ft.constant(test_image,dtype=tf.float32)
im=tf.expand_dims(im,axis=0) # because of batch size
print(tf.argmax(model(im),axis=-1)) # with built-in preprocessing layer
print(CLASS_NAMES[tf.argmax(model(im),axis=-1).numpy()[0]]) # with built-in preprocessing layer