!pip install -Uqq fastbook
!pip install emoji --upgrade
import fastbook
from fastbook import *
import getpass
import emoji

# Nice way to prompt for API key so we don't have to store it
key = getpass.getpass()

tiger_types = 'bengal', 'white', 'teddy'
path = Path('tigers')

path.exists()
# don't forget parenthesis after method calls...

if not path.exists():
    path.mkdir()
    for o in tiger_types:
        dest = (path/o)
        dest.mkdir(exist_ok=True)
        results = search_images_bing(key, f'{o} tiger')
        download_images(dest, urls=results.attrgot('contentUrl'))

# Clean the datasets up by getting rid of dead links
fns = get_image_files(path)
failed = verify_images(fns)
failed.map(Path.unlink)

# Check the docs
??DataBlock

??Resize

# Getting a Dataloaders from a dataset is a 2-step process. 
# First we initialise the Datablock to tell it what to expect the dataset we're about to give it.
tigers = DataBlock(
    
    # This specifies the types of input (images) and output/predictions (categories) we're working with
    # INput is called the INdependent variable
    # Output is called the Dependent variable
    blocks=(ImageBlock, CategoryBlock),
    
    # This tells the datablock how to read the raw data. 
    # Here we'll using get_image_files() as above
    get_items=get_image_files,
    
    # We'll split the dataset into 2 sets: training and validation
    # The learner will train itself on the training set and validate itself using the validation set.
    # The validation set will take 20% of the master set.
    # The seed is set at 42 so re-running against the same set will produce the same training/validation set for consistency
    splitter=RandomSplitter(valid_pct=0.2, seed=42),
    
    # get_y tells it how to label or categorise each item in the dataset.
    # parent_label is a built-in function that returns the name of the parent dir
    # Our tiger dataset is currently divided into subdirs tigers/bengali, tigers/white, tigers/teddy
    # Some datasets follow other conventions like having everything in the one dir but the filename 
    # has a particular format to differentiate categories. You could write a small custom function to parse each name
    # and return the label
    get_y=parent_label,
    
    # We need to resize our images to a uniform size.
    # This is so they can be passed around in a uniformly sized array to the GPU
    item_tfms=Resize(128))

# Now that we have our DataBlock template ready to go, we can give it our raw data
dataloaders = tigers.dataloaders(path)

??dataloaders.valid.show_batch

dataloaders.valid.show_batch()

??aug_transforms

# We can augment the data by telling the DataBlock to do some random resizing and cropping on it's training set
# tigers.new is a handy way to update an existing DataBlock rather than starting from scratch
# Just override the params you want updated
tigers = tigers.new(
    # Randomly crop to various points of the image to artificially introduce more variety to the training set.
    # This will make it a bit more robust and (hopefully) able to recognise a wider variety of images
    item_tfms=RandomResizedCrop(224, min_scale=0.5),
    # The aug_transforms will create skewed images in the training set for the same reason, more variety
    batch_tfms=aug_transforms())

dataloaders = tigers.dataloaders(path)

learner = cnn_learner(dataloaders, resnet18, metrics=error_rate)

# 'fine_tune' adds a random layer to the head of the pre-trained cnn (resnet18) and asks it to go through 20 epochs (rounds) of training
# There is a 'fit' method too but that would start training the whole CNN from scratch which would lose the benefit transfer learning 
# that's already been done with resnet18
learner.fine_tune(20)

# Check out the documentation for cnn_learner
doc(cnn_learner)

interp = ClassificationInterpretation.from_learner(learner)
interp.plot_confusion_matrix()

interp.plot_top_losses(5, nrows=1)