In [ ]:

    

# Import useful packages

%matplotlib inline
import dlt
import numpy as np
import chainer as C


    

In [ ]:

    

# (TODO) Implement a simple classifier manually

def simple_classifier(image, debug=True):
    average_pixel_value = np.mean(image)
    if debug:
        print("Average pixel:", average_pixel_value)
    # TODO - try to detect "a" or "b" just from average_pixel_value
    return "a"

dlt.CustomInput(simple_classifier)


    

In [ ]:

    

# Load up & show the data

train = dlt.load_hdf5("/data/uji/train.hdf")
valid = dlt.load_hdf5("/data/uji/valid.hdf")
print("  Training: %s" % train)
print("Validation: %s" % valid)
train.show()


    

In [ ]:

    

# Explore the data

print("vocab:", train.vocab)
print()
print("--- first training example: ---")
print("target:", train.y[0])
print("character:", train.vocab[train.y[0]])
print("input:", train.x[0])


    

In [ ]:

    

# (TODO) Evaluate your simple classifier

for label, image in zip(train.y, train.x):
    character = train.vocab[label]
    if character == "a" or character == "b":
        guess = simple_classifier(image, debug=False)
        if guess == character:
            pass  # TODO

accuracy = 0  # TODO

print("simple_classifier binary accuracy = {:.1%}".format(accuracy))


    

In [ ]:

    

# (TODO) Select data, define & train a model


    

In [ ]:

    

# (Optional) Demo your trained classifier

def classify(image):
    # TODO - replace with real scores
    # Hint - if you need a batch dimension, try: image.reshape(1, -1)
    scores = np.random.randn(1, len(train.vocab))
    print("TODO - classify image (image.shape %s -> scores.shape %s)" % (image.shape, scores.shape))

    return train.vocab[np.argmax(scores)]

assert classify(valid.x[0, :]) in train.vocab, "quick test"
dlt.CustomInput(classify)