In this week’s exercise, you will use TFDS module for performing extract, transform and load (ETL) tasks on the Rock-Paper-Scissors dataset.
Upon completion of the exercise, you will have
Note : The assignment uses TF version 2 so if you run this notebook on TF 1.x, some things might not work.
from os import getcwd
import tensorflow as tf
import tensorflow_datasets as tfds
Remember to one hot encode the labels as you have 3 classes - Rock, Paper and Scissors.
You can use Tensorflow’s one_hot function (tf.one_hot) to convert categorical variables to one-hot vectors.
Useful parameters -
indices - A tensor containing all indicesdepth - A scalar defining the depth of the one hot dimension.# EXERCISE: encoding the labels using your own function for one-hot encoding
def my_one_hot(feature, label):
# Encode the labels to one-hot using tf.one_hot with depth equal to total
# number of classes here which are rock, paper and scissors
one_hot = tf.one_hot(indices=label, depth=3)
return feature, one_hot
# TESTING THE FUNCTION
_,one_hot = my_one_hot(["a","b","c","a"],[1,2,3,1])
print(one_hot)
tf.Tensor(
[[0. 1. 0.]
[0. 0. 1.]
[0. 0. 0.]
[0. 1. 0.]], shape=(4, 3), dtype=float32)
tf.Tensor(
[[0. 1. 0.]
[0. 0. 1.]
[0. 0. 0.]
[0. 1. 0.]], shape=(4, 3), dtype=float32)
You will be using tfds.load()] method to load the dataset. The dataset is already downloaded and unzipped for you in the data folder but if you are running on your local machine and do not have the dataset downloaded, it will first download and save the dataset to your tensorflow directory and then load it.
Useful parameters -
split - Which split of the data to load (e.g. ‘train’, ‘test’ [‘train’, ‘test’], ‘train[80%:]’,…)
data_dir - Directory to read/write data. Defaults to the value of the environment variable TFDS_DATA_DIR, if set, otherwise falls back to “~/tensorflow_datasets”
as_supervised- If True, the returned tf.data.Dataset will have a 2-tuple structure (input, label) according to builder.info.supervised_keys. If False the default, the returned tf.data.Dataset will have a dictionary with all the features
Note - Therock_paper_scissors:3.*.* dataset is already downloaded for you so if you specify the major version thisway while loading, it will try to find the dataset from the directory and load it. If none is present or the current dataset has been upgraded to a new major version, then it will try to download the new dataset to the directory.
# EXERCISE: Loading the rock, paper and scissors train and test dataset using tfds.load.
# Use data_dir=filepath as the dataset is already downloaded for you
filePath = f"{getcwd()}/data"
train_data = tfds.load(name="rock_paper_scissors:3.*.*", data_dir=filePath, split="train", as_supervised=True)
val_data = tfds.load(name="rock_paper_scissors:3.*.*", data_dir=filePath, split="test", as_supervised=True)
# Testing train_data and val_data if loaded correctly
train_data_len = len(list(train_data))
val_data_len = len(list(val_data))
print(train_data_len)
print(val_data_len)
2520
372
2520
372
You will apply the my_one_hot() encoding function to the train and validation data using map function. It will apply the custom function to each element of the dataset and returns a new dataset containing the transformed elements in the same order as they appeared in the input.
# EXERCISE: one-hot encode the train and validation labels using the function you defined earlier
# HINT - use map function https://www.tensorflow.org/api_docs/python/tf/data/Dataset#map
train_data = train_data.map(lambda x, y: my_one_hot(x, y))
val_data = val_data.map(lambda x, y: my_one_hot(x, y))
print(type(train_data))
<class 'tensorflow.python.data.ops.dataset_ops.MapDataset'>
<class 'tensorflow.python.data.ops.dataset_ops.MapDataset'>
Do remember that tfds.load() has a parameter called with_info which if True will return the tuple (tf.data.Dataset, tfds.core.DatasetInfo) containing the info associated with the builder.
# EXERCISE: Check the information about the dataset to see the dataset image shape
# HINT: Use with_info=True and data_dir
_,info = tfds.load(name="rock_paper_scissors:3.*.*", data_dir=filePath, with_info=True)
# DO NOT EDIT THIS
print(info.features['image'].shape)
(300, 300, 3)
(300, 300, 3)
Now you will define a simple 1-layer CNN model which will learn to classify the images into rock, paper and scissor!
# EXERCISE: Train a simple CNN model on the dataset
train_batches = train_data.shuffle(100).batch(10)
validation_batches = val_data.batch(32)
model = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(16, (3, 3), activation='relu', input_shape=(300, 300, 3)),
tf.keras.layers.MaxPooling2D((2, 2)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(3, activation="softmax")
])
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 298, 298, 16) 448
max_pooling2d (MaxPooling2D (None, 149, 149, 16) 0
)
flatten (Flatten) (None, 355216) 0
dense (Dense) (None, 3) 1065651
=================================================================
Total params: 1,066,099
Trainable params: 1,066,099
Non-trainable params: 0
_________________________________________________________________
# Now click the 'Submit Assignment' button above.
Remember to submit your assignment before you uncomment and run the next cell.
# # OPTIONAL EXERCISE: Compile and fit your model - use categorical loss and choose optimizer as Adam
EPOCH = 3
# You should get decent enough training accuracy in 3-4 epochs itself as this one layer model will heavily overfit
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["accuracy"])
history = model.fit(train_batches, epochs= EPOCH , validation_data=validation_batches, validation_steps=1)
print("Final Training Accuracy:-",history.history['accuracy'][-1])
print("Final Validation Accuracy:-",history.history['val_accuracy'][-1])
Epoch 1/3
252/252 [==============================] - 16s 45ms/step - loss: 275.1075 - accuracy: 0.8833 - val_loss: 2.7942 - val_accuracy: 0.7188
Epoch 2/3
252/252 [==============================] - 5s 17ms/step - loss: 0.0172 - accuracy: 0.9944 - val_loss: 3.2443 - val_accuracy: 0.7188
Epoch 3/3
252/252 [==============================] - 5s 18ms/step - loss: 0.0060 - accuracy: 0.9988 - val_loss: 3.1339 - val_accuracy: 0.7188
Final Training Accuracy:- 0.9988095164299011
Final Validation Accuracy:- 0.71875
%%javascript
<!-- Save the notebook -->
IPython.notebook.save_checkpoint();
<IPython.core.display.Javascript object>
%%javascript
<!-- Shutdown and close the notebook -->
window.onbeforeunload = null
window.close();
IPython.notebook.session.delete();
<IPython.core.display.Javascript object>