In this exercise, you will use the Tensorflow Dataset’s Splits API and its concepts which you looked at in the week 2 lecture videos. Also
Also, you will look at some additional ways of loading things using Tensorflow Hub using the cats_vs_dogs v4 dataset.
Finally, you will use transfer learning using a pretrained feature vector from mobilenet to define a new classification model in the end.
Upon completion of the exercise, you will have
Note : The assignment uses TF version 2.1.0 and TFDS version 3.2.1 so if you run this notebook on TF 1.x or some other version of TFDS, some things might not work.
import tensorflow as tf
import tensorflow_datasets as tfds
from os import getcwd
print(tf.__version__)
print(tfds.__version__)
2.9.1
4.7.0
The next code block will download the mobilenet model from TensorFlow Hub, and use its learned features, extracted as feature_extractor and set to be fine tuned by making them trainable.
It is already downloaded for you but feel free to use the commented part to download the latest version from the tfhub.dev website when experimenting on your local machine.
import tensorflow_hub as hub
model_selection = ("mobilenet_v2", 224, 1280)
handle_base, pixels, FV_SIZE = model_selection
IMAGE_SIZE = (pixels, pixels)
filePath = f"{getcwd()}/data"
# You can also use directly to download from the source.
# MODULE_HANDLE ="https://tfhub.dev/google/tf2-preview/{}/feature_vector/4".format(handle_base)
# feature_extractor = hub.KerasLayer(MODULE_HANDLE,
# input_shape=IMAGE_SIZE + (3,))
# The data is already downloaded for you at the filepath
feature_extractor = hub.KerasLayer(filePath+'/mobilenet_v2_feature_vector',input_shape=IMAGE_SIZE + (3,))
feature_extractor.trainable = True
You need to use subsets of the original cats_vs_dogs data, which is entirely in the ‘train’ split. I.E. ‘train’ contains 25000 records with 1738 corrupted images to in total you have 23,262 images.
You will split it up to get
These 3 recordsets should be called train_examples, validation_examples and test_examples respectively.
Note: Remember to use cats_vs_dogs:4.*.* as dataset because 4.0 support the new Splits API.
# EXERCISE: Split the dataset
splits = ["train[:10%]", "train[90%:95%]", "train[95%:]"]
# Remember to use `cats_vs_dogs:4.*.*`
# https://www.tensorflow.org/datasets/catalog/cats_vs_dogs
# It has been downloaded for you so use the data_dir parameter
# else it will try to download the dataset and give you an error here
splits = tfds.load(name="cats_vs_dogs:4.*.*",
data_dir=filePath,
split=splits)
(train_examples, validation_examples, test_examples) = splits
# Testing lengths of the data if they are loaded correctly. Do not edit the code below
train_len = len(list(train_examples))
validation_len = len(list(validation_examples))
test_len = len(list(test_examples))
print(train_len)
print(validation_len)
print(test_len)
2326
1163
1163
Expected Output
2326
1163
1163
Now, you will take few of the examples from train set and shuffle them initially.
Then, you will map a custom function format_image formats the image by resizing them first to (224, 224) as that is the input image size for mobilenet and post resizing, it normalizes the image by dividing each pixel by 255.
Finally, you will create train, test and validation batches with size 16 here because of memory constraints. Do not edit the BATCH_SIZE in the code cell and while submitting the assignment.
Feel free to play around the BATCH_SIZE and other parameters if you are running this locally just for experimenting.
num_examples = 500 # DO NOT CHANGE THIS FOR THE GRADER, UPDATE AND USE IT WHEN PLAYING AROUND AND TRAINING IT LOCALLY.
num_classes = 2
# EXERCISE: shuffle and map the batches
# This will turn the 3 sets into batches
# so you can train and load batches
def format_image(features):
image = features['image']
image = tf.image.resize(image, IMAGE_SIZE) / 255.0
return image, features['label']
BATCH_SIZE = 16 # DO NOT EDIT
# For training batches, shuffle the examples by num_examples, map using the function defined above
# and batching using the batch_size.
# For validation and test batches, just avoid shuffling and follow the rest as training batch example
train_batches = train_examples.map(format_image).batch(BATCH_SIZE).shuffle(num_examples)
validation_batches = validation_examples.map(format_image).batch(BATCH_SIZE)
test_batches = test_examples.map(format_image).batch(BATCH_SIZE)
Here, you will use the mobilenet feature vector which you loaded before from Tensorflow Hub to create a new model for training.
This is a simple model where you are just using the feature vectors and adding the final dense layer to get the cat/dog classification.
Note: Always be careful of the input and output dimensions for a model loaded for transfer learning and use summary to check the dimensions.
# EXERCISE: Define the model
# The new model will take the features from the mobilenet via transfer learning
# And add a new dense layer at the bottom, with 2 classes -- for cats and dogs
model = tf.keras.Sequential([
feature_extractor,
tf.keras.layers.Dense(num_classes)
])
# Model summary to test your model layers, output shape and number of parameters.
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
keras_layer (KerasLayer) (None, 1280) 2257984
dense (Dense) (None, 2) 2562
=================================================================
Total params: 2,260,546
Trainable params: 2,226,434
Non-trainable params: 34,112
_________________________________________________________________
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
keras_layer (KerasLayer) (None, 1280) 2257984
_________________________________________________________________
dense (Dense) (None, 2) 2562
=================================================================
Total params: 2,260,546
Trainable params: 2,226,434
Non-trainable params: 34,112
# Now click the 'Submit Assignment' button above.
Training is not in the scope of this assignment but you can go ahead and train the network to achieve decent accuracy of 90% and above by training for epochs less than 5.
Note:This would take quite a lot of time to train on CPU (30-40 minutes per epoch) while here it can take 2-3 minutes per epoch.
Remember to submit your assignment before you uncomment and run the next cells.
# Compile the model with adam optimizer and sparse categorical crossentropy,
# and track the accuracy metric
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['acc'])
# Train it for a number of epochs. You should not need many (max 5)
# Train on the train_batches, and validation on the validation_batches
EPOCHS = 4
history = model.fit(train_batches, epochs=EPOCHS, validation_data=validation_batches,verbose=1)
model.summary()
Epoch 1/4
146/146 [==============================] - 72s 246ms/step - loss: 1.5058 - acc: 0.6879 - val_loss: 3.4384 - val_acc: 0.6019
Epoch 2/4
146/146 [==============================] - 22s 127ms/step - loss: 0.8041 - acc: 0.5606 - val_loss: 0.7793 - val_acc: 0.5993
Epoch 3/4
146/146 [==============================] - 21s 120ms/step - loss: 0.7805 - acc: 0.5542 - val_loss: 0.7804 - val_acc: 0.5899
Epoch 4/4
146/146 [==============================] - 21s 123ms/step - loss: 0.7804 - acc: 0.5546 - val_loss: 0.7803 - val_acc: 0.5907
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
keras_layer (KerasLayer) (None, 1280) 2257984
dense (Dense) (None, 2) 2562
=================================================================
Total params: 2,260,546
Trainable params: 2,226,434
Non-trainable params: 34,112
_________________________________________________________________
# Evaluate the model on the test batches
eval_results = model.evaluate(test_batches)
# And print the results.
for metric, value in zip(model.metrics_names, eval_results):
print(metric + ': {:.4}'.format(value))
73/73 [==============================] - 2s 26ms/step - loss: 0.7803 - acc: 0.5709
loss: 0.7803
acc: 0.5709
%%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();