Keras is a high level framework that runs on top of AI library such as Tensorflow, Theano, or CNTK. The key feature of Keras is that it allow to switch out the underlying library without performing any code changes. Keras contains commonly used neural-network building blocks such as layers, optimizer, activation functions etc and keras has support for convolutional and recurrent neural networks. In addition keras contains datasets and some pre-trained deep learnig applications that make it easier to learn for beginners. Essentially Keras is democrasting deep learning by reducing barrier into deep learning.
1. Keras: A Python framework for Deep Learning
CA-691
Mohit Kumar
Roll No. - 20961
M.Sc. (Comp. Applications)
(Course Seminar)
INDIANAGRICULTURALSTATISTICSRESEARCH
INSTITUTE,NEWDELHI
1
2. Contents
• Introduction
• Keras models
• Keras layers
• Image data augmentation
• Keras applications
• Transfer learning
• Keras datasets
• Keras callbacks
• Case study
• Summary
• References
2
3. Introduction
• Francois Chollet, the author of Keras, says:
“The framework was developed with a focus on enabling fast experimentation.
Being able to go from idea to result with the least possible delay is key to doing
good research.”
• Keras is open source framework written in Python
• ONEIROS ( Open Ended Neuro-Electronic Intelligent Robot OS)
• Released on 27 March 2015 by Francois Chollet
• Contains neural-network building blocks like layers, optimizer,
activation functions
• Support CNN and RNN
3
4. Introduction …
• Contains datasets and some pre-trained deep learning applications.
• Model check-pointing, early stopping
• Uses libraries TensorFlow, Theano, CNTK as backend, only one at a time
• Backend does all computations
• Keras call backend functions
• Works for both CPU and GPU
Fig: Keras architecture
4
5. Keras features
• Rapid prototyping-
1. Build neural network with minimal lines of code
2. Build simple or complex neural networks within a few minutes
• Flexibility-
1. Sometime it is desired to define own metrics, layers, a cost
function, Keras provide freedom for the same.
5
6. Keras models
• Two types of built in models
1. Sequential
2. Functional
All models have following common properties
1. Inputs that contain a list of input tensors
2. Layers, which comprise the model graph
3. Outputs, a list of output tensors.
6
7. Sequential Model
• It is linear stack of layers
• Output of previous layer is input
to the next layer.
• Create models by stacking one
layer on top of other
• Useful in situation where task is
not complex
• Provides higher level of
abstraction
Fig: An example of sequential model
7
8. Functional Model
• It define more complex models
• Such as directed acylic graphs
• Multi input output models
• Model with shared layers
• Possible to connect a layer with
any other layer
Fig: An example of functional model 8
10. Model methods
1. Compile: It is used to configure model. It accept following parameters
• Optimizer:
• This specifies type of optimiser to use in back-propagation algorithm
• SGD, Adadelta, Adagrad , Adam , Nadam optimizer and many others.
• Loss:
• It is the objective function
• It track losses or the drift from the function during training the model.
• For regression: mean squared error, mean absolute error etc.
• For classification: Categorical cross-entropy, binary cross entropy
• Different loss functions for different outputs
10
11. Model methods…
• Metrics:
• It is similar to objective function.
• Results from metric aren’t used when training model.
• It specifies the list of metrics that model evaluate during training and testing.
• The commonly used metric is the accuracy metric.
• Possible to specify different metrics for different output
11
12. Model methods…
2. Fit
This is the second important method
It train the model for specified epochs
It accept the following important arguments
• x: numpy array of training data
• y: numpy array of target data
• batch_size:
1. It specifies the number of samples per gradient update ,by default 32.
2. 32 samples of training data are fed into the model at a single time.
12
13. Model methods…
• epochs:
• An epoch is an iteration
• It specifies number of times training data is feed to the model.
• validation_split:
• Fraction of the data that is used as validation.
• Validation data is selected from the end samples
• At the end of each epoch, loss and metrics are calculated for this data.
• validation_data:
• shuffle:
13
14. Keras layers
• It consist of different types of layers used in deep learning such as:
1. Dense layer:
• A Dense layer is fully connected neural network layer
(Ramchoun et al, 2016)
14
15. Keras Layers…
2. Convolutional layer:
• Mostly used in computer vision.
• It extract features from the input image.
• It preserves the spatial relationships between pixels
• It learn image features using small squares of input data
• Finally, the obtained matrix is known as the feature map
15
17. Keras Layers…
3. Pooling layer
• Also called as subsampling or down-sampling layer
• Pooling reduces the dimensionality of feature map
• Retains the most important information
• There are many types of pooling layers such as
• MaxPooling and AveragePooling
• In case of max pooling, take the largest element from the rectified feature
map within that window.
• In average pooling, average of all elements in that window is taken.
17
19. Keras Layers …
4. Recurrent layer
• Basic building block of RNN
• This is mostly used in sequential and time series modelling.
5. Embedding layers
• Required when input is text
• These are mostly used in Natural Language Processing.
6. Batch Normalisation layer
• Normalize the activations of the previous layer at each batch
• Applies a transformation that maintains the mean activation close to 0 and
the activation standard deviation close to 1.
19
20. Image Data Augmentation
• Performance of deep learning neural networks often improves with
the amount of data available
• It artificially expand training dataset by creating modified versions of
images
• It create transformed versions of images in the training dataset that
belong to the same class as the original image.
• Transformation include operations from the field of image
manipulation such as shifts, flips, zooms, rotation etc.
20
22. Keras Applications
• Keras applications are deep learning models
• All these are image classifiers
• These applications are available with pre-trained weights
• Weights are downloaded automatically when instantiating application
• Stored at ~/. Keras/models
• These are different variation of pre-trained neural networks
• Each has its architecture, size, speed, accuracy etc.
22
24. Keras Applications…
1. VGGNET:
• Introduced by Simonyan and Zisserman in their 2014 paper
• Very deep convolutional networks for large scale image recognition
• 3×3 convolutional layers stacked on top of each other in increasing depth.
• Reducing volume size is handled by max pooling.
• Two fully-connected layers, each with 4,096 nodes are then followed by a
softmax classifier
24
26. Continued
2. RESNET
• Introduced by He et. al in their 2015 paper deep residual learning for image
recognition.
• works on the core idea “identity shortcut connection”
• Add the original input to the output of the convolution block
26
28. Keras Applications…
3. INCEPTION:
• The “Inception” architecture was first introduced by Szegedy et al.
in their 2014 paper- Going Deeper with convolutions
• It is a convolutional neural network (CNN) which is 27 layers deep
• It has some interesting layers called the inception layers
• The inception layer is a combination of various layers such as :
• 1×1 Convolutions layers
• 3×3 Convolutional layer and 5×5 layer
• Output filter concatenated into single output vector forming the input of
the next image
28
30. Transfer learning
• A model that have been trained for a particular task is reused as a
base or starting point for other model
• Some smart researchers built models on large image datasets like
ImagNet, Open images, and COCO
• They decided to share their models to the public for reuse
• This prevents the need to train an Image Classifier from scratch again
30
31. Transfer learning …
• To train an Image classifier that will achieve near or above human
level accuracy on image classification, it is required to have
1. Massive amount of data
2. Large computational power
3. Lots of time
• Since this is a big problem for people with little or no resources
• There is no need for transfer learning if we have
• Large dataset that is quite different from the ones above
• Have large computational power
31
32. Keras Datasets
• Keras contains various datasets that are used to build neural
networks. The datasets are described below
1. Boston House Pricing dataset:
• It contains 13 attributes of houses of Boston suburbs in the late 1970s
• Used in regression problems
2. CIFAR10:
• It is used for classification problems.
• This dataset contains 50,000 32×32 colour images
• Images are labelled over 10 categories
• 10,000 test images.
32
33. Keras Datasets …
3. CIFAR100: Same as CIFAR10 but it has 100 categories
4. MNIST:
• This dataset contains 60,000 28×28 greyscale images of 10 digits
• Also include 10,000 test images.
5. Fashion-MNIST:
• This dataset is used for classification problems.
• This dataset contains 60,000 28×28 greyscale images of 10 categories, along
with 10,000 test images
6. IMDB movie reviews data:
• Dataset of 25,000 movies reviews from IMDB
• labelled by sentiment (positive/negative).
7. Reuters newswire topics classification:
• Dataset of 11,228 newswires from Reuters, labelled over 46 topics
33
34. Keras Callbacks
• A callback is a set of functions to be applied at any given stages of the
training procedure
• Used to get an internal view or statistics of the model during training
• A list of callbacks are passed to the fit method of Sequential or Model
class
• There are different types of callbacks performing different operations
1. History:
1. This call-back is automatically applied to every keras model
2. History object is returned by the fit method of model.
34
35. Keras callbacks …
2. ModelCheckPoint:
This callback save the model after every epoch to a specified path
The advantage of this callback is that if model training is stopped due to any
reason, then model will automatically saved to the disk.
3. EarlyStopping:
1. This callback stop training when a monitored quantity stopped improving
2. Important parameter are
• Quantity to be monitored
• Patience, number of epochs with no improvement after which training will stop.
35
36. Keras Callbacks …
4. TensorBoard:
It is a visualization tool
This callback write a log for TensorBoard which allow to visualize dynamic
graph of training and test metrics
36
37. Case Study
• Image classifier developed using keras
• Web application is developed using Python’s flask framework
• Dataset contains 8032 images belongs to 8 classes such as
• Banana, Corn, Daisy, Fig, Jackfruit, Lemon, Orange, and Pomegranate
• ResNet50 model is used for transfer learning
• This model is image classifier for 1000 classes
• Above 8 classes also belongs to these 1000 classes.
• ResNet50 model’s last layer is used for classification
• To develop new model, last layer’s trainable property is set to False
• Then a new dense layer is added with 8 units 37
39. Case Study…
• Keras callbacks early stopping and model checkpoints are applied
• Validation loss is monitored with patience value 3
• An epoch with minimum validation loss value, will be saved to disk
• During training of model image data is augmented
• Model was set to run for 15 epochs
• Model was early stopped at 10th epoch because validation loss didn’t
decrease for 3 successive epochs from 8, 9 and 10th
• Model achieved training accuracy 0.7930 and validation accuracy
0.8418
39
42. Summary
• Keras is open source framework written in Python
• Easy and fast prototyping
• Runs seamlessly on both CPU and GPU
• Support both CNN and RNN
• Keras models
• Image data augmentation
• Keras callbacks
42
44. References
• Chollet, F. (2017). Xception: Deep learning with depthwise separable
convolutions. In Proceedings of the IEEE conference on computer vision and
pattern recognition (pp. 1251-1258).
• F. Chollet. Keras. https://github.com/fchollet/keras, 2015.
• Gulli, A., & Pal, S. (2017). Deep Learning with Keras. Packt Publishing Ltd.
• He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image
recognition. In Proceedings of the IEEE conference on computer vision and pattern
recognition (pp. 770-778).
• J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale
Hierarchical Image Database. In CVPR09, 2009
44
45. References
• Ioffe, S., & Szegedy, C. (2015). Batch normalization: Accelerating deep network
training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167.
• Perceptron: Architecture Optimization and Training. IJIMAI, 4(1), 26-30.
• Ramchoun, H., Idrissi, M. A. J., Ghanou, Y., & Ettaouil, M. (2016). Multilayer
Perceptron: Architecture Optimization and Training. IJIMAI, 4(1), 26-30.
• Simonyan, K., & Zisserman, A. (2014). Very deep convolutional networks for large-
scale image recognition. arXiv preprint arXiv:1409.1556.
• Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A.
(2015). Going deeper with convolutions. In Proceedings of the IEEE conference on
computer vision and pattern recognition (pp. 1-9).
45