Cv mini project (1)
- 3. Two ways 1. Using tensorflow API 2. Basic Classifier from scratch 3
- 4. PLAN 01 Dataset 02 Import Libraries 03 Build the CNN Model 04 Train and validate 4
- 5. Dataset 1 5
- 6. About Dataset Dataset: Airplane and car images Source : Kaggle.com Number of images: 1000 6
- 8. Libraries, Tools and Technologies used. 8 ● Cv2 ● Spyder ● Anaconda ● NumPy ● Sklearn ● Matplotlib ● Keras
- 9. ● Split the data into two parts using sklearn - Training data and validation data ● Stored the resized images in array X and the label for it in array y. - Resize to uniform dimension 150 by 150 for height and width and 3 channels(i.e., for RGB) - Array Y will store label 0 for CAR and 1 for Airplane. - X is now an array of image pixel values - Y is a list of corresponding labels. 9 Data preprocessing
- 11. CNN ▪ Convolutional neural network to train the model. 11 But, What is convolutional neural network? ● It’s deep, feed-forward artificial neural network.(FFNN are also c/as Multilayer perceptrons). ● made up of neurons that have learnable weights and biases. ● Each neuron receives some inputs, performs a dot product. ● unlike a regular Neural Network, the layers of a ConvNet have neurons arranged in 3 dimensions: width, height, depth.
- 12. 12 A visualization is worth a thousand words Activation map or feature map
- 13. CNN ▪ Convolutional layer acts as a feature extractor. ▪ To extract features like edges, corners etc., 13
- 15. Activation Function ReLU - Rectified linear units(non-linear activation function) A(x) = max(0,x) Sigmoid Function suffers from vanishing Gradient problem. ❏ Sparsity - efficient in computations, ❏ more intuitive ❏ independent ❏ No complicated math. PS : Find notes and important links with click to add notes section of the ppt.15 BUT WHY ONLY ReLU?
- 16. Pooling layer ▪ Using all features in classification is computationally hard when the image size is large. ▪ For example 224x224, 480x480 size images. ▪ Prone to overfitting. ▪ Pooling (also called subsampling or downsampling) reduces the dimensionality of each feature map. ▪ But retains the most important information ▪ Max pooling 16
- 17. Flatten layer and Dropout Layer ▪ Flatten Layer - Convert 2D array to single linear vector. ▪ Dropout layer - ▫ Dropout randomly drops some layers in a neural networks and then learns with the reduced network. ▫ This way, the network learns to be independent and not reliable on a single layer. Bottom-line is that it helps in overfitting. ▫ 0.5 means to randomly drop half of the layers. 17
- 18. Model.summary() Formula to get number of parameters = (((kernel size * stride)*channels)+1) *filters) = (((3*3*1)*3))+1)*32 =((9*3)+1)*32 =28*32 = 896
- 19. Visualization of layers Use big image 19
- 20. Model Fitting using Back Propagation ▪ Back propagate errors(error calculation with loss function) ▪ Loss optimization (optimizer-update weights) ▪ Calculate gradient. ▪ With Learning rate 0.001 update the weight. 20 Positive Gradient Decrease weight Negative Gradient Increase weight
- 21. Optimizer and entropy Entropy - binary_crossentropy since there are two classes (Loss function) Optimizer - RMSprop with a learning rate 0.0001. ▪ Minimize the loss incurred using RMSprop optimizer. 21
- 22. 22
- 23. 23 Snapshot of epochs Number of Epochs=10 Training set has 768 images with batch size of 32. Hence Number of iterations to complete one epoch = 768/32= 24
- 24. 24 When Softmax was used for final fully connected layer
- 25. Using Tensorflow API ▪ Install object detection API ▪ Download pre-trained object detection model(COCO) ▪ Load images to be classified. ▪ Apply object detection on the load images. ▪ Based on the pre-trained model COCO dataset it tries to predict the object in the image 25
- 26. Future Scope Transfer Learning http://cs231n.stanford.edu/repor ts/2017/pdfs/411.pdf By - Yangyang Yu, Olivier Jin, Daniel Hsu , Stanford University TL (Transfer learning) is a popular training technique used in deep learning; where models that have been trained for a task are reused as base/starting point for another model. 26
- 27. References ▪ Presentation template by SlidesCarnival ▪ https://towardsdatascience.com/a-guide-to-an-efficient-way-to-build-neural-network- architectures-part-ii-hyper-parameter-42efca01e5d7 ▪ https://towardsdatascience.com/activation-functions-and-its-types-which-is-better- a9a5310cc8f ▪ https://www.tensorflow.org/tutorials/keras/basic_classification ▪ https://www.google.com/search?q=convolved+gif&rlz=1C1RLNS_enIN787IN787&tbm=isch&s ource=iu&ictx=1&fir=1VHtR2R2SFEVMM%253A%252CXjlYoIPax9NDkM%252C_&vet=1&us g=AI4_-kQpnIDbwPbGc- UEQVXxHxuXIZ_CRA&sa=X&ved=2ahUKEwilyfryn73hAhXkmuYKHT22A38Q9QEwB3oECA YQEg#imgrc=1VHtR2R2SFEVMM: 27
Editor's Notes
- #13: Where pooling layer reduces the resolution and fc layer acts as classifier which o/p image size = (imagesize + 2*padding - kernel size)/stride)+1 Stride,padding etc.. Smaller the filter size greater local information captured Receptive field technical jargons
- #14: CNNs, like neural networks, are made up of neurons with learnable weights and biases. Each neuron receives several inputs, takes a weighted sum over them, pass it through an activation function and responds with an output. The whole network has a loss function and all the tips and tricks that we developed for neural networks still apply on CNNs. https://medium.com/technologymadeeasy/the-best-explanation-of-convolutional-neural-networks-on-the-internet-fbb8b1ad5df8
- #15: https://towardsdatascience.com/a-guide-to-an-efficient-way-to-build-neural-network-architectures-part-ii-hyper-parameter-42efca01e5d7 Activation function to decide which activations to keep and which to be discarded….in simple terms layers will understand their language so we apply some activation on feature map When we do not have the activation function the weights and bias would simply do a linear transformation. A linear equation is simple to solve but is limited in its capacity to solve complex problems. A neural network without an activation function is essentially just a linear regression model. The activation function does the non-linear transformation to the input making it capable to learn and perform more complex tasks. Wide and shallow and narrow and deep https://towardsdatascience.com/activation-functions-and-its-types-which-is-better-a9a5310cc8f Non-linearity adapt with variety of data and to differentiate between the output. https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6
- #16: The problem is that in some cases, the gradient will be vanishingly small, effectively preventing the weight from changing its value. In the worst case, this may completely stop the neural network from further training. https://towardsdatascience.com/activation-functions-and-its-types-which-is-better-a9a5310cc8f https://www.quora.com/What-is-special-about-rectifier-neural-units-used-in-NN-learning
- #17: https://medium.com/nybles/create-your-first-image-recognition-classifier-using-cnn-keras-and-tensorflow-backend-6eaab98d14dd https://medium.com/technologymadeeasy/the-best-explanation-of-convolutional-neural-networks-on-the-internet-fbb8b1ad5df8 Overfitting - when our model memorizes the training data. The model will perform excellently at training time but fail at test time. allowing for assumptions to be made about features contained
- #18: In such methods, each of the neural network's weights receives an update proportional to the partial derivative of the error function with respect to the current weight in each iteration of training.
- #19: Formula to get number of parameters = kernel size * stride)*channels)+1)*filters
- #21: https://medium.com/datathings/neural-networks-and-backpropagation-explained-in-a-simple-way-f540a3611f5eThis is very similar to genetic algorithms where after each generation we apply a small mutation rate and the fittest survives.
- #24: https://stats.stackexchange.com/questions/233658/softmax-vs-sigmoid-function-in-logistic-classifier