Machine learning project
Download as PPTX, PDF1 like245 views
The document describes a machine learning project to classify images of dogs and cats using different models. The team used random forest initially, achieving 58% accuracy. They then used a neural network and CNN, reaching 60% and 79.45% accuracy, respectively. Transfer learning with pre-trained VGG19 and ResNet50 models significantly improved accuracy to 90.61% and 95.8%. Finally, fine-tuning pretrained ResNet50 by tweaking the last 15 convolutional layers achieved the highest accuracy of 95.55%
Machine learning project
- 1. MACHINE LEARNING PROJECT ECKOVATION TOPIC:- DOGS VS CATS CLASSIFICATION TEAM NAME : “ERROR_404” SUBMITTED BY:- HARSH JAIN (01151203116) PARV BHARTI(02151203116) SHANKAR(028512031156) VISHAL DUBEY(O3451203116) SUDHANSHU GOEL(03751203116)
- 2. AIM AIM :- • Our goal was to classify images of dogs and cats from the given data of 25000 images. • In our examples we will use two sets of pictures, which we got from Kaggle 12500 cats and 12500 dogs.
- 3. Approach • We Started with Random Forest • Then we applied Deep Neural network • We then applied Convolution Neural Network • Then We Applied CNN with pre-trained model (Transfer Learning) • Last but not least we applied fine tuned pre-trained model
- 4. Random Forest • The Accuracy when applying Random Forest was 58%.
- 5. Neural network • In Neural network the input feature is represented by nodes in input layers and output by nodes in output layer.
- 6. What happens in deep neural network? • Each pixel of image is represented by a node in input layer and are multiplied by an appropriate weight and then a activation function is applied on the result which act as the input of next layer and this process is carried out till the output layer. • The activation function make output non-linear and neural network can be applied to higher order logics. • The error is computed by comparing actual output and output of neural network.
- 7. How Weights are found? • First batches of images are fed into the network with random weights then output with those weights are calculated and error is determined. Then the error is back-propagated and weights are adjusted. This process is continued till error is reduced to minimal.
- 8. Our Experience with neural network • When we applied neural network we got an accuracy of 60% which is not par and is almost identical to Random forest
- 9. CONVOLUTIONAL NEURAL NETWORK (CNN) • Convolutional Neural Networks are a form of Feedforward Neural Networks. Given below is a schema of a typical CNN. The first part consists of Convolutional and max-pooling layers which act as the feature extractor. The second part consists of the fully connected layer which performs non-linear transformations of the extracted features and acts as the classifier. •
- 10. HOW CNN IMAGE CLASSIFIER WORK • The convolutional layer can be thought of as the eyes of the CNN. The neurons in this layer look for specific features. If they find the features they are looking for, they produce a high activation. • Convolution can be thought of as a weighted sum between two signals ( in terms of signal processing jargon ) or functions ( in terms of mathematics ).
- 11. • How convolution at a place is calculated
- 12. CNNs learn Hierarchical features
- 13. Our Experience with CNN • We got an accuracy of 79.45
- 14. What Next? • Since we have very less data compared to what is required in deep neural networks. Even if we get the data, it takes a large amount of time to train the network ( hundreds of hours ). • We will use pre-trained models. The models have been trained on millions of images and for hundreds of hours on powerful GPUs.
- 15. VGG19 And Resnet50 Models
- 16. Accuracy with pre-trained model • We got a accuracy of 90.61% with VGG19 and 95.8% with Resnet50
- 17. Fine tunning Pre-trained model • Earlier we only slightly modified the dense layer in pre-trained model to fit out problem but doesn’t trained convolution layer. We freezed convolution layer in keras as just tweaked dense layers. • Now we will train some convolution layers also.
- 18. Accuracy with Fine tuned model • We got a accuracy of 95.55% by tweaking last 15 convolution layers in Resnet50 pre trained model
- 19. CONCLUSION As we learn deep neutral network helps us to manage large amount of data with greater accuracy. It is the latest algorithm used to classify any kind of data it can be text or images. CNN is the advanced version of nueral network for classifying thousands of images having very high pixels rate and it has been best algorithm till date. Nueral works on many layers as we saw before rather than other algorithm didn’t work or divide data into layers which makes it more accurate and efficient.
- 20. Link to our kaggle kernel Random Forest : https://www.kaggle.com/harshhj25/dogs-vs-cats Neural Network:https://www.kaggle.com/jhashanku007/cat-vs-dog-p rediction?scriptVersionId=4645747 CNN : https://www.kaggle.com/jhashanku007/fork-of-cat-vs-dog-prediction- cnn?scriptVersionId=4660054 CNN with Pre tuned model: https://www.kaggle.com/jhashanku007/cat-vs-dog-cnn-with- different-pre-trained-model/versions CNN with tweaked layers: https://www.kaggle.com/jhashanku007/cat-vs-dog-cnn-with- different-pre-trained-model?scriptVersionId=4686576