DeepSeeNet — :eyes: DeepSeeNet is a deep learning framewor...
🌿 Problem Statement A smart city needs smart mobility, and to achieve this objective, the travel should be made convenient through sustainable transport solutions. Transportation system all over the world is facing unprecedented challenges in the current scenario of increased population, urbanization and motorization. Farewell to all difficulties as reinforcement learning along with deep learning can now make it simpler for consumers. In this paper we have applied reinforcement learning techniques for a self-driving agent in a simplified world to aid it in effectively reaching its destinations in the allotted time. We have first investigated the environment, the agent operates in, by constructing a very basic driving implementation. Once the agent is successful at operating within the environment, we can then identify each possible state the agent can be in when considering such things as traffic lights and oncoming traffic at each intersection. With states identified, we can implement a Q-Learning algorithm for the self-driving agent to guide the agent towards its destination within the allotted time. Finally, we can improve upon the Q-Learning algorithm to find the best configuration of learning and exploration factors to ensure the self-driving agent is reaching its destinations with consistently positive results. Our aim is also to find optimum values of parameters of the fitting function alpha, gamma and epsilon, so that the agent can work in an optimized way with the most optimum parameter values. Hence, a comparative analysis has also been conducted. Methodology used The solution to the smart cab objective is deep reinforcement learning in a simulated environment. The smart cab operates in an ideal, grid-like city (similar to New York City), with roads going in the North-South and East-West directions. Other vehicles will certainly be present on the road, but there will be no pedestrians to be concerned with. At each intersection there is a traffic light that either allows traffic in the North-South direction or the East-West direction. We have assumed that the smart cab is assigned a route plan based on the passengers' starting location and destination. The route is split at each intersection into waypoints, and the smart cab, at any instant, is at some intersection in the world. Therefore, the next waypoint to the destination, assuming the destination has not already been reached, is one intersection away in one direction (North, South, East, or West). The smart cab has only an egocentric view of the intersection it is at: It can determine the state of the traffic light for its direction of movement, and whether there is a vehicle at the intersection for each of the oncoming directions. For each action, the smart cab may either stay idle at the intersection, or drive to the next intersection to the left, right, or ahead of it. Finally, each trip has a time to reach the destination which decreases for each action taken (the passengers want to get there quickly). If the allotted time becomes zero before reaching the destination, the trip has failed. The smart cab will receive positive or negative rewards based on the action it has taken. Expectedly, the smart cab will receive a small positive reward when making a good action, and a varying amount of negative reward dependent on the severity of the traffic violation it would have committed. Based on the rewards and penalties the smart cab receives, the self-driving agent implementation should learn an optimal policy for driving on the city roads while obeying traffic rules, avoiding accidents, and reaching passengers' destinations in the allotted time. Environment: The smartcab operates in an ideal, grid-like city (similar to New York City), with roads going in the North-South and East-West directions. Other vehicles will certainly be present on the road, but there will be no pedestrians to be concerned with. At each intersection there is a traffic light that either allows traffic in the North-South direction or the East-West direction. U.S. Right-of-Way rules apply: On a green light, a left turn is permitted if there is no oncoming traffic making a right turn or coming straight through the intersection. On a red light, a right turn is permitted if no oncoming traffic is approaching from your left through the intersection. To understand how to correctly yield to oncoming traffic when turning left.
🍂 Automatic detection of Diabetic Retinopathy using Deep Learning. Net2net technique of architecture used for training fundus images dataset from Kaggle. VGG16 is used as the teacher model and a self constructed dilated convolution block is the student model. These two modules are concatenated and global average pool is applied to give final predictions. LAB format of images to split into 3 colour channels was used as preprocessing and Adaptive Histogram Equalisation to the L-channel applied before recombining into original RGB format. Hyper-parameters: Optimiser : Stochastic gradient descent with momentum 0.9 Loss Function : Categorical Crossentropy Learning Rate : 0.0001 which was reduced by a factor of 0.125 after 10 epochs Activation Function : ReLU for Conv layers and Softmax for output layer. Epochs : 20 Batch size : 32 Class Weight : Appropriate weightage applied while training according to dataset of each class as imbalance in dataset.
🍂 The repository regroup all the work of my 3 months internship in the Laboratory Jean-Kuntzmann as research trainee. The main topic is a simulation of human memory while reading a text. There is also a second part about prediction of eyes movements thank to a deep learning algorithm.
1 ⭐ (0)cachett/TextComprehensionModel
🍂 Caption generation is a challenging artificial intelligence problem where a textual description must be generated for a given photograph. It requires both methods from computer vision to understand the content of the image and a language model from the field of natural language processing to turn the understanding of the image into words in the right order. Recently, deep learning methods have achieved state-of-the-art results on examples of this problem. Deep learning methods have demonstrated state-of-the-art results on caption generation problems. What is most impressive about these methods is a single end-to-end model can be defined to predict a caption, given a photo, instead of requiring sophisticated data preparation or a pipeline of specifically designed models. In this tutorial, you will discover how to develop a photo captioning deep learning model from scratch. After completing this tutorial, you will know: How to prepare photo and text data for training a deep learning model. How to design and train a deep learning caption generation model. How to evaluate a train caption generation model and use it to caption entirely new photographs.