<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Shivangg Tripathi on Medium]]></title>
        <description><![CDATA[Stories by Shivangg Tripathi on Medium]]></description>
        <link>https://medium.com/@shivangg?source=rss-4c0997cfd5d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*ffebhAHzWUxdbKSF.</url>
            <title>Stories by Shivangg Tripathi on Medium</title>
            <link>https://medium.com/@shivangg?source=rss-4c0997cfd5d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 04:28:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@shivangg/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How can being selfless be the most selfish thing to do.]]></title>
            <link>https://medium.com/@shivangg/how-can-being-selfless-be-the-most-selfish-thing-to-do-3b6b56afb40b?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b6b56afb40b</guid>
            <category><![CDATA[work]]></category>
            <category><![CDATA[selfishness]]></category>
            <category><![CDATA[selflessness]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Tue, 17 Sep 2019 00:25:57 GMT</pubDate>
            <atom:updated>2019-09-17T00:27:24.540Z</atom:updated>
            <content:encoded><![CDATA[<h4>Why money can’t buy you happiness.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/0*EJx3Yi3mqnP6SojQ" /></figure><p>It can’t. It&#39;s simple to understand. But it can buy you pleasure, though. But that’s also limited. After buying the best of the things, you would still become unhappy after some time when the dopamine goes down. Also, an observation can be made that you can’t achieve the newer levels of happiness highs if you never hit the newer lows of sadness.</p><p>So, materialistic happiness (which all can be bought) is limited. Therefore, to lead a happy life what one can aim for is a never-ending mission like ending poverty or hunger or ignorance. Such problems are so big that it&#39;s very difficult to solve them completely. Also, if you did solve them, it would truly be a great service to the world by making the life of everyone better in some way. Struggling (or hustling as they say) with these challenging problems can give new lows and thus pave a path of higher levels of happiness.</p><p>This can be a good framework to stay happy, thus by being selfless enough to work on hard human problems, one can achieve the selfish goal of higher levels of happiness.</p><p>The above are my reflections after watching: <a href="https://www.youtube.com/watch?v=0ZFjNlvmqsI">https://www.youtube.com/watch?v=0ZFjNlvmqsI</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b6b56afb40b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Using AlphaGo’s Deep Q-learning for learning dexterity]]></title>
            <link>https://medium.com/@shivangg/using-alphagos-deep-q-learning-for-learning-dexterity-679f457b7670?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/679f457b7670</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[q-learning]]></category>
            <category><![CDATA[dqn]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Tue, 05 Mar 2019 01:54:36 GMT</pubDate>
            <atom:updated>2019-03-05T01:54:36.023Z</atom:updated>
            <content:encoded><![CDATA[<p>This is the writeup for the Deep RL project based on the NVidia open source project “jetson-reinforcement” developed by <a href="https://github.com/dusty-nv">Dustin Franklin</a>. <strong>The goal of the project is to create a DQN agent and define reward functions to teach a robotic arm to carry out two primary objectives:</strong></p><ol><li>Have any part of the robot arm touch the object of interest, with at least a 90% accuracy.</li><li>Have only the gripper base of the robot arm touch the object, with at least an 80% accuracy.</li></ol><p>Hyperparameters tuning was the most challenging and time taking part of the project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/478/0*uhy0j2GW2Z2nIbeH.gif" /></figure><h3>Reward Functions</h3><p>The design of the reward function is the single most important task to accomplish any objective using reinforcement learning. Spending time on designing a good reward function saves a lot of time in later training and tuning the hyperparameters. The agent always tries to optimize the cumulative reward which can sometimes lead to undesired unplanned behavior. This is also called the <a href="https://en.wikipedia.org/wiki/Cobra_effect">Cobra effect</a>.</p><h3>Rewards for the interim states</h3><p>The distance from the gripper to the goal prop was measured using the distGoal(gripper, prop) method. Change of distance to goal was called delta = distGoal - lastGoal. As evident, negative delta implied the gripper was moving towards the goal pose.</p><p><a href="https://en.wikipedia.org/wiki/Moving_average#Exponential_moving_average">Exponential moving average</a> of the delta avgGoalDelta was used in the reward function.</p><pre>alpha = 0.7;<br>avgGoalDelta = avgGoalDelta * (1 - alpha) + delta * alpha;</pre><p>Higher alpha value meant lesser weightage to the previous values of distribution compared to the later values.</p><pre>reward = -C1 * avgGoalDelta<br>// -C1 is a negative constant</pre><p>It’s important to note that this is a positive reward function, i.e., as the gripper moves towards the goal prop, the rewards will be positive.</p><p>The important point to note in the case of positive rewards is if the agent can accumulate excessive positive reward without termination <em>by simply staying in its place</em>. This does not happen because if the delta stays <strong>0</strong> (position not changing), the rewards will (<em>smoothly</em>) become <strong>0</strong>. Thus, training with this reward function will not lead to the robot getting stuck to a position and still accumulate large rewards.</p><p>The faster the robotic gripper moves towards the prop, the more the rewards will be. This is because the value of the delta will be higher. Thus, this reward function incentivizes the fast robot gripper and prop collision.</p><h3>Rewards for the terminal states</h3><p>For task 1, there were 2 terminal states:</p><ol><li>Ground contact: The episode was terminated when the gripper touched the ground(<em>or was within 0.05 meters from the ground</em>). This was a case of dangerous outcome as the collision with the ground will lead to wear and tear of the arm. This was given a penalty of REWARD_LOSS = -0.10.</li><li>Collision with prop: The episode terminated when any part of the arm touched the goal prop. This was an intended outcome and was, thus, given a reward of REWARD_WIN = 0.20.</li></ol><p>For task 2, there were 3 terminal states:</p><ol><li><strong>Ground contact:</strong> Similar to the previous case, this was given a penalty of REWARD_LOSS = -0.10.</li><li><strong>Non-gripper collision with prop:</strong> When a non-gripper part of the arm collides with the goal prop, the agent is partly reaching the near the goal. Thus, this was given a reward of REWARD_WIN/2 = 0.10.</li><li><strong>Gripper Prop collision:</strong> When the gripper touches the goal prop, this was exact intended action and the agent was rewarded with REWARD_WIN = 0.20.</li></ol><p>Also, if any episode reached 100 frames, it was terminated with a penalty of REWARD_LOSS = -0.10.</p><h3>Type of Robot Joint Control Used</h3><p>Both velocity as well as the position control was experimented with. Position control performed well for both the task whereas the velocity control did not work well with the second task as it required more precise positioning of the gripper near the prop.</p><p>The advantage of the velocity control was that the arm had fewer chances of getting stuck at any one position because some <strong>non-zero constant output</strong> would almost always be given by the agent whereas in the case of position control, some <strong>non-zero constant output</strong> would mean that the arm will stay stuck in a particular position.</p><p>Joint position control was a good compromise because this was a direct way of controlling the arm. <strong>The important point to note was that the agent was rewarded not on its current distance from the goal but on the difference of successive distances from the goal(<em>deltas</em>).</strong> This reduced the chances of the robotic arm getting stuck because that would make the reward -&gt; 0 (smoothly) and agent always trying to maximize cumulative rewards would not get stuck in a single position.</p><p>The joint position control was used which performed well for both the tasks of the project.</p><h3>Hyperparameters</h3><p>The INPUT_WIDTH and INPUT_HEIGHT were kept equal to the size of the input image, i.e., 64 x 64.</p><p>Adam was tried with various parameters but RMSprop turned out to be better and was used to achieve the required accuracy for both the tasks.</p><p>Learning rate was set high enough to learn fast enough to reduce the number of episodes required to achieve the required accuracy. With higher rates, the accuracy was not reached and with lower rates, the number of episodes needed increasing by order of hundreds.</p><p>The degree of randomness though was set to true was done with a low EPS_END of 0.005 to prevent the agent the agent from trying random trajectories every so often, taking more time and episodes to reach required accuracy.</p><p>Larger batch size leads to convergence to narrow goal cases. This means that the agent will work only for cases in which the prop is in the narrow neighborhood of the prop training region. The agent will fail to generalize for the wide range of goal positions that it may encounter. Thus, a smaller batch size was used to generalize for various distance goal instances.</p><p>The LSTM size was chosen depending upon the length of the sequence that will need to be learned. Episodes terminated at 100 episode frames but the agent was able to come to a conclusive result(ground or prop collision) by the 30–40 episode frames. Also, making size a multiple of 32 may actually speed up the training by a small amount. Thus, a value of 64 was used for this task.</p><h3>Results</h3><p>The first task to achieve was:</p><blockquote><em>Getting the robotic arm to touch the goal prop with more than 90% accuracy.</em></blockquote><p>The watermarked screenshot of the <strong>task 1 accuracy graph</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Oc-QQKgz7erBQnuk.png" /></figure><p>It took close to 700 iterations to reach the 90% accuracy mark. It reached to 91.67% by the 1285th iteration.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zt8gz0iNOEqmaoXt.png" /></figure><p>The task 2 was more refined:</p><blockquote><em>Getting only the robotic gripper to touch the goal prop with more than 80% accuracy.</em></blockquote><p>This took more time and episodes but was a more meaningful task to accomplish for the agent. It touched the 80% accuracy mark after around 2400 iterations. The watermarked screenshot of the <strong>task 2 accuracy graph</strong>:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*xJ-_H6n4581ikQB2.png" /></figure><p>By the 2808th iteration, the accuracy was at 81.23%</p><h3>Building from Source (NVidia Jetson TX2)</h3><p>Run the following commands from terminal to build the project from source:</p><pre>$ sudo apt-get install cmake<br>$ git clone <a href="http://github.com/udacity/RoboND-DeepRL-Project">http://github.com/udacity/RoboND-DeepRL-Project</a><br>$ cd RoboND-DeepRL-Project<br>$ git submodule update --init<br>$ mkdir build<br>$ cd build<br>$ cmake ../<br>$ make</pre><p>During the cmake step, Torch will be installed so it can take awhile. It will download packages and ask you for your sudo password during the install.</p><h3>Future Work</h3><p>To improve the current results,</p><ul><li>Find a better reward function.</li></ul><p>In the current reward function, until the very end of episode, the agent is blind as to where it is heading. At the very end, it may encounter a collision with a gripper(high reward) or a ground(high penalty). The reward function should also consider the current distance between the gripper and the ground.</p><ul><li>Bright coloring the gripper and the prop.</li></ul><p>This will make it easier for the CNN in DQN algorithm to find the relation of gripper and prop position with the reward received.</p><p>DQN is a novel end-to-end reinforcement learning agent that can perform diverse range of tasks with superhuman level of mastery without human intervention. This is a major technical step in the quest of general AI. Paradoxically, the algorithms that power this machine learning are manually designed. Instead, if this learning could to automated, this could open up new opportunities. <a href="https://bair.berkeley.edu/blog/2017/09/12/learning-to-optimize-with-rl/">Learning to optimize</a> and <a href="https://www.google.com/search?tbo=p&amp;tbm=bks&amp;q=isbn:1461555299">Learning to Learn by Thrun &amp; Pratt, 2012)</a> was a step in this direction which I would like to implement as they used RL to learn the optimizer.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=679f457b7670" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dark secrets of the ML witchcraft]]></title>
            <link>https://medium.com/@shivangg/dark-secrets-of-the-ml-witchcraft-2bdfd08d26d1?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bdfd08d26d1</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[secrets]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Mon, 18 Feb 2019 14:46:38 GMT</pubDate>
            <atom:updated>2019-02-18T14:46:38.686Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/780/0*7acfLQ3YzhTR62wR.jpg" /><figcaption>Image source: <a href="https://www.magic-love-spells.com/old-witchcraft-secrets/">https://www.magic-love-spells.com/old-witchcraft-secrets/</a></figcaption></figure><p>This article is based on the research paper named <a href="https://homes.cs.washington.edu/~pedrod/papers/cacm12.pdf">A Few Useful Things to Know about Machine Learning</a> by Pedro Domingos.</p><p>Machine learning as a discipline allows its practitioner to solve problems just by generalizing from provided example solutions. The biggest advantage of this approach is that this prevents the need for explicitly writing rules of the solution.</p><h4>When we don’t write a write the rules for the solution, we, by definition, can’t introduce bugs in that solution.</h4><h3>Problem</h3><p>Needless to say, this sorcery comes with its own heap of shortcomings. Without the possession of the darker “folk secrets”, the ML projects end up in these scenarios:</p><ol><li>Taking way more time than estimated.</li><li>Giving sub-optimal results</li></ol><p>Let’s start revealing <strong>some</strong> of these dark secrets to making this witchcraft more powerful for all those practitioners out in the wild.</p><h3>Secrets to avoid the above problems</h3><h3>Learning = Representation + Evaluation + Optimization</h3><p>The three components of learning how to solve a problem are:</p><ol><li>Representation: Modelling of the problem and solution in the mathematical terms so the machine can understand it)</li><li>Evaluation: Scoring of the found solutions. Finding an error function whose output is inversely proportional to the gap between the desired and computed solution.</li><li>Optimization: A mechanism for finding the model with the best score(<em>or the least error).</em></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QToSiHtDdYRKfF1gPCLtbw.png" /><figcaption>Image source: <a href="https://homes.cs.washington.edu/~pedrod/papers/cacm12.pdf">https://homes.cs.washington.edu/~pedrod/papers/cacm12.pdf</a></figcaption></figure><p>Most texts go in-depth to explain the method for choosing the Representation component but neglect the other two. This leads to suboptimal results leaving the practitioner wondering where he/she went wrong.</p><h3>It’s the Generalization That Matters</h3><p>The ML model should be able to generalize after seeing all the examples from the dataset. The biggest mistake newcomers make is to train the model on <strong>all of the data. </strong>This gives real low error rates and they think the problem is solved until the model is tested on real data.</p><p>To solve this problem, the available data should be divided into <strong>training</strong> and <strong>testing</strong> data. After training on the training data, the model can be tested on the test data.</p><p><strong>Even after this precaution, the ML model may be contaminated. </strong>This is when you use the results of your model on the test data to tune the model hyperparameters. On doing this, the test data information leaks into your model through your model hyperparameter tuning.</p><p>To save your model from this contamination, cross-validation should be employed. Here, the total available data is divided into multiple units, training is done with one model kept aside for testing and averaging the results to see how the model does with a particular set of parameters.</p><p>To solve this problem, the available data should be divided into <strong>training</strong> and <strong>testing</strong> data. After training on the training data, the model can be tested on the test data.</p><h3>Data Alone is Not Enough</h3><p>The goal of ML models is to generalize from the given data. But if the model does not model an assumption beyond the data, it cannot generalize beyond the data. It will only be as good as pure random guessing.</p><p>Luckily, the assumptions, like :</p><ol><li>Similar data have similar classes</li><li>Limited complexity</li><li>Limited dependencies</li></ol><p>are enough for most of the general tasks. Thus, the ML algorithms have proven successful. ML models perform induction, the opposite of deduction, where they use a small amount of data input to yield a larger amount of data outputs. Although induction is way more powerful than deduction, as they require little information, they still do require some information. This quote from Pedro Domingos sums it up aptly.</p><blockquote>Learning is more like farming, which lets nature do most of the work. Farmers combine seeds with nutrients to grow crops. Learners combine knowledge with data to grow programs.</blockquote><h3>Feature Engineering is Key</h3><p>Finding the features that:</p><ol><li>Are independent of each other.</li><li>Correlate well to the output</li></ol><p>is the key to efficiently solving the problem with ML models.</p><blockquote>Purely raw data is often not recommended for training the model because the features may be too tightly interwoven for the model to decouple and learn from.</blockquote><p>Feature engineering is difficult because it is highly domain-specific whereas the ML models are general purpose.</p><p>Also, when trying to find the features that should be incorporated for learning, it should be noted that certain features don’t make sense in isolation but are important when taken together.</p><p>Thus, the features need to be carefully found. Also, one may think about training with a large combination of features but this is not only time-consuming but also causes overfitting.</p><h3>More Data Beats a Cleverer Algorithm</h3><p>There are two ways to increase the accuracy of the trained ML model which are:</p><ol><li>Use a cleverer learning algorithm</li><li>Get more data to train the model on</li></ol><p>This implies that a dumb algorithm with a larger amount of data would beat a cleverer algorithm with a lesser amount of data. In essence, ML is the process of making the model do complex work by just giving it more data.</p><p>Every problem in CS faces the constraints of <strong>Computation</strong> and <strong>Memory</strong>. ML problems have an added constraint of <strong>Training Data</strong>. The bottleneck in solving the problem has changed from decade to decade. Currently, with the mammoth amount of data getting generated by users, data has ceased to be the bottleneck.</p><p>Apart from all of the above, the number of human cycles is the most significant bottleneck. For better results, the organization should have the infrastructure that makes experimenting with:</p><ol><li>Different learning algorithms</li><li>Different data sources</li><li>Different learning problems</li></ol><p>easy and efficient. This will reduce the human effort and human cycles required to arrive at an optimal solution.</p><h3>Conclusion</h3><p>As with any great skill, ML has its own share of “wisdom” associated with it. Possessor of such wisdom separates a suboptimal ML model to the one that works great on the problem!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bdfd08d26d1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Perception Pipeline To Pick & Place objects]]></title>
            <link>https://medium.com/@shivangg/perception-pipeline-to-pick-place-objects-1d6fca31d05?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d6fca31d05</guid>
            <category><![CDATA[roses]]></category>
            <category><![CDATA[pr2]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[3d-object-detection]]></category>
            <category><![CDATA[perception]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Wed, 03 Oct 2018 11:41:08 GMT</pubDate>
            <atom:updated>2018-10-03T11:41:08.487Z</atom:updated>
            <content:encoded><![CDATA[<p>The mains steps in this perception Pipeline</p><h3>RGBD camera feed from ASUS Xtion</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*BmhzDax8YQ_n-ocy.PNG" /></figure><h3>1. Pipeline for filtering and RANSAC plane fitting implemented.</h3><h3>Statistical Outlier Filter</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kFVwMkDkxHiSkpzw.PNG" /></figure><p>This is used to remove the points that do not satisfy the given criteria. In this, we set 2 things:</p><ol><li>Number of neighbors to consider.</li><li>The distance beyond which the point is considered an outlier.</li></ol><h3>Vowel Grid Downsampling</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*NqvvpZZE2VGh37pZ.PNG" /></figure><p>This is required to downsample the incoming data. This helps in processing the data that is feature rich. Downsampling of data leads to more spatial date resolution. This provides us with a sparsely sampled point data cloud thus reducing the processing power needed to analyze the incoming data.</p><h3>Pass Filtering ( Cropping )</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*v7sNjrM9YureqGn0.PNG" /></figure><p>Specifying an axis leads to cropping along that direction with the specified magnitude. I put a passthrough filter for the</p><ol><li>x axis to filter the surrounding bins.</li><li>z axis to filter the stand of the table</li></ol><h3>Random Sample Consensus ( RANSAC )</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CDar6tecQWdyGBnZ.PNG" /></figure><p>RANSAC is used to cluster the points when :</p><ol><li>The particular shape of the cluster is known.</li><li>Number of clusters is known.</li></ol><p>We knew there is only 1(number) table present as well as that its planar(shape). Thus, RANSAC was used.</p><h4>Objects are in the outlier of the RANSAC plane fitting model</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*2igz0y-yHoPmhZ4u.PNG" /></figure><h3>2. Pipeline including clustering for segmentation implemented.</h3><h3>DBSCAN (Euclidean clustering)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*t8c6tuoiuxr6R4fc.PNG" /></figure><p>Used when the number of clusters is unknown(This resembles the real life scenario where this is almost always true ). Given the radius d and number of points to consider N, it segments the points into points, thereby segmenting the camera RGBD feed to separate objects.</p><h3>2. Features extracted and SVM trained. Object recognition implemented.</h3><h3>Features To extracted</h3><p>Good features extraction leads to better and robust object recognition.</p><p>(cube) RGB = RG and b axis where a point represents the value of R, G and B in the point.</p><p>(cylinder) HSV = Hue is the color(angular displacement), saturation is intensity of color(radial displacement) and value is the aural brightness (height).</p><p>Training with HSV leads to better results than RGB counterparts because it has low dependence on brightness levels.</p><h4>color histogram (uses color information as feature)</h4><ol><li>Divide the range of values into different bins.</li><li>Put the incoming values into their respective bins.</li><li>Normalizing can help to match objects with different size.</li><li>Compare the obtained color distribution with object to be found.</li></ol><h4>Surface normals (uses shape information as feature)</h4><p>Just like color histograms, here, shape histograms are made. In each bar(bin), the surface normals of same direction are placed. Sphere has a uniform constant surface normal distribution.</p><p>The histogram code can be found in the file <a href="https://github.com/shivangg/pr2detectit/blob/master/sensor_stick/src/sensor_stick/features.py">features.py</a>.</p><h3>Support Vector Machine (SVM)</h3><p>Features are stored as normalized histogram (unit variance and zero mean) and given to the SVM for training.</p><ol><li>Given features and labels of data points, it is trained and computes a decision boundary.</li><li>Uses the decision boundary to predict the new label of new data points.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jexehNh7jBovauIY.PNG" /></figure><h3>Steps taken to use the SVM:</h3><ol><li>Read the input images.</li><li>Extract the color (feature) vector.</li><li>Normalize the vector to zero mean and unit variance.</li><li>Define a label vector.</li><li>Shuffle and split the data into training and testing sets.</li><li>Define the classifier.</li><li>Train the classifier.</li></ol><p>SpecificationsNameValueIterations99Bins32accuracy90% (approx)</p><h4>Objects Detection</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*punPZD53gOBJpyH6.PNG" /></figure><p>The trained SVM model classifies the captures point out into an object. The accuracy of the model depends on the quality and number of object data samples the SVM was trained on. The output_*.yaml files for test\*.world can be found under pr2_robot/scripts/. Only the glue was not getting detected in test3.world because a book was in front of it. On deleting the book in front, the glue was successfully detected. The output file of this scenario is labeled output_3_glue_detected.yaml.</p><h3>Conclusion:</h3><h4>Future Work</h4><p>This project can be improvised by incorporating the pick and place operations. This will also improve the object recognition because some of the objects are <em>behind</em> others, like in test3.world, glue is behind the book if the book is picked up and placed(elsewhere or in the dropbox), the glue will be easily detected as evident from the <a href="https://github.com/shivangg/pr2detectit/blob/master/pr2_robot/scripts/output_3_glue_detected.yaml">output_3_glue_detected.yaml</a>.</p><p>Also, <em>collision detection</em> is an important aspect for the <em>efficient and safe manipulations</em> by the robot.</p><h4>Fails</h4><p>This project might fail if:</p><ol><li>Number of tables unknown; RANSAC needs to know the number of cluster to be effective.</li><li>If not trained for an object point cloud; this can be improved by training the SVM on the dataset of lots of real world objects’ point cloud.</li><li>Passthrough filters need to be changed for different scenarios to best filter to analyze only the objects of value.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1d6fca31d05" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Robotics Inference using Nvidia DIGITS]]></title>
            <link>https://medium.com/@shivangg/robotics-inference-using-nvidia-digits-d1ee59ac1662?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/d1ee59ac1662</guid>
            <category><![CDATA[digital-marketing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[transfer-learning]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[nvidia]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Sat, 15 Sep 2018 18:17:55 GMT</pubDate>
            <atom:updated>2018-09-16T20:55:19.282Z</atom:updated>
            <content:encoded><![CDATA[<p>In this project, a network will be trained for classifying real-world objects into predefined classes. In addition to training a network on the supplied dataset, a different network will be chosen and trained using self-acquired data. The quality and quantity of data acquired will be discussed.</p><p>The network would be able to successfully classify the objects in the self-collected data into 4 classes with reasonable accuracy.</p><h3>Introduction</h3><p>The classification of objects in the 2D image had been tried to be solved by a lot of classical methods but deep learning was able to become the new State of the Art technology. Deep learning shines when there is a lot of data.</p><p>The advantages of deep learning :</p><ol><li>Scales well with an increase in data.</li><li>No in-depth feature detection rules of the data need to be hardcoded.</li><li>Adaptable and transferable to the new problem datasets.</li></ol><p>In the defense of the classical algorithms[<a href="https://arxiv.org/pdf/1802.00036">1</a>], classical algorithms are better than deep learning in some cases:</p><ol><li>Needs no training time.</li><li>Financially and computationally cheap to run.</li><li>Easier to implement: The parameter tuning in the case of the classical algorithms is more straightforward because of the thorough understanding of the inner working of them. In case of the deep learning, even the researchers do not understand the inner working of the DL network(<em>almost a black box</em>).</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/609/0*LGzeFpeEy8PdC0yl.png" /></figure><p><em>Image Source: </em><a href="https://towardsdatascience.com/deep-learning-vs-classical-machine-learning-9a42c6d48aa"><em>https://towardsdatascience.com/deep-learning-vs-classical-machine-learning-9a42c6d48aa</em></a></p><p>To investigate the effectiveness of the inference models with random everyday objects that are used extensively by the humans, it was decided to classify the everyday objects on the work desk for this inference task.</p><p>Classes — Wallet, Mouse, NailCutter, Nothing</p><p>All the inference tasks essentially have similar steps that need to be followed. Nvidia DIGITS is a system to smoothen this workflow. This project will employ the Nvidia DIGITS workflow for developing the inference system.</p><p>Nvidia DIGITS is a way to quickly:</p><ul><li>Build a dataset that the network can take as input.</li><li>Define, customize and visualize a network architecture.</li><li>Train the network with the dataset.</li></ul><p>Nvidia DIGITS provides a cleaner interface to do all of the above, through the web browser.</p><p>The common inference tasks are:</p><ol><li>Classification: Answers the <em>what</em> in the image</li><li>Detection: Answers the <em>where</em> in the image</li><li>Segmentation: Pixel-level labeling of the images.</li></ol><h3>Formulation</h3><h3>Transfer Learning</h3><p>In practice, it is very rare that a CovNet is trained from scratch (with random initialization) because of the unavailability of the vast amount of dataset and computing power needed.</p><blockquote><em>Andrew Ng once explained neural network as it being like a rocket ship. Rocket ship needs a powerful engine and lot of fuel. Powerful engine but less fuel fails to make it to the orbit. Weak engine but lot of fuel fails to even liftoff. Analogously, DL models are engines and dataset is the fuel.</em></blockquote><p>Transfer learning is the finetuning of a network pretrained on a large dataset of the problem similar to the original problem.</p><p>In the project, transfer learning was leveraged. A GoogLeNet that was pretrained on the ImageNet(1000 Image classes) was used for classifying the everyday objects in the self-acquired data.</p><p>The <a href="https://arxiv.org/pdf/1409.4842.pdf">GoogLeNet</a> was used because it has 22 layers and 12 times less parameters than the AlexNet because of:</p><ol><li>The absence of the fully connected later at the end.</li><li>Use of the concept of a local mini network inside a bigger network called the Inference module. The inference module played the major role in the parameter reduction. This solved the problem of deciding which convolution to use — 1x1, 3x3, 5x5, 7x7. It enabled the network to use all of the convolutions concatenated depth wise.</li></ol><p>The naive Inference module used the depthwise concatenation of the 1x1, 3x3, 5x5 convolutional layers along with a 3x3 max pooling. The max pooling layer led to the depth staying the same or increasing in the output layer of the Inference module. This problem was solved by 1x1 convolutional layer with the 3x3, 5x5 and more importantly the max pooling layer. The number of 1x1 convolutional layer helped reduce the dimensionality of the output layer because the depth of output layer will just be the number of 1x1 convolutional layers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/0*90YwGgbyv0vePwdv.png" /></figure><p><em>Image Source: Inception module from “Going Deeper with Convolutions.”</em></p><p>GoogLeNet was faster and more accurate than the AlexNet. The idea around the development of GoogLeNet was that it should be able to run even in the smartphone.(<em>calculation budget around 1.5 billion multiply-adds on prediction</em>)</p><p>Also, SGD performed better than the RMSProp and was used in the finetuning of the network.</p><h3>Data Acquisition</h3><h3>Number of images</h3><h4>Number of Training Images</h4><p>MouseWalletNail CutterNothing3172166425</p><p>Training Images Distribution: This distribution of images for each class was used for training the network.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*HpDEwkSMyMraZjCi.png" /></figure><p>Number of Training Image in Each Class</p><h4>Number of Test Images</h4><p>MouseWalletNail CutterNothing10572218</p><p>Test Images Distribution: This distribution of images was used for the validation of the network. The network validates its accuracy after each run from these images.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*Sm9KXWfa74FvhhHP.png" /></figure><p>Number of Training Image in Each Class</p><p>The collected image data was augmented by using the rotation of:</p><ul><li>90 degrees clockwise</li><li>90 degrees counterclockwise</li><li>180 degrees</li></ul><p>As a result, the number of images in the dataset quadrupled.</p><p>As it can be seen here, the number of images for the Wallet and Mouse class is much more than the Nail Cutter. This will be explained in detail under the <a href="https://github.com/shivangg/inferX#Discussion">Discussion</a> section.</p><h3>Size of images</h3><p>The images were captured after setting the image aspect ratio to 1:1 so that the images are square in shape. This is because the network that we will be training on accepts 256x256 images. If the images are square, they can be easily resized to 256 pixels maintaining the aspect ratio. Had the aspect ratio been different from 1:1, it would lead to <em>the unnecessary cropping or squashing</em> of the image data causing loss of important data.</p><p>Also, each such square image had the file size of 7 - 13 kB.</p><h3>Image Color Type</h3><p>Real World objects are in color and the grayscale conversions of them will conflict with each other leading to 2 objects looking the same. Thus, the RGB images of the objects were captured and a network that can handle the RGB(<em>GoogLeNet</em>) was used to successfully classify them.</p><h3>Image Collection Method</h3><p>The smartphone camera was used for the collection of the images because:</p><ol><li>The same camera would be used for the test of the classification accuracy.</li><li>Lending to the ubiquitous nature of the smartphone cameras, the trained network will have the better probability of successfully classifying the objects in a different condition. This is because at least the testing and training data capture device stayed the same, i.e, smartphone camera.</li></ol><p>It should be duly noted that the environmental conditions in which the classification will need to successfully work should overlap with the environment in which the training image would be captured.</p><p>For example, if the classification needs to work <em>in a lighted room with diffused sunlight from the window</em> then the data acquisition should be done in the <em>same environment</em>.</p><p>MouseWalletNail CutterNothing</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*XALdEn7Bi1TqqSKw.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*DgIhfO1OlxCWIKTJ.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*9bODX-wZRgN4qTEm.jpg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*GMdDGRXwyo9rPnmP.jpg" /></figure><p>As evident, the images:</p><ol><li>Have an aspect ratio of 1:1</li><li>Diffused light that leads to shiny surface</li></ol><p>The test conditions will be similar to the above features.</p><p>In the supplied data, the images were also square.</p><p>Supplied Data</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*u1mQxdUvhA9r5Hrn.png" /></figure><h3>Results</h3><h4>Model for classifying supplied dataset</h4><p>Used GoogLeNet with SGD with the supplied data that achieved :</p><ol><li>The accuracy of 75.4098%</li><li>Inference time of 5.11062 ms using TensorRT</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/838/0*1urlsO6aDwctFvVX.png" /></figure><p>The accuracy and inference time of the network trained on the supplied data</p><p>The Caffe model trained on the supplied data can be found <a href="https://github.com/shivangg/inferX/blob/master/suppliedDataModel">here</a>.</p><p>The five millisecond inference time is good enough to be deployed for real-time classification for use in the robotics workbench.</p><p>Candy BoxBottle</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lT2V1MZCNm_Pa-AY.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zuHTT4a4kgayCOko.png" /></figure><p>It should be noted that when training the GoogLeNet with the ADAM, the accuracy dropped to 67.213%.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/882/0*mIOmP4MNUKeFRbxG.png" /></figure><p>The accuracy of the network decreased when trained with ADAM instead of SGD</p><h4>Model for classifying self-acquired dataset</h4><p>In the <a href="https://github.com/shivangg/inferX/blob/master/customDataModel">custom model</a>, the following important features can be seen as a function of epochs:</p><ul><li>Validation Loss</li><li>Validation Accuracy</li><li>Training Loss</li><li>Training Accuracy</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WbA2CHG9PABg7Apl.png" /></figure><p>Loss and Accuracy of Train and Validation data with SGD</p><p>The custom model was also tested on separate images the result which the network was able to successfully classify with good accuracy.</p><p>Separately Tested on Different Images</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/783/0*zbmLa2rk3ZJaK3PT.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/783/0*LVsOOnLn_td8qRHw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/783/0*GBxfbmWCZIq7m3_1.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/783/0*9ZHli7QvU-omYz0F.png" /></figure><p>As evident from the separate tested images, the finetuned GoogLeNet was able to successfully classify the objects in the problem set with good accuracy.</p><h3>Visualizations</h3><p>As an advantage of using Nvidia DIGITS, the network can be easily visualized. The weights and activations from the pooling layers, convolutional layers, Inception layer can be easily visualized.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*Q7Y61Fmzw4smtQsY.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/297/0*itetcIRSU0sLkvGd.png" /></figure><p>Input image of nail cutterInception layer(<em>near final layers</em>) Output Activations</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/906/0*WjlpsW7R-FxpT6G2.png" /></figure><p>Output Activation from Convolutional layer near input layer</p><h3>Discussion</h3><p>It was interesting to see the classification accuracy improvement over the epochs as the network was finetuned by SGD.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/797/0*ET1pP1zoedKKlKpM.png" /></figure><p>performance of a snapshotted earlier epoch less tuned network having a hard time classifying</p><p>When training was started with only 25 images in each class to test the effect of less data on the network, the network was able to classify the Nothing class but could not differentiate between the Wallet and Mouse. This was because the objects Wallet and Mouse were almost identical in terms of:</p><ul><li>Color</li><li>Surface shine</li></ul><p>This is the reason more data had to be collected for these 2 classes to As it can be seen from the earlier epochs snapshot of the model, there is more problem classifying between Wallet and Mouse than the nailCutter. As the epochs increased, this problem was minimized as the network got better at classification by learning the underlying patterns in the image.</p><h3>Accuracy vs Inference Time</h3><p>As the accuracy increase,</p><ol><li>The number parameters required increases leading to a bigger sized network.</li><li>More parameters mean more time to train the network.</li><li>The bigger network will be difficult to deploy on the embedded system which comes with less computing power.</li></ol><p>But the higher accuracy means better classification. This can be a dependable and reliable solution that can handle a wider spectrum of data.</p><p>To reduce the inference time, the network needs to be smaller, i.e, less number of network parameters.</p><ol><li>This narrows the learning capability of the network.</li></ol><p>Therefore, the trade-off between accuracy and inference time needs to be made. This should be highly dependent on the task at hand.</p><p>If a general solution is needed that can infer from the wide range of data inputs with the network running with good compute power on a workstation, accuracy can be increased.</p><p>If a very task-specific solution is needed, that needs to work fast on the embedded system with limited computing power in real time, the inference time should be reduced to be in the order of milliseconds.</p><h3>Conclusion</h3><p>The project good results in the supplied as well as the custom self-acquired data. This was essentially a result of:</p><ul><li>Quality data collection of the objects.</li><li>Smooth and fast development of ideas using the Nvidia DIGITS workflow.</li><li>And most importantly, transfer learning which enabled one with limited data and compute power to build a good classification network.</li></ul><p>This inference model can be deployed on a small robotic arm that can manipulate everyday objects on the table. The problems with this approach is that:</p><ol><li>Since no depth information is available, it will be difficult to manipulate the table objects as the trajectory calculations depend on the final pose to be reached.</li><li>Finer manipulations will require the computation the grasping pose that is still a open area of research.</li></ol><h3>Future Work</h3><p>In the future, rather than using a classification network, a detection network can be used to output the object coordinates in the image. Although this would require each image in the dataset to be annotated with a bounding box.</p><p>Also, it will be interesting to see the performance of such networks with the point cloud data to LIDARs, stereographic and RGB-D camera.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d1ee59ac1662" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Pumping up my arsenal for productivity]]></title>
            <link>https://medium.com/@shivangg/pumping-up-my-arsenal-for-productivity-3b8ca4ec35e1?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b8ca4ec35e1</guid>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[vimium]]></category>
            <category><![CDATA[vim]]></category>
            <category><![CDATA[mouse]]></category>
            <category><![CDATA[mx-master]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Wed, 15 Aug 2018 23:38:10 GMT</pubDate>
            <atom:updated>2018-08-15T23:38:10.849Z</atom:updated>
            <content:encoded><![CDATA[<p>This is the time to think about how I input to the computer for getting work done by it. The Jeff Atwood <a href="https://blog.codinghorror.com/the-code-keyboard/">blog on keyboard</a> made me think about it.</p><p>I wanted to know which mouse could I buy to increase my productivity. There were some very great looking (and ergonomic) options like the Logitech MX Master, MX Anywhere, unconventional Anker Vertical Mouse. They all looked like noble choices(with noble prices).</p><p>As always, I went to reddit to find better reviews about these. I found what mouse do the seasoned programmers like to use for their everyday use. Answers ranged in all the spectrum. Ranging from “buying the cheapest one available” to “buying the pricy(but deserving) Logitech MX Master 2S”. One user suggested to using the <strong>graphical tablet </strong>instead of a mouse altogether. Hmm, not bad. But this also is not very cheap.</p><p>Then came a suggestion from <em>timmyotc </em>that changed my perspective</p><blockquote>This is tangentially related to the thread, but I try to avoid mouse usage, even when browsing SO and the web. I downloaded Vimium. My life hasn’t been the same since.</blockquote><p>I realised the only part of my workflow where a mouse is required is when I’m using the browser to search for answers. So, I tried <a href="http://vimium.github.io/">vimium</a> out. Simple addition as a chrome extension and easy to understand keys if you already have an fair idea of using <em>vim</em>.</p><p>I just want to say, <em>My life hasn’t been the same.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b8ca4ec35e1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Realtime Plotting made simple using gnuplot]]></title>
            <link>https://medium.com/@shivangg/realtime-plotting-made-simple-using-gnuplot-aebb5252de5c?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/aebb5252de5c</guid>
            <category><![CDATA[aws]]></category>
            <category><![CDATA[gnuplot]]></category>
            <category><![CDATA[bash]]></category>
            <category><![CDATA[terminal]]></category>
            <category><![CDATA[plotting]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Wed, 15 Aug 2018 20:06:18 GMT</pubDate>
            <atom:updated>2018-08-15T20:06:18.028Z</atom:updated>
            <content:encoded><![CDATA[<p>So, I was stuck in tuning the hyperparameters for a DQN agent when a problem struck me! I needed graphs to reliably compare as well as view the whole performance of the agent in all the episodes. <em>gnuplots</em> was the way forward. Lightweight, simple and fast.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/300/0*CyQTDkLTiHpomB9K.png" /><figcaption>gnuplots</figcaption></figure><p>The syntax is as follows:</p><pre>plot &#39;plotRewards.dat&#39; using 1:2 with lines<br>pause 0.5 <br>reread<br>set autoscale </pre><p>The plotRewards.datis the file storing the data that was written it the relevent program in the format of every datax y on a new line.</p><p>The pauselets you specify the break <em>in seconds </em>(Self-explanatory).</p><p>Coming to reread is the most important part for <strong>realtime plotting </strong>as this tells <em>gnuplots</em> to re-read the above specified file again.</p><p>As the new values keep coming, they go out of range. Setting the autoscale using set autoscale enables autoscaling in both the axis.</p><p>This is really helpful for quick insight into the loads of data which is getting published in the terminal faster than can be sanely understood.</p><p>Please Clap if you found this helpful!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aebb5252de5c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The SIFT algorithm( Scale Invariant Feature Transform )]]></title>
            <link>https://medium.com/@shivangg/the-sift-algorithm-scale-invariant-feature-transform-fb8dd68145b2?source=rss-4c0997cfd5d------2</link>
            <guid isPermaLink="false">https://medium.com/p/fb8dd68145b2</guid>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Shivangg Tripathi]]></dc:creator>
            <pubDate>Sun, 08 Jul 2018 07:53:26 GMT</pubDate>
            <atom:updated>2018-07-08T07:53:26.282Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Q. Why use this algorithm?</strong><br>A. It is useful for *feature detection* and *feature description*.</p><p><strong>Q. What are features? Why should I care about them?</strong><br>A. Features are <strong>points of interest</strong> about the object that can be tracked. If we detect the correct features and track them, we can achieve a number of goals: <strong>Object detection</strong>: Using a training image, the good features of the object can be stored. When viewing the test image, we can look for the same features. The detection of the same features gives us the recognised object as well as its location in the test image.</p><p><strong>Q. That’s cute. But what features are considered good?</strong><br>A. The more robust the features, the better they are.</p><p><strong>Q. Robust against what? </strong><br>A. The features that were detected in the object should not change unless the object changes. So, the pixel values might change when:<br>1. The image is scaled<br>2. Brightness is changed<br>3. Noise is increased</p><p>but the ideal features should not change. <br>The ideal features detected should be robust against changes in:<br>1. Image scale: The object can be viewed not only from different angles but also from different distance from the camera. <br>2. Illumination.<br>3. Noise.</p><p><strong>Q. Enough on detecting features but what is feature description?</strong><br>A. Feature description is the *unique* and *robust* representation of the pixel that make up the feature. They are stored in the database which are then used to compare with test image features for object recognition.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fb8dd68145b2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>