<?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 Jaykumaran R on Medium]]></title>
        <description><![CDATA[Stories by Jaykumaran R on Medium]]></description>
        <link>https://medium.com/@jaykumaran2217?source=rss-959983f4da26------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*WZS1oh6tr0XO-DSrosfiYg.gif</url>
            <title>Stories by Jaykumaran R on Medium</title>
            <link>https://medium.com/@jaykumaran2217?source=rss-959983f4da26------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 15:03:02 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jaykumaran2217/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 to Fix an Ubuntu System Stuck on the Boot Screen]]></title>
            <link>https://medium.com/@jaykumaran2217/how-to-fix-an-ubuntu-system-stuck-on-the-boot-screen-2206d48b880b?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/2206d48b880b</guid>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[nvidia-driver]]></category>
            <category><![CDATA[device-drivers]]></category>
            <category><![CDATA[ubuntu]]></category>
            <category><![CDATA[gpu]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Sat, 01 Nov 2025 18:59:27 GMT</pubDate>
            <atom:updated>2025-11-01T19:05:46.934Z</atom:updated>
            <content:encoded><![CDATA[<p>If your Ubuntu system fails to start and drops you into a grub&gt; prompt or ends with a <strong>&quot;Kernel panic&quot;</strong> error, you are in the right place.<br> This guide provides a direct, technical solution to fix your system <strong>without losing any personal data</strong>.</p><p>The instructions are based on the common failure sequence shown in your images, which is typically caused by a <strong>corrupted boot configuration</strong> after a driver or kernel update.</p><h3>Analysis of the Problem</h3><p>May fix if one of following erros,</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A8NqQ5CPX1sjvdJ10HkelQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iNHTXdhXcZlDjGipk5t3uw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*no0dtMJP254xHbHW9G44-w.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YQoJhbZwUiCp2ocA3SvrYA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LeIirHwnUhrDzPmN0hMiSg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ez1kDk4x39JNhRZOTAh39Q.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CXwiCShxrdMpijsTj6KgeQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NjPGvgK_9iUioJ41ViAiFg.jpeg" /></figure><ol><li><strong>GRUB Rescue Prompt (</strong><strong>grub&gt;):</strong><br> Your system&#39;s bootloader (GRUB) starts, but it cannot find its configuration file (grub.cfg) or the Linux kernel. This is the first sign that the boot process is broken.</li><li><strong>error: unknown filesystem:</strong><br> When you try to manually inspect your drives from the GRUB prompt, it fails to recognize the filesystems correctly, preventing you from locating the necessary boot files.</li><li><strong>Kernel panic - not syncing: VFS: Unable to mount root fs:</strong><br> This is the critical failure. It means the kernel started loading but could not find or access your main Ubuntu partition.<br> This is usually because the <strong>initramfs</strong> (a temporary root system loaded at boot) is missing the essential drivers (e.g., for your NVMe SSD) to continue the boot process.</li><li><strong>ACPI BIOS Errors:</strong><br> These messages are common warnings and are <strong>not</strong> the cause of the failure. They can be safely ignored.</li></ol><p><strong>Conclusion:</strong><br> The problem is a broken link between the <strong>bootloader (GRUB)</strong>, the <strong>Linux kernel</strong>, and the <strong>initial RAM disk (</strong><strong>initramfs)</strong>.<br> The solution is to use a <strong>Live Ubuntu USB</strong> to get temporary access to your system and <strong>rebuild these broken components</strong>.</p><h3>Step-by-Step Repair Instructions</h3><p>Follow these instructions exactly to repair your system.<br> This process is safe for your personal data.</p><h3>Part 1: Boot from a Live Ubuntu USB</h3><ol><li><strong>Create a Bootable USB:</strong><br> On a separate, working computer, download an Ubuntu ISO from the <a href="https://ubuntu.com/download/desktop">official website</a> and use a tool like <strong>BalenaEtcher</strong> or <strong>Rufus</strong> to write it to a USB drive.</li><li><strong>Start Your Computer from the USB:</strong><br> Plug the USB into the broken computer, turn it on, and press the boot menu key (often <strong>ESC</strong>, <strong>F2</strong>, or <strong>F12</strong>).<br> Select your USB drive from the list.</li><li><strong>Choose “Try Ubuntu”:</strong><br> When the menu appears, select this option.<br> This will launch a full Ubuntu desktop session from the USB drive without altering your system.</li><li><strong>Connect to the Internet.</strong></li></ol><h3>Part 2: Access Your Installed System with Chroot</h3><p>Open a terminal (Ctrl + Alt + T) in the live session and run the following commands.</p><ol><li><strong>Find Your Partition Names:</strong></li></ol><pre>lsblk</pre><ol><li><strong>Explanation:</strong> Lists your drives. Identify your main Ubuntu partition (likely the largest ext4 partition, e.g., /dev/nvme0n1p4) and your EFI partition (a small fat32 partition, e.g., /dev/nvme0n1p2).</li><li><strong>Mount Your Main Partition:</strong> (Replace nvme0n1p4 if yours is different)</li></ol><pre>sudo mount /dev/nvme0n1p4 /mnt</pre><ol><li><strong>Explanation:</strong> Makes your installed system’s files available in the /mnt folder.</li><li><strong>Mount Your EFI Partition:</strong> (Replace nvme0n1p2 if yours is different)</li></ol><pre>sudo mount /dev/nvme0n1p2 /mnt/boot/efi</pre><ol><li><strong>Explanation:</strong> Mounts the boot partition.<br> If you get an error that the directory doesn’t exist, run:</li></ol><pre>sudo mkdir -p /mnt/boot/ef</pre><p>and try again.</p><ol><li><strong>Prepare the Chroot Environment:</strong></li></ol><pre>for d in /dev /proc /sys /run; do sudo mount --bind $d /mnt$d; done</pre><ul><li><strong>Explanation:</strong> Allows the live session to interact with your system’s hardware and processes.</li></ul><p><strong>2. Enter Your System:</strong></p><ul><li>sudo chroot /mnt</li></ul><p><strong>Explanation:</strong> This command “changes root” and places your terminal session inside your broken installation, allowing you to run commands as if you had booted normally.</p><h3>Part 3: Execute the Repairs</h3><p>You are now in control of your installed system.<br> Run these commands to fix the boot components.</p><ol><li><strong>Remove Old NVIDIA Drivers:</strong></li></ol><pre>apt-get purge --autoremove &#39;*nvidia*&#39;</pre><ol><li><strong>Explanation:</strong> Completely removes potentially corrupted graphics drivers, a common cause of the issue.</li><li><strong>Reinstall the Kernel and Rebuild Initramfs:</strong></li></ol><pre>apt-get install --reinstall linux-image-generic</pre><ol><li><strong>Explanation:</strong> Reinstalls the kernel package, which automatically triggers a rebuild of the initramfs with the correct drivers needed to mount your root filesystem at boot.</li><li><strong>Update the GRUB Configuration:</strong></li></ol><pre>update-grub</pre><ol><li><strong>Explanation:</strong> Scans for the reinstalled kernel and generates a new, correct boot menu configuration file.</li><li><strong>Reinstall the GRUB Bootloader:</strong> (Use the disk name, not the partition)</li></ol><pre>grub-install /dev/nvme0n1</pre><ol><li><strong>Explanation:</strong> Reinstalls the bootloader to the primary disk, fixing any potential corruption.</li></ol><h3>Part 4: Finalize and Reboot</h3><ol><li><strong>Leave Your System:</strong></li></ol><pre>exit</pre><ol><li><strong>Explanation:</strong> Exits the chroot environment and returns you to the live USB&#39;s terminal.</li><li><strong>Unmount the Drives:</strong></li></ol><ul><li>sudo umount -R /mnt</li></ul><ol><li><strong>Explanation:</strong> Cleanly unmounts your computer’s partitions.</li><li><strong>Restart the Computer:</strong></li></ol><ul><li>reboot</li></ul><ol><li><strong>Explanation:</strong> Reboots your machine. <strong>Remember to remove the USB drive</strong> as it starts up.</li></ol><p>Your system should now boot correctly into the Ubuntu login screen.<br> You can then safely reinstall your graphics drivers using the “Additional Drivers” tool.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2206d48b880b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Install 50 Series Nvidia Drivers on Ubuntu]]></title>
            <link>https://medium.com/@jaykumaran2217/install-50-series-nvidia-drivers-on-ubuntu-fae366d33073?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/fae366d33073</guid>
            <category><![CDATA[5060-ti]]></category>
            <category><![CDATA[nvidia-driver]]></category>
            <category><![CDATA[ubuntu]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Mon, 22 Sep 2025 18:06:05 GMT</pubDate>
            <atom:updated>2025-09-22T18:31:05.040Z</atom:updated>
            <content:encoded><![CDATA[<p>Tested with <strong>Ubuntu 20 Focal — i9 14th gen 5060 Ti</strong>, hopefully should work on other Ubuntu machines as well. `570-open` from Software Updates -&gt; Additional drivers → Apply -&gt; Restart worked for me.</p><p>The following steps are carried out in a clean Ubuntu20 clean build. One can give a try with an existing installation as well. But ensure you dont have existing nvidia* or libnvidia driver residuals.</p><p>And cant stress enough, Backup your data…</p><p>Reference: <a href="https://gist.github.com/MihailCosmin/affa6b1b71b43787e9228c25fe15aeba">https://gist.github.com/MihailCosmin/affa6b1b71b43787e9228c25fe15aeba</a></p><p>Make sure you don’t do `sudo apt autoremove` if you have ROS packages installed as they might be removed. Run only upto</p><blockquote>### to verify your gpu is cuda enable check<br>lspci | grep -i nvidia</blockquote><blockquote>### If you have previous installation remove it first. <br>sudo apt purge nvidia* -y<br>sudo apt remove nvidia-* -y<br>sudo rm /etc/apt/sources.list.d/cuda*<br>sudo apt autoclean -y<br>sudo rm -rf /usr/local/cuda*</blockquote><blockquote># system update<br>sudo apt update &amp;&amp; sudo apt upgrade -y</blockquote><blockquote># install other import packages<br>sudo apt install g++ freeglut3-dev build-essential libx11-dev libxmu-dev libxi-dev libglu1-mesa libglu1-mesa-dev</blockquote><blockquote># first get the PPA repository driver<br>sudo add-apt-repository ppa:graphics-drivers/ppa<br>sudo apt update</blockquote><blockquote># find recommended driver versions for you<br>ubuntu-drivers devices</blockquote><p>On a side note: Never do Ubuntu Version Updates via Ubuntu Software Installer (Example, 20 -&gt; 22 -&gt;24 )as it will result in issues. Recommended to do a Ubuntu version upgrade with live usb pendrive after backup data.</p><h3>1. Disable Secure Boot</h3><p>First and foremost reason for not properly installing nvidia drivers. For MSI → use the <strong>Search</strong> option at the top in BIOS.</p><h3>2. Update System</h3><pre>sudo apt update &amp;&amp; sudo apt upgrade -y<br>sudo apt install build-essential</pre><h3>3. Get the PPA repository driver</h3><pre>sudo add-apt-repository ppa:graphics-drivers/ppa<br>sudo apt update</pre><h3>4. Check drivers</h3><pre>sudo ubuntu-drivers devices</pre><p>Mostly ≥ 570 drivers will be listed. You can ignore nouveau unknown chip error.</p><p>Then go to <strong>Software Updates → Additional drivers</strong></p><ul><li><strong>Normal machine</strong> → test first with 570 proprietary, tested</li><li><strong>Server grade GPU (Founder)</strong> → try 570-server-open</li><li><strong>Latest GPUs</strong> where proprietary drivers aren’t stable → go with 570-open</li></ul><p>⚠️ If any issue is there → always test in a <strong>clean Ubuntu build</strong>.</p><h3>Miscellaneous</h3><p>If you <strong>couldn’t access keyboard or peripherals</strong> while in BIOS, update the grub:</p><pre>sudo nano /etc/default/grub</pre><p>Change these:</p><pre>GRUB_TIMEOUT=5    <br>GRUB_TIMEOUT_STYLE=menu<br>GRUB_CMDLINE_LINUX_DEFAULT=&quot;quiet splash usbcore.autosuspend=-1&quot;</pre><p>Then:</p><pre>sudo update-grub<br>sudo reboot now</pre><h3>2) dkms --config -a error</h3><p>Run this first:</p><pre>sudo dkms --config -a</pre><p>If still errors → install headers and retry:</p><pre>sudo apt install linux-headers-$(uname -r)<br>sudo apt --fix-broken install</pre><h3>Final Note on Kernel Updates</h3><p>While the following approaches may work fine (dated <strong>Mar 2025</strong>) → <a href="https://youtu.be/o5deOXLDpZw?feature=shared">YouTube Reference</a></p><p>Updating to a <strong>mainline kernel</strong> can break compatibility — for example, headers for <strong>kernel 6.13</strong> may fail like 6.13-header-generic not found because the mainline install gives 6.13-headers without generic suffix. I tried symlink of the same , it resulted in some dependency errors stating the current system kernel libraries are too old and I couldnt proceed further.<br> <br> If you really want to run a newer kernel, use the <strong>Mainline Kernel App</strong> so you can choose a specific version. But I had no luck with that running successfully.</p><p>👉 <strong>Best practice:</strong> avoid updating kernels unnecessarily.<br> If you must take the risk, <strong>backup your data twice</strong> before proceeding.</p><p>Also see: <a href="https://shirker.medium.com/installing-nvidia-rtx-50-series-drivers-on-debian-ubuntu-9836d35c8b94">Medium Reference</a><br> This shows an approach without touching kernel upgrades. It involves updating <strong>dkms-related packages</strong>.<br> ⚠️ Only try this method if you don’t want to do a clean build.</p><p>Keep <strong>clean build</strong> as the last (yet most reliable) option.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fae366d33073" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Medical Instance Segmentation with Torchvision — Mask RCNN Custom Finetuning.]]></title>
            <link>https://medium.com/@jaykumaran2217/medical-instance-segmentation-with-torchvision-mask-rcnn-custom-finetuning-c7b69ca6ea19?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/c7b69ca6ea19</guid>
            <category><![CDATA[segmentation]]></category>
            <category><![CDATA[mask-rcnn]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[medical]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Thu, 04 Apr 2024 11:56:02 GMT</pubDate>
            <atom:updated>2024-04-05T04:17:30.683Z</atom:updated>
            <content:encoded><![CDATA[<h3>Medical Instance Segmentation with Mask RCNN Custom Finetuning.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/988/1*nrgDoZ0EV3AsRCaAtrjJzQ.png" /><figcaption>Instance Segmentation Masks</figcaption></figure><p><strong>Mask R-CNN</strong> extends <a href="http://paperswithcode.com/method/faster-r-cnn">Faster R-CNN</a> to solve instance segmentation tasks. It achieves this by adding a branch for predicting an object mask in parallel with the existing branch for bounding box recognition. In principle, Mask R-CNN is an intuitive extension of Faster R-CNN, but constructing the mask branch properly is critical for good results.</p><p>Most importantly, Faster R-CNN was not designed for pixel-to-pixel alignment between network inputs and outputs. This is evident in how <a href="http://paperswithcode.com/method/roi-pooling">RoIPool</a>, the <em>de facto</em> core operation for attending to instances, performs coarse spatial quantization for feature extraction. To fix the misalignment, Mask R-CNN utilises a simple, quantization-free layer, called <a href="http://paperswithcode.com/method/roi-align">RoIAlign</a>, that faithfully preserves exact spatial locations.</p><p><strong>Dataset Overview:</strong><br>Dataset link:<br><a href="https://www.kaggle.com/datasets/ipateam/nuinsseg/data">https://www.kaggle.com/datasets/ipateam/nuinsseg/data</a></p><p>The NuInsSeg dataset contains more than 30k manually segmented nuclei from 31 human and mouse organs and 665 image patches extracted from H&amp;E-stained whole slide images. We also provide ambiguous area masks for the entire dataset to show in which areas manual semantic/instance segmentation were impossible.</p><p>Human organs:<br>cerebellum, cerebrum (brain), colon (rectum), epiglottis, jejunum, kidney, liver, lung, melanoma, muscle, oesophagus, palatine tonsil, pancreas, peritoneum, placenta, salivary gland, spleen, stomach (cardia), stomach (pylorus), testis, tongue, umbilical cord, and urinary bladder</p><p>Mouse organs:<br>cerebellum, cerebrum, colon, epiglottis, lung, melanoma, muscle, peritoneum, stomach (cardia), stomach (pylorus), testis, umbilical cord, and urinary bladder)</p><p>Let’s setup our training pipeline.</p><p><strong>Scripts:</strong></p><pre>!pip install -q kaggle<br>from google.colab import files<br>uploaded = files.upload()</pre><pre>%%capture<br>!mkdir ~/.kaggle<br>!mv kaggle.json ~/.kaggle/<br>!chmod 600 ~/.kaggle/kaggle.json</pre><pre>%%capture<br>!kaggle datasets download -d ipateam/nuinsseg<br>!unzip -q ./nuinsseg.zip -d ./Dataset-Nuclei</pre><pre>!pip install -q pycocotools</pre><pre>import os<br>import copy<br>import random<br>import json<br>import glob<br>import cv2<br>import numpy as np<br>import time<br>import matplotlib.pyplot as plt<br>import matplotlib.patches as patches<br>import requests<br>from   zipfile import ZipFile<br>import argparse<br>import shutil<br><br><br>from PIL import Image<br>from sklearn.model_selection import train_test_split<br><br>import torch<br>import torch.utils.data<br>from torch import nn<br>import torchvision<br><br><br>from torchvision import transforms as T<br><br>from pycocotools import mask as coco_mask<br>from pycocotools.coco import COCO</pre><p><strong>DATA PREPARATION</strong></p><pre>&quot;&quot;&quot;<br>Clean up of the Nuclei dataset<br>Can be downloaded from: https://www.kaggle.com/datasets/ipateam/nuinsseg?resource=download<br><br>The dataset comes with a lot of subdirectories that can be very useful depending<br>on the task of your interest. For our Mask RCNN and Ultralytics training, we just<br>need the raw images and labeled masks that we will convert to appropriate formats.<br>So let us delete all unwanted data and keep only the raw images, binary masks<br>and label masks.<br><br>Keep the directories with names &#39;tissue images&#39;, &#39;mask binary without border&#39;, and &#39;label masks modify&#39;<br><br>&quot;&quot;&quot;<br><br><br>def prune_subdirectories(base_dir,keep_dirs):<br>    #Iterate through each subdirs in base dir<br>    for root_dir in os.listdir(base_dir):<br>        root_path = os.path.join(base_dir,root_dir)<br>        if os.path.isdir(root_path):<br>          #  print(f&quot;Processing: {root_path}&quot;)<br>           #List all subdirs inside the current root dir<br>           for sub_dir in os.listdir(root_path):<br>               sub_path = os.path.join(root_path, sub_dir)<br>               #If the subdirectory isn&#39;t in the keep list, delete it<br>               if os.path.isdir(sub_path) and sub_dir not in keep_dirs:<br>                  # print(f&quot;Deleting: {sub_path}&quot;)<br>                  shutil.rmtree(sub_path)<br>               elif os.path.isdir(sub_path):<br>                  # print(f&quot;Keeping: {sub_path}&quot;)<br>                  pass<br><br><br>base_directory = &quot;Dataset-Nuclei&quot;<br>directories_to_keep = [&#39;tissue images&#39;, &#39;mask binary witthout border&#39;, &#39;label masks modify&#39;]<br><br>prune_subdirectories(base_directory,directories_to_keep)<br><br></pre><pre># Install Pip Tree globally<br>%%capture<br>!sudo apt install tree</pre><pre>!tree --dirsfirst -L 2  &quot;Dataset-Nuclei&quot;</pre><pre>Dataset-Nuclei<br>├── human bladder<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human brain<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human cardia<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human cerebellum<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human epiglottis<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human jejunum<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human kidney<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human liver<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human lung<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human melanoma<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human muscle<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human oesophagus<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human pancreas<br>│   ├── label masks modify<br>│   ├── tissue images<br>│   └── Thumbs.db<br>├── human peritoneum<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human placenta<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human pylorus<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human rectum<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human salivory gland<br>│   ├── label masks modify<br>│   ├── tissue images<br>│   └── Thumbs.db<br>├── human spleen<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human testis<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human tongue<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human tonsile<br>│   ├── label masks modify<br>│   └── tissue images<br>├── human umbilical cord<br>│   ├── label masks modify<br>│   └── tissue images<br>├── mouse fat (white and brown)_subscapula<br>│   ├── label masks modify<br>│   ├── tissue images<br>│   └── Thumbs.db<br>├── mouse femur<br>│   ├── label masks modify<br>│   └── tissue images<br>├── mouse heart<br>│   ├── label masks modify<br>│   ├── tissue images<br>│   └── Thumbs.db<br>├── mouse kidney<br>│   ├── label masks modify<br>│   ├── tissue images<br>│   └── Thumbs.db<br>├── mouse liver<br>│   ├── label masks modify<br>│   └── tissue images<br>├── mouse muscle_tibia<br>│   ├── label masks modify<br>│   └── tissue images<br>├── mouse spleen<br>│   ├── label masks modify<br>│   └── tissue images<br>└── mouse thymus<br>    ├── label masks modify<br>    └── tissue images<br><br>93 directories, 5 files</pre><pre>#Convert Label Masks to Coco Json Single File annotation<br><br>&quot;&quot;&quot;<br>With this code, we will convert our labeled mask image annotations to coco json<br>format so they can be used in training Mask R-CNN.<br><br>&quot;&quot;&quot;</pre><pre>def get_image_mask_pairs(data_dir):<br>    image_paths = []<br>    mask_paths = []<br><br>    for root,_,files in os.walk(data_dir):<br>        if &#39;tissue images&#39; in root:<br>            for file in files:<br>                if file.endswith(&#39;.png&#39;):<br>                    image_paths.append(os.path.join(root,file))<br>                    mask_paths.append(os.path.join(root.replace(&#39;tissue images&#39;,&#39;label masks modify&#39;), file.replace(&#39;.png&#39;,&#39;.tif&#39;)))<br>    return image_paths, mask_paths<br><br>def mask_to_polygons(mask,epsilon=1.0):<br>    contours,_ = cv2.findContours(mask,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)<br>    polygons = []<br>    for contour in contours:<br>        if len(contour) &gt; 2:<br>           poly = contour.reshape(-1).tolist()<br>           if len(poly) &gt; 4: #Ensures valid polygon<br>              polygons.append(poly)<br>    return polygons<br><br><br><br><br><br>def process_data(image_paths, mask_paths, output_dir,train=True):<br>    annotations = []<br>    images = []<br>    image_id = 0<br>    ann_id = 0<br><br>    for img_path, mask_path in zip(image_paths, mask_paths):<br>        image_id += 1<br>        img = cv2.imread(img_path)<br>        mask = cv2.imread(mask_path, cv2.IMREAD_UNCHANGED)<br><br>        # Copy image to output directory<br>        shutil.copy(img_path, os.path.join(output_dir, os.path.basename(img_path)))<br><br>        images.append({<br>            &quot;id&quot;: image_id,<br>            &quot;file_name&quot;: os.path.basename(img_path),<br>            &quot;height&quot;: img.shape[0],<br>            &quot;width&quot;: img.shape[1]<br>        })<br><br>        unique_values = np.unique(mask)<br>        for value in unique_values:<br>            if value == 0:  # Ignore background<br>                continue<br><br>            object_mask = (mask == value).astype(np.uint8) * 255<br>            polygons = mask_to_polygons(object_mask)<br><br>            for poly in polygons:<br>                ann_id += 1<br>                annotations.append({<br>                    &quot;id&quot;: ann_id,<br>                    &quot;image_id&quot;: image_id,<br>                    &quot;category_id&quot;: 1,  # Only one category: Nuclei<br>                    &quot;segmentation&quot;: [poly],<br>                    &quot;area&quot;: cv2.contourArea(np.array(poly).reshape(-1, 2)),<br>                    &quot;bbox&quot;: list(cv2.boundingRect(np.array(poly).reshape(-1, 2))),<br>                    &quot;iscrowd&quot;: 0<br>                })<br><br>    coco_input = {<br>        &quot;images&quot;: images,<br>        &quot;annotations&quot;: annotations,<br>        &quot;categories&quot;: [{&quot;id&quot;: 1, &quot;name&quot;: &quot;Nuclei&quot;}]<br>    }<br><br>    if train:<br>        with open(os.path.join(output_dir, &#39;instances_train2017.json&#39;), &#39;w&#39;) as f:<br>            json.dump(coco_input, f)<br>    else:<br>        with open(os.path.join(output_dir, &#39;instances_val2017.json&#39;), &#39;w&#39;) as f:<br>            json.dump(coco_input, f)<br><br>#instances_train2017.json</pre><pre>def main():<br>    data_dir = &#39;Dataset-Nuclei&#39;<br>    output_dir = &#39;Nuclei_Instance_COCO_input&#39;<br>    train_dir = os.path.join(output_dir, &#39;train2017&#39;) #we will stick with naming from official repo scripts<br>    val_dir = os.path.join(output_dir, &#39;val2017&#39;)<br><br>    # Create output directories<br>    os.makedirs(train_dir, exist_ok=True)<br>    os.makedirs(val_dir, exist_ok=True)<br><br>    image_paths, mask_paths = get_image_mask_pairs(data_dir)<br><br>    # Split data into train and val<br>    train_img_paths, val_img_paths, train_mask_paths, val_mask_paths = train_test_split(image_paths, mask_paths, test_size=0.2, random_state=42)<br><br>    # Process train and val data<br>    process_data(train_img_paths, train_mask_paths, train_dir,train=True)<br>    process_data(val_img_paths, val_mask_paths, val_dir,train=False)<br><br>if __name__ == &#39;__main__&#39;:<br>    main()</pre><pre>!mkdir ./Nuclei_Instance_COCO_input/annotations<br>!mv ./Nuclei_Instance_COCO_input/train2017/instances_train2017.json  ./Nuclei_Instance_COCO_input/annotations/<br>!mv ./Nuclei_Instance_COCO_input/val2017/instances_val2017.json  ./Nuclei_Instance_COCO_input/annotations/<br><br>!mkdir -p ./Nuclei_Instance_COCO_input/outputs/training/nuclei_instance_out<br>!mkdir -p ./Nuclei_Instance_COCO_input/outputs/inference/nuclei_instance_out</pre><pre>!tree --dirsfirst -L 1 &quot;Nuclei_Instance_COCO_input&quot;</pre><pre>Nuclei_Instance_COCO_input<br>├── annotations<br>├── outputs<br>├── train2017<br>└── val2017<br><br>4 directories, 0 files</pre><pre>!tree --dirsfirst -L 1 &quot;Nuclei_Instance_COCO_input/annotations&quot;</pre><pre>Nuclei_Instance_COCO_input/annotations<br>├── instances_train2017.json<br>└── instances_val2017.json<br><br>0 directories, 2 files</pre><pre>Visualise Coco Label - Dataset visualise - Only Contours<br><br><br>def display_images_with_coco_annotations(image_paths, annotations, display_type=&#39;both&#39;, filled=True):<br>    fig, axs = plt.subplots(2, 2, figsize=(10, 10))<br><br>    for ax, img_path in zip(axs.ravel(), image_paths):<br>        # Load image using OpenCV and convert it from BGR to RGB color space<br>        image = cv2.imread(img_path)<br>        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)<br><br>        ax.imshow(image)<br>        ax.axis(&#39;off&#39;)  # Turn off the axes<br><br>        # Get image filename to match with annotations<br>        img_filename = os.path.basename(img_path)<br>        img_id = next(item for item in annotations[&#39;images&#39;] if item[&quot;file_name&quot;] == img_filename)[&#39;id&#39;]<br><br>        # Filter annotations for the current image<br>        img_annotations = [ann for ann in annotations[&#39;annotations&#39;] if ann[&#39;image_id&#39;] == img_id]<br><br>        green_color = (0, 0,1)  # RGB for green<br><br>        for ann in img_annotations:<br>            # Display bounding box<br>            if display_type in [&#39;bbox&#39;, &#39;both&#39;]:<br>                bbox = ann[&#39;bbox&#39;]<br>                rect = patches.Rectangle((bbox[0], bbox[1]), bbox[2], bbox[3],<br>                                         linewidth=1, edgecolor=green_color,<br>                                         facecolor=&#39;none&#39;)<br>                ax.add_patch(rect)<br><br>            # Display segmentation polygon<br>            if display_type in [&#39;seg&#39;, &#39;both&#39;]:<br>                for seg in ann[&#39;segmentation&#39;]:<br>                    poly = [(seg[i], seg[i+1]) for i in range(0, len(seg), 2)]<br>                    if filled == False:<br>                        polygon = patches.Polygon(poly, closed=True, edgecolor=green_color, fill=False)<br>                    else:<br>                        polygon = patches.Polygon(poly, closed=True, edgecolor=green_color, fill=True, facecolor=green_color, alpha=0.7)<br><br>                    ax.add_patch(polygon)<br><br>    plt.tight_layout()<br>    plt.show()</pre><pre># Load COCO annotations<br>with open(&#39;Nuclei_Instance_COCO_input/annotations/instances_train2017.json&#39;,&#39;r&#39;) as f:<br>     annotations = json.load(f)<br><br># Get all image files<br>image_dir = &quot;Nuclei_Instance_COCO_input/train2017&quot;<br>all_images_files = [os.path.join(image_dir,img[&#39;file_name&#39;]) for img in annotations[&#39;images&#39;]]<br>random_image_files = random.sample(all_images_files,4)<br><br># Choose between &#39;bbox&#39;, &#39;seg&#39;, or &#39;both&#39;<br>display_type = &quot;seg&quot;<br>display_images_with_coco_annotations(random_image_files,annotations,display_type,filled=False)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/988/1*SNXodM3QfyHfN535A1LPNQ.png" /><figcaption>Visualization of Coco Annotations</figcaption></figure><p>From torchvision download training scripts for detection that we will be using for training Mask RCNN:<br><a href="https://github.com/pytorch/vision/tree/main/references/detection">https://github.com/pytorch/vision/tree/main/references/detection</a></p><p>Create folder scripts in current working directory of colab and have these torchvision scripts.</p><pre>class_names = &quot;&quot;&quot;<br>INSTANCE_CATEGORY_NAMES = [<br>    &#39;__background__&#39;,<br>    &#39;Nuclei&#39;,<br>]<br>&quot;&quot;&quot;<br><br>with open(&#39;class_names.py&#39;, &#39;w&#39;) as file:<br>    file.write(class_names)</pre><pre>!tree --dirsfirst -L 2 &quot;..&quot;</pre><pre>..<br>├── Nuclei_Instance_COCO_input<br>│   ├── annotations<br>│   ├── outputs<br>│   ├── train2017<br>│   └── val2017<br>├── scripts<br>│   ├── class_names.py<br>│   ├── coco_eval.py<br>│   ├── coco_utils.py<br>│   ├── engine.py<br>│   ├── group_by_aspect_ratio.py<br>│   ├── presets.py<br>│   ├── train.py<br>│   ├── transforms.py<br>│   └── utils.py<br>└── nuinsseg.zip<br><br>6 directories, 10 files</pre><p>In train.py script change the following :</p><pre>#train.py<br><br>import torch.nn as nn<br>from class_names import INSTANCE_CATEGORY_NAMES as class_names<br><br>#Next  to this line<br>model = torchvision.models.get_model(<br>        args.model, weights=args.weights, weights_backbone=args.weights_backbone, num_classes=num_classes, **kwargs<br>    )<br>model.roi_heads.box_predictor.cls_score = nn.Linear(in_features=1024, out_features=len(class_names),bias=True)<br>model.roi_heads.box_predictor.bbox_pred = nn.Linear(in_features=1024, out_features=len(class_names)*4,bias=True)<br>model.roi_heads.mask_predictor.mask_fcn_logits = nn.Conv2d(256, len(class_names),kernel_size=(1,1),stride=(1,1))<br><br>model.to(device)</pre><pre>!python train.py \<br>--data-path ../Nuclei_Instance_COCO_input \<br>--model maskrcnn_resnet50_fpn_v2 \<br>--weights MaskRCNN_ResNet50_FPN_V2_Weights.COCO_V1 \<br>--batch-size 4 \<br>--epochs 10 \<br>--lr 0.005 \<br>--output-dir ../Nuclei_Instance_COCO_input/outputs/training/nuclei_instance_out \<br>--amp \<br>--lr-steps 10 \<br>--data-augmentation lsj \<br>--use-copypaste \<br>--workers 0</pre><p><strong>INFERENCE</strong></p><pre>model = torchvision.models.detection.maskrcnn_resnet50_fpn_v2(pretrained=False)<br><br>from class_names import INSTANCE_CATEGORY_NAMES as class_names<br><br>model.roi_heads.box_predictor.cls_score = nn.Linear(in_features=1024,out_features=len(class_names),bias=True)<br>model.roi_heads.box_predictor.bbox_pred = nn.Linear(in_features=1024,out_features=len(class_names)*4,bias=True)<br>model.roi_heads.mask_predictor.mask_fcn_logits = nn.Conv2d(256,len(class_names),kernel_size=(1,1),stride=(1,1))</pre><pre>OUT_DIR = &quot;../Nuclei_Instance_COCO_input/outputs&quot;<br># ckpt = torch.load(os.path.join(OUT_DIR,&quot;training/nuclei_instance_out/checkpoint.pth&quot;))<br>ckpt = torch.load(&quot;../Nuclei_Instance_COCO_input/outputs/training/nuclei_instance_out/checkpoint.pth&quot;)<br>model.load_state_dict(ckpt[&#39;model&#39;])<br><br>device = torch.device(&#39;cuda&#39; if torch.cuda.is_available() else &#39;cpu&#39;)<br>model.to(device).eval()<br>print(model)</pre><pre>np.random.seed(0)<br><br>def get_outputs(image, model, threshold):<br>    # print(&quot;Get out Image shape&quot;,image.shape)<br>    with torch.no_grad():<br>        # forward pass of the image through the model.<br>        outputs = model(image)<br><br>    # get all the scores<br>    scores = list(outputs[0][&#39;scores&#39;].detach().cpu().numpy())<br>    # index of those scores which are above a certain threshold<br>    thresholded_preds_inidices = [scores.index(i) for i in scores if i &gt; threshold]<br>    thresholded_preds_count = len(thresholded_preds_inidices)<br>    # get the masks<br>    masks = (outputs[0][&#39;masks&#39;]&gt;0.5).squeeze().detach().cpu().numpy()<br>    # discard masks for objects which are below threshold<br>    masks = masks[:thresholded_preds_count]<br><br>    # get the bounding boxes, in (x1, y1), (x2, y2) format<br>    boxes = [[(int(i[0]), int(i[1])), (int(i[2]), int(i[3]))]  for i in outputs[0][&#39;boxes&#39;].detach().cpu()]<br>    # discard bounding boxes below threshold value<br>    boxes = boxes[:thresholded_preds_count]<br>    # get the classes labels<br>    labels = [class_names[i] for i in outputs[0][&#39;labels&#39;]]<br>    return masks, boxes, labels</pre><pre>np.random.seed(0)<br><br>def get_outputs(image, model, threshold):<br>    # print(&quot;Get out Image shape&quot;,image.shape)<br>    with torch.no_grad():<br>        # forward pass of the image through the model.<br>        outputs = model(image)<br><br>    # get all the scores<br>    scores = list(outputs[0][&#39;scores&#39;].detach().cpu().numpy())<br>    # index of those scores which are above a certain threshold<br>    thresholded_preds_inidices = [scores.index(i) for i in scores if i &gt; threshold]<br>    thresholded_preds_count = len(thresholded_preds_inidices)<br>    # get the masks<br>    masks = (outputs[0][&#39;masks&#39;]&gt;0.5).squeeze().detach().cpu().numpy()<br>    # discard masks for objects which are below threshold<br>    masks = masks[:thresholded_preds_count]<br><br>    # get the bounding boxes, in (x1, y1), (x2, y2) format<br>    boxes = [[(int(i[0]), int(i[1])), (int(i[2]), int(i[3]))]  for i in outputs[0][&#39;boxes&#39;].detach().cpu()]<br>    # discard bounding boxes below threshold value<br>    boxes = boxes[:thresholded_preds_count]<br>    # get the classes labels<br>    labels = [class_names[i] for i in outputs[0][&#39;labels&#39;]]<br>    return masks, boxes, labels</pre><pre>def draw_segmentation_map(image, masks, boxes, labels):<br>    alpha = 1.0<br>    beta = 0.5 # transparency for the segmentation map<br>    gamma = 0 # scalar added to each sum<br>    #convert the original PIL image into NumPy format<br>    image = np.array(image)<br>    # convert from RGN to OpenCV BGR format<br>    image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)<br>    for i in range(len(masks)):<br>        # apply a random color mask to each object<br>        color = [random.randint(0, 255) for _ in range(3)]<br>        if masks[i].any() == True:<br>            red_map = np.zeros_like(masks[i]).astype(np.uint8)<br>            green_map = np.zeros_like(masks[i]).astype(np.uint8)<br>            blue_map = np.zeros_like(masks[i]).astype(np.uint8)<br>            red_map[masks[i] == 1], green_map[masks[i] == 1], blue_map[masks[i] == 1] = color<br>            # combine all the masks into a single image<br>            segmentation_map = np.stack([red_map, green_map, blue_map], axis=2)<br>            # apply mask on the image<br>            cv2.addWeighted(image, alpha, segmentation_map, beta, gamma, image)<br><br>            # draw the bounding boxes around the objects<br>            cv2.rectangle(image, boxes[i][0], boxes[i][1], color=color,<br>                        thickness=2)<br>            # put the label text above the objects<br>            # cv2.putText(image , labels[i], (boxes[i][0][0], boxes[i][0][1]-10),<br>            #             cv2.FONT_HERSHEY_SIMPLEX, 1, color,<br>            #             thickness=2, lineType=cv2.LINE_AA)<br>    return image</pre><pre>validation_path = &quot;../Nuclei_Instance_COCO_input/val2017&quot;<br>image_paths = glob.glob(os.path.join(validation_path,&#39;*png&#39;))<br>transform = T.Compose([<br>    T.ToTensor()<br>])<br><br>fig = plt.figure(figsize=(10,25),layout=&quot;constrained&quot;)<br><br>for image_path in image_paths:<br>    # print(image_path)<br>    image = Image.open(image_path).convert(&quot;RGB&quot;)<br>    # keep a copy of the original image for OpenCV functions and applying masks<br>    orig_image = image.copy()<br><br>    image = transform(image)<br>    #add a batch dimension<br>    image = image.unsqueeze(0).to(device)<br>    # print(&quot;Image shape&quot;,image.shape)<br><br>    masks,boxes,labels = get_outputs(image,model,threshold=0.5)<br><br>    result = draw_segmentation_map(orig_image,masks,boxes,labels)<br><br>    save_path = f&quot;{OUT_DIR}/inference/nuclei_instance_out{image_path.split(os.path.sep)[-1].split(&#39;.&#39;)[0]}.jpg&quot;<br>    cv2.imwrite(save_path, result)<br><br>    #visualise images<br>    plt.imshow(result)<br>    plt.axis(&#39;off&#39;)<br><br>plt.show()</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/895/1*EMO98NFEHWN9pcBZKcxSdw.png" /><figcaption>Mask RCNN Prediction</figcaption></figure><p>Conclusion:</p><p>Checkout training notebook:</p><p>Git:</p><p><a href="https://github.com/Jaykumaran/Medical_Imaging_Referenced/blob/main/Nuclei_Instance_Segmentation/Torchvision_MaskRCNN.ipynb">https://github.com/Jaykumaran/Medical_Imaging_Referenced/blob/main/Nuclei_Instance_Segmentation/Torchvision_MaskRCNN.ipynb</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c7b69ca6ea19" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Prompt Engineering Resources and Guide]]></title>
            <link>https://medium.com/@jaykumaran2217/prompt-engineering-resources-and-guide-e98731c69c59?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/e98731c69c59</guid>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[prompt-engineering]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Sun, 07 Jan 2024 08:58:10 GMT</pubDate>
            <atom:updated>2024-01-07T08:58:10.844Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lH66vTM8xW_Xx4Z5UEUbuA.png" /></figure><p><strong>PROMPT ENGINEERING</strong></p><p><strong>Best Guide:</strong></p><p><a href="https://www.promptingguide.ai/">https://www.promptingguide.ai/</a></p><p><strong>26 Prompt Techniques:</strong></p><p><a href="https://www.youtube.com/watch?v=chAQGTBMXXQ">https://www.youtube.com/watch?v=chAQGTBMXXQ</a></p><p><strong>OpenAI Guide:</strong></p><p><a href="https://platform.openai.com/docs/guides/prompt-engineering/six-strategies-for-getting-better-results">https://platform.openai.com/docs/guides/prompt-engineering/six-strategies-for-getting-better-results</a></p><p><a href="https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-openai-api">https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-openai-api</a></p><p><a href="https://help.openai.com/en/collections/3675942-prompt-engineering">https://help.openai.com/en/collections/3675942-prompt-engineering</a></p><p><strong>Cookbook:</strong></p><p><a href="https://github.com/dair-ai/Prompt-Engineering-Guide">https://github.com/dair-ai/Prompt-Engineering-Guide</a></p><p><strong>Effectiveness of prompt Structure in Healthcare:</strong><br> <a href="https://arxiv.org/abs/2305.13614">https://arxiv.org/abs/2305.13614</a></p><p><strong>Best channel to learn LLM:</strong></p><p>Umar Jamil <a href="https://www.youtube.com/@umarjamilai">https://www.youtube.com/@umarjamilai</a> -Best For LLM Architectures</p><p>AI Anytime <a href="https://www.youtube.com/@AIAnytime">https://www.youtube.com/@AIAnytime</a></p><p>code_your_own_ai <a href="https://www.youtube.com/@code4AI">https://www.youtube.com/@code4AI</a></p><p>1littlecoder <a href="https://www.youtube.com/@1littlecoder">https://www.youtube.com/@1littlecoder</a></p><p>PromptEngineer <a href="https://www.youtube.com/@engineerprompt">https://www.youtube.com/@engineerprompt</a></p><p>Samwitteveenai <a href="https://www.youtube.com/@samwitteveenai">https://www.youtube.com/@samwitteveenai</a></p><p>James Briggs <a href="https://www.youtube.com/results?search_query=james+briggs">https://www.youtube.com/results?search_query=james+briggs</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e98731c69c59" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM Evaluators and Monitoring quick guide]]></title>
            <link>https://medium.com/@jaykumaran2217/llm-evaluators-and-monitoring-quick-guide-fa7a0cefb1b7?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa7a0cefb1b7</guid>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[evaluation]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Sun, 07 Jan 2024 08:55:10 GMT</pubDate>
            <atom:updated>2024-01-07T08:55:10.607Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>LLM EVALUATIONS:</strong></p><p><strong>RAGAS :</strong></p><p><strong>Docs</strong>: <a href="https://docs.ragas.io/en/latest/howtos/customisations/llms.html">https://docs.ragas.io/en/latest/howtos/customisations/llms.html</a></p><p><strong>Tutoria</strong>l: <a href="https://youtu.be/mEv-2Xnb_Wk?feature=shared">https://youtu.be/mEv-2Xnb_Wk?feature=shared</a></p><p><strong>Notebooks:</strong></p><p><a href="https://github.com/explodinggradients/ragas/tree/main/docs/howtos">https://github.com/explodinggradients/ragas/tree/main/docs/howtos</a></p><p><strong>With OpenAI</strong></p><p><a href="https://colab.research.google.com/drive/1TZo2sgf1YFzI4_U-tGppg_ylHAR3MXF_?usp=sharing">https://colab.research.google.com/drive/1TZo2sgf1YFzI4_U-tGppg_ylHAR3MXF_?usp=sharing</a></p><p><strong>With HF models and VLLM:</strong></p><p><a href="https://github.com/Jaykumaran/RAGAS-Langchain-Evaluation-VLLM-HFTextInference/blob/main/Working_Copy_of_Evaluating_Advanced_Retrieval_Methods_in_VLLM_%2B_LangChain_with_RAGAS%2BLiteLLM.ipynb">https://github.com/Jaykumaran/RAGAS-Langchain-Evaluation-VLLM-HFTextInference/blob/main/Working_Copy_of_Evaluating_Advanced_Retrieval_Methods_in_VLLM_%2B_LangChain_with_RAGAS%2BLiteLLM.ipynb</a></p><p><strong>LANGSMITH:</strong></p><p><strong>Docs:</strong> <a href="https://docs.smith.langchain.com/overview">https://docs.smith.langchain.com/overview</a></p><p><strong>ARIZE PHOENIX</strong>:</p><p><strong>Docs</strong>: <a href="https://docs.arize.com/phoenix/">https://docs.arize.com/phoenix/</a></p><p><strong>Blog:</strong></p><p><a href="https://arize.com/blog-course/llm-evaluation-the-definitive-guide/">https://arize.com/blog-course/llm-evaluation-the-definitive-guide/</a></p><p>****************************************************************************************************</p><p><strong>WEIGHTS AND BIASES:</strong></p><p><a href="https://wandb.ai/ayush-thakur/llm-eval-sweep/reports/How-to-Evaluate-Compare-and-Optimize-LLM-Systems--Vmlldzo0NzgyMTQz">https://wandb.ai/ayush-thakur/llm-eval-sweep/reports/How-to-Evaluate-Compare-and-Optimize-LLM-Systems--Vmlldzo0NzgyMTQz</a></p><p><strong>Other evaluation Blogs:</strong></p><p><a href="https://www.analyticsvidhya.com/blog/2023/05/how-to-evaluate-a-large-language-model-llm/">https://www.analyticsvidhya.com/blog/2023/05/how-to-evaluate-a-large-language-model-llm/</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fa7a0cefb1b7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLM Guardrails Ultimate Guide]]></title>
            <link>https://medium.com/@jaykumaran2217/llm-guardrails-ultimate-guide-4559d3a850ca?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/4559d3a850ca</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[guardrail]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Sun, 07 Jan 2024 08:53:35 GMT</pubDate>
            <atom:updated>2024-01-07T08:53:35.723Z</atom:updated>
            <content:encoded><![CDATA[<p>Guardrails Implementations and Cookbook</p><p><strong>GUARDRAILS</strong></p><p><strong>NEMOGUARDRAILS -Nvidia:</strong></p><p>Guardrails (or “rails” for short) are specific ways of controlling the output of a large language model, such as not talking about politics, responding in a particular way to specific user requests, following a predefined dialog path, using a particular language style, extracting structured data, and more.</p><p><strong>Read any Medium blog for free</strong>(as some blogs referred here are paywalled ):</p><p><a href="https://www.youtube.com/shorts/T52uUUPkRNI">https://www.youtube.com/shorts/T52uUUPkRNI</a></p><p><strong>Start with:</strong> Youtube Tutorial then ref docs and blogs</p><p><a href="https://youtu.be/SwqusllMCnE?feature=shared">https://youtu.be/SwqusllMCnE?feature=shared</a></p><p><a href="https://youtu.be/3Vg5bEtaepQ?feature=shared">https://youtu.be/3Vg5bEtaepQ?feature=shared</a></p><p><a href="https://youtu.be/alDcHItL6yo?feature=shared">https://youtu.be/alDcHItL6yo?feature=shared</a></p><p><strong>Above tutorial associated Notebooks:</strong></p><p><a href="https://www.pinecone.io/learn/nemo-guardrails-intro/">https://www.pinecone.io/learn/nemo-guardrails-intro/</a></p><p><a href="https://github.com/pinecone-io/examples/tree/master/learn/generation/chatbots/nemo-guardrails">https://github.com/pinecone-io/examples/tree/master/learn/generation/chatbots/nemo-guardrails</a></p><p><strong>Must read:</strong></p><p><a href="https://ksankar.medium.com/generative-ai-guardrails-functional-compositional-to-address-the-enhanced-treat-surface-43366685e466">https://ksankar.medium.com/generative-ai-guardrails-functional-compositional-to-address-the-enhanced-treat-surface-43366685e466</a></p><p><strong>Docs:</strong></p><p><a href="https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/docs">https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/docs</a></p><p><strong>NemoGuardrails Architecture Diagram:</strong></p><p><a href="https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/docs/architecture">https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/docs/architecture</a></p><p><strong>Setting up Pipelines for various HF models:</strong></p><p><a href="https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/examples/configs/llm">https://github.com/NVIDIA/NeMo-Guardrails/tree/develop/examples/configs/llm</a></p><p><strong>Colang or Canonical form vs Semantic:</strong></p><p>· Canonical language is the “how” of saying things, while semantic is the “what” of it.</p><p>· Canonical language is preferred for clarity and precision, while semantic analysis considers all possible meanings and interpretations.</p><p><strong>NemoGuardrails Library:</strong></p><p><a href="https://github.com/NVIDIA/NeMoGuardrails/blob/develop/docs/user_guides/guardrails-library.md">https://github.com/NVIDIA/NeMoGuardrails/blob/develop/docs/user_guides/guardrails-library.md</a></p><p><strong>llm_call function calling:</strong></p><p><a href="https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/nemoguardrails/actions/llm/utils.py">https://github.com/NVIDIA/NeMo-Guardrails/blob/develop/nemoguardrails/actions/llm/utils.py</a></p><p><strong>NemoGuardrails Implementations:</strong></p><p><strong>With OpenAI:</strong></p><p><a href="https://towardsdatascience.com/configuring-nemo-guardrails-your-way-an-alternative-method-for-large-language-models-c82aaff78f6e">https://towardsdatascience.com/configuring-nemo-guardrails-your-way-an-alternative-method-for-large-language-models-c82aaff78f6e</a></p><p><strong>RAG with OpenAI and NemoGuardrails:</strong></p><p><a href="https://medium.aiplanet.com/retrieval-augmented-pipeline-with-actions-using-nemo-gaurdrails-447b84a5334b">https://medium.aiplanet.com/retrieval-augmented-pipeline-with-actions-using-nemo-gaurdrails-447b84a5334b</a></p><p><strong>With Llama2:</strong></p><p><a href="https://blog.marvik.ai/2023/10/09/enhancing-llama2-conversations-with-nemo-guardrails-a-practical-guide/">https://blog.marvik.ai/2023/10/09/enhancing-llama2-conversations-with-nemo-guardrails-a-practical-guide/</a></p><p>Above blog repo: <a href="https://github.com/marvik-ai/llama2-nemo-guardrails">https://github.com/marvik-ai/llama2-nemo-guardrails</a></p><p>Issues faced with above implementation community discussion thread:</p><p><a href="https://github.com/NVIDIA/NeMo-Guardrails/issues/238">https://github.com/NVIDIA/NeMo-Guardrails/issues/238</a></p><p>My version of above repo fixed:</p><p><a href="https://github.com/Jaykumaran/NemoGuardrails/tree/main/GuardRAILS%20LLAMA2/llama2-nemo-guardrails">https://github.com/Jaykumaran/NemoGuardrails/tree/main/GuardRAILS%20LLAMA2/llama2-nemo-guardrails</a> #Read Markdown headings in notebooks carefully to understand setup.</p><p><strong>Guardrails AI:</strong></p><p><strong>Docs</strong>: <a href="https://www.guardrailsai.com/docs">https://www.guardrailsai.com/docs</a></p><p><strong>Example Notebooks:</strong></p><p><a href="https://github.com/guardrails-ai/guardrails/tree/main/docs/examples">https://github.com/guardrails-ai/guardrails/tree/main/docs/examples</a></p><p><strong>SEMANTIC ROUTER: (Different from Guardrails)</strong></p><p>Tutorial: <a href="https://www.youtube.com/watch?v=ro312jDqAh0">https://www.youtube.com/watch?v=ro312jDqAh0</a></p><p>Notebooks: <a href="https://github.com/aurelio-labs/semantic-router/tree/main/docs">https://github.com/aurelio-labs/semantic-router/tree/main/docs</a></p><p>Blogs: <a href="https://medium.com/ai-insights-cobet/beyond-basic-chatbots-how-semantic-router-is-changing-the-game-783dd959a32d">https://medium.com/ai-insights-cobet/beyond-basic-chatbots-how-semantic-router-is-changing-the-game-783dd959a32d</a></p><p><strong>LLAMA GUARD: (Purple Llama)</strong></p><p><a href="https://towardsdatascience.com/safeguarding-your-rag-pipelines-a-step-by-step-guide-to-implementing-llama-guard-with-llamaindex-6f80a2e07756">https://towardsdatascience.com/safeguarding-your-rag-pipelines-a-step-by-step-guide-to-implementing-llama-guard-with-llamaindex-6f80a2e07756</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4559d3a850ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[VLLM Serving Common Errors Compilation for RAGAS evaluation: Alternative to OpenAI Api key]]></title>
            <link>https://medium.com/@jaykumaran2217/vllm-serving-common-errors-compilation-for-ragas-evaluation-alternative-to-openai-api-key-8aa020e979a3?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/8aa020e979a3</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[vllm]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Thu, 28 Dec 2023 09:58:28 GMT</pubDate>
            <atom:updated>2023-12-28T10:06:09.761Z</atom:updated>
            <content:encoded><![CDATA[<h3>VLLM free Serving Common Errors Compilation for RAGAS evaluation: Alternative to OpenAI Api key</h3><p>Working Implementation Code :<a href="https://bitly.ws/37ECz">https://bitly.ws/37Ezo</a></p><p><a href="https://github.com/Jaykumaran/RAGAS-Langchain-Evaluation-VLLM-HFTextInference">GitHub - Jaykumaran/RAGAS-Langchain-Evaluation-VLLM-HFTextInference: Implementation of RAGAS Evaluation using HuggingFaceModels</a></p><p>Errors I faced during succesful Vllm serving to use with ChatOpenAI</p><ul><li><a href="https://github.com/vllm-project/vllm/issues/1479">Running out of memory with TheBloke/CodeLlama-7B-AWQ · Issue #1479 · vllm-project/vllm</a></li><li><a href="https://pytorch.org/get-started/previous-versions/">Previous PyTorch Versions</a></li><li><a href="https://docs.vllm.ai/en/latest/getting_started/quickstart.html">Quickstart - vLLM</a></li><li><a href="https://docs.ragas.io/en/latest/howtos/customisations/llms.html">Bring your own LLMs | Ragas</a></li></ul><p><a href="https://www.youtube.com/results?search_query=vllm+evaluation">https://www.youtube.com/results?search_query=vllm+evaluation</a></p><p><a href="https://python.langchain.com/docs/integrations/llms/huggingface_textgen_inference">Huggingface TextGen Inference | 🦜️🔗 Langchain</a></p><p><a href="https://youtu.be/6CRrhxpF8WI?feature=shared">https://youtu.be/6CRrhxpF8WI?feature=shared</a></p><ul><li><a href="https://docs.litellm.ai/docs/providers/vllm">VLLM | liteLLM</a></li><li><a href="https://python.langchain.com/docs/integrations/llms/huggingface_textgen_inference">Huggingface TextGen Inference | 🦜️🔗 Langchain</a></li><li><a href="https://github.com/explodinggradients/ragas/issues/398">Issue with &quot;Bring Your Own LLMs&quot; · Issue #398 · explodinggradients/ragas</a></li><li><a href="https://vilsonrodrigues.medium.com/serving-falcon-models-with-text-generation-inference-tgi-5f32005c663b">Serving Falcon models with 🤗 Text Generation Inference (TGI)</a></li></ul><p><a href="https://private-user-images.githubusercontent.com/3515260/287610399-7ee90ac2-c5c1-4e79-9d36-ad00c1e24b38.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.LsAkXsTRMsqabTexfziNA-cgy8O8CcVWaXmE3SnlHTs">https://private-user-images.githubusercontent.com/3515260/287610399-7ee90ac2-c5c1-4e79-9d36-ad00c1e24b38.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.LsAkXsTRMsqabTexfziNA-cgy8O8CcVWaXmE3SnlHTs</a></p><ul><li><a href="https://medium.com/@thinhda/streamlining-llm-applications-with-litellm-proxy-a-comprehensive-guide-875122c00974">Streamlining LLM Applications with LiteLLM Proxy: A Comprehensive Guide</a></li><li><a href="https://github.com/vllm-project/vllm/issues/549#issue-1817047555">AssertionError: data parallel group is already initialized · Issue #549 · vllm-project/vllm</a></li></ul><p><a href="https://colab.research.google.com/github/explodinggradients/ragas/blob/main/docs/howtos/customisations/llms.ipynb#scrollTo=85e313f2-e45c-4551-ab20-4e526e098740">https://colab.research.google.com/github/explodinggradients/ragas/blob/main/docs/howtos/customisations/llms.ipynb#scrollTo=85e313f2-e45c-4551-ab20-4e526e098740</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8aa020e979a3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[YoloV8 Custom PoseKeypoint Detection Tutorial:]]></title>
            <link>https://medium.com/@jaykumaran2217/yolov8-custom-posekeypoint-detection-tutorial-04be87ee915e?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/04be87ee915e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[robotics-automation]]></category>
            <category><![CDATA[yolo]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Tue, 03 Oct 2023 06:35:33 GMT</pubDate>
            <atom:updated>2023-12-30T07:02:46.839Z</atom:updated>
            <content:encoded><![CDATA[<h3>YoloV8 Custom PoseKeypoint Detection Tutorial:Robotic Arm Keypoint Pose Tracking : Yolo Custom Pose Keypoint Training</h3><h3>Enhancing Chess-Playing Robots with Custom Pose Keypoint Detection</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/240/1*l_n8rBw7vzrpX8h6XtrTPQ.gif" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/639/1*bSB6Skxdyr6xFPdsU6TyDw.png" /></figure><p><strong>Git:</strong><a href="https://github.com/Jaykumaran/Robotic_Arm_Keypoint_Tracking_YoloV8-Pose/tree/main">https://github.com/Jaykumaran/Robotic_Arm_Keypoint_Tracking_YoloV8-Pose/tree/main</a></p><p><strong>Introduction</strong></p><p>Chess has long been a game of strategy and intellect, challenging players to think several moves ahead. In recent years, advancements in robotics and artificial intelligence have given rise to chess-playing robots capable of competing against human opponents. However, to truly excel in the game, these robots need more than just the ability to move chess pieces. They require a deeper understanding of the game board, the positions of the pieces, and the ability to plan and execute complex maneuvers. This is where custom pose keypoint detection comes into play.</p><p><strong>Understanding Custom Pose Keypoint Detection</strong></p><p>Custom pose keypoint detection is a computer vision technique that involves identifying and tracking specific points, or keypoints, on an object. In the case of a chess-playing robot arm, these keypoints could represent the positions of chess pieces, the orientation of the chessboard, or even the configuration of the robot arm itself.</p><p><strong>Annotation using CVAT:</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FkOlEC30hK7o%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DkOlEC30hK7o&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FkOlEC30hK7o%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f773188df3d7a4b950096c0996ed81c3/href">https://medium.com/media/f773188df3d7a4b950096c0996ed81c3/href</a></iframe><p><strong>Step by Step:</strong></p><p><strong>Step1 :</strong> Create a Project-&gt;Task-&gt;Import Dataset</p><p><strong>Step2:</strong> Create a Skeleton for the Project ,this helps to avoid tireless process of annotating individual images.Make sure all of your images are of dimensions 640x640 as YOLO Pose training accepts only this dimensions.(Important this is done before annotation import in CVAT).U can use batch img online resizers like by unchecking lock aspect ratio.</p><p><a href="https://redketchup.io/bulk-image-resizer">Bulk Image Resizer - Edit Multiple Images Online | RedKetchup</a></p><p><strong>Step3:</strong> At bottom there will be option called Skeleton choose that to your current image and adjust points to your need.</p><p><strong>Step4:</strong> Click save and go to task section and Export dataset for task in CocoKeypoints Format.</p><p><strong>Step5:</strong> Arrange your images in format train-&gt; images and labels</p><p><strong>Step6:</strong> Run your json in given python script to convert Coco.json to YOLO format.This automatically gives Normalised Annotation.</p><p>Normalised keypoints should be within range of 0 to 1 .Otherwise training wouldnt happen with Yolo Pose</p><pre>import json<br>import os<br><br>def convert_coco_to_yolo(coco_json_path, output_dir, image_width, image_height):<br>    # Load COCO JSON file<br>    with open(coco_json_path, &#39;r&#39;) as f:<br>        coco_data = json.load(f)<br><br>    # Create output directory if it doesn&#39;t exist<br>    if not os.path.exists(output_dir):<br>        os.makedirs(output_dir)<br><br>    # Iterate over each image in the dataset<br>    for image_data in coco_data[&#39;images&#39;]:<br>        image_id = image_data[&#39;id&#39;]<br>        image_name = image_data[&#39;file_name&#39;]<br>        keypoints_list = []<br><br>        # Find annotations for the current image<br>        for annotation in coco_data[&#39;annotations&#39;]:<br>            if annotation[&#39;image_id&#39;] == image_id:<br>                keypoints = annotation[&#39;keypoints&#39;]<br>                keypoints_list.append(keypoints)<br><br>        # Skip images without annotations<br>        if not keypoints_list:<br>            continue<br><br>        # Create YOLO annotation file<br>        annotation_file_name = os.path.splitext(image_name)[0] + &#39;.txt&#39;<br>        annotation_file_path = os.path.join(output_dir, annotation_file_name)<br>        with open(annotation_file_path, &#39;w&#39;) as f:<br>            for keypoints in keypoints_list:<br>                # Find bounding box coordinates<br>                x_min = min(keypoints[0::3])<br>                y_min = min(keypoints[1::3])<br>                x_max = max(keypoints[0::3])<br>                y_max = max(keypoints[1::3])<br><br>                # Normalize bounding box coordinates to range [0, 1]<br>                x_center = (x_min + x_max) / (2 * image_width)<br>                y_center = (y_min + y_max) / (2 * image_height)<br>                width = (x_max - x_min) / image_width<br>                height = (y_max - y_min) / image_height<br><br>                # Write the annotation to the YOLO file<br>                f.write(f&#39;{0} {round(x_center, 6)} {round(y_center, 6)} {round(width, 6)} {round(height, 6)} &#39;)<br><br>                # Append normalized keypoints to the annotation<br>                for i in range(0, len(keypoints), 3):<br>                    x = round(keypoints[i] / image_width, 6)<br>                    y = round(keypoints[i + 1] / image_height, 6)<br>                    v = round(keypoints[i + 2], 6)<br>                    f.write(f&#39;{x} {y} {v} &#39;)<br>                f.write(&#39;\n&#39;)<br><br>    print(&#39;Conversion complete.&#39;)<br><br># Example usage<br>coco_json_path = &quot;C:\\Users\\jaikr\\Downloads\\Subset\\annotations\\person_keypoints_default.json&quot;<br>output_dir = &quot;C:\\Users\\jaikr\\Downloads\\Subset640&quot;<br>image_width = 640 #Recommended for YOLOV8_Pose<br>image_height = 640 #Recommended for YOLOV8_Pose<br>print(output_dir)<br>convert_coco_to_yolo(coco_json_path, output_dir, image_width, image_height)</pre><pre>#Output format expected for eg img.txt files for each img<br>0 0.485896 0.332236 0.255865 0.248009 0.357964 0.351019<br>2 0.545229 0.208231 2 0.613828 0.456241 2<br>#where <br>0:class name<br>0.485896 0.332236 0.255865 0.248009 :Bounding boxes<br>0.357964 0.351019 2 #first keypoint[0] x,y,visible(whether keypoint will be visible<br>0.545229 0.208231 2 kpt[1]<br>0.613828 0.456241 2 kpt[2]</pre><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FSxSXIrUiJ-g%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DSxSXIrUiJ-g&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FSxSXIrUiJ-g%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/02dd8fe37e38cb3ee9039f1f44fd322e/href">https://medium.com/media/02dd8fe37e38cb3ee9039f1f44fd322e/href</a></iframe><h4>Step7: Training in colab using</h4><pre>Folder structure should follow<br><br>Main dir<br> -Train<br>   -images<br>   -labels<br><br> -Val<br>   -images<br>   -labels</pre><pre>!pip install ultralytics<br><br>from ultralytics import YOLO<br><br># Load a model<br><br>model = YOLO(&#39;yolov8n-pose.pt&#39;)  # load a pretrained model (recommended for training)<br><br><br># Train the model<br>results = model.train(data=&#39;data.yaml&#39;, epochs=100, imgsz=640)</pre><pre>Data.yaml should be of this:<br><br>#change this to your file structure<br><br>train: /content/RobotArm/train/images<br>val: /content/RobotArm/val/images<br>test: /content/RobotArm/test/images<br>nc: 1<br>names:<br> 0: joints<br># Keypoints<br>kpt_shape: [3, 3]  # number of keypoints,number of dims (2 for x,y or 3 for x,y,visible)<br>#( in my robot arm its 3 kpts,and 3 for visible kpts)</pre><pre><br><br>```<br>names:<br>    0: joints<br>```<br>  <br> I ve used number/index as if multiple classes were there means we can augment<br><br>```<br>names:<br>     0: joints<br>     1: skeleton<br>     2: face_keypoints<br>```<br>  <br> For proper mapping of classes used while annoattaion and assigning the,m while training is important in multi class.</pre><h4><strong>Prediction using best.pt in local:</strong></h4><pre>import matplotlib.pyplot as plt<br>from ultralytics import YOLO<br>from PIL import Image, ImageDraw<br><br>model = YOLO(&quot;C:\\Users\\jaikr\\Downloads\\Final\\best.pt&quot;)<br>results = model.predict(source=&quot;C:\\Users\\jaikr\\Downloads\\Subset640\\train\\images\\WIN_20230915_20_27_08_Pro.jpg&quot;)<br><br>for r in results:<br>    print(r.keypoints)<br><br>    # this line is changed<br>    keypoints = r.keypoints.xy.int().numpy()  # get the keypoints<br>    img_array = r.plot(kpt_line=True, kpt_radius=6)  # plot a BGR array of predictions<br>    im = Image.fromarray(img_array[..., ::-1])  # Convert array to a PIL Image<br><br>    draw = ImageDraw.Draw(im)<br>    draw.line([(keypoints[0][0][0], keypoints[0][0][1]), (keypoints[0][1][0],<br>            keypoints[0][1][1]), (keypoints[0][2][0], keypoints[0][2][1])],<br>             fill=(0, 0,255), width=5)<br>    im.show()<br><br>#change the keypoints order and no.of keypoints accordingly<br>#As keypoints returns a tensor in r.keypoints we convert extract the kptpoints</pre><p>By Stacking Images to Video Sequence:</p><pre>import os<br>import cv2<br>from ultralytics import YOLO<br>from PIL import Image, ImageDraw<br>import subprocess<br>import time  # Add the time module<br><br>model = YOLO(&quot;C:\\Users\\jaikr\\Downloads\\Final\\best.pt&quot;)<br><br># Folder containing input images<br>input_folder = &quot;C:\\Users\\jaikr\\Downloads\\Final\\train\\images&quot;<br>output_folder = &quot;output_images&quot;  # Output folder for saving images<br>output_video = &quot;output_video.mp4&quot;  # Output video file name<br><br># Create the output folder if it doesn&#39;t exist<br>os.makedirs(output_folder, exist_ok=True)<br><br># Get a list of image files in the input folder<br>image_files = [f for f in os.listdir(input_folder) if f.endswith((&quot;.jpg&quot;, &quot;.jpeg&quot;, &quot;.png&quot;))]<br><br># Sort the image files to maintain order<br>image_files.sort()<br><br>for idx, image_file in enumerate(image_files):<br>    image_path = os.path.join(input_folder, image_file)<br><br>    try:<br>        results = model.predict(source=image_path)<br><br>        for r in results:<br>            print(r.keypoints)<br><br>            # This line is changed<br>            keypoints = r.keypoints.xy.int().numpy()  # Get the keypoints<br>            img_array = r.plot(kpt_line=True, kpt_radius=6)  # Plot a BGR array of predictions<br>            im = Image.fromarray(img_array[..., ::-1])  # Convert array to a PIL Image<br><br>            draw = ImageDraw.Draw(im)<br>            draw.line([(keypoints[0][0][0], keypoints[0][0][1]), (keypoints[0][1][0],<br>                                                                  keypoints[0][1][1]),<br>                       (keypoints[0][2][0], keypoints[0][2][1])], fill=(255, 0, 0), width=5)<br><br>            # Save the image with a sequence number<br>            output_path = os.path.join(output_folder, f&quot;output_image_{idx:04d}.png&quot;)<br>            im.save(output_path)<br><br>        print(f&quot;Processed image &#39;{image_file}&#39;.&quot;)<br><br>    except Exception as e:<br>        print(f&quot;Error processing image &#39;{image_file}&#39;: {e}&quot;)<br>        continue  # Continue to the next image if an error occurs<br><br>print(&quot;Image processing completed.&quot;)<br><br># Use OpenCV to create a video from the saved images with a delay of 0.5 seconds<br>frame_array = []<br>for i in range(len(image_files)):<br>    img_path = os.path.join(output_folder, f&quot;output_image_{i:04d}.png&quot;)<br><br>    # Check if the image file exists<br>    if not os.path.exists(img_path):<br>        print(f&quot;Image &#39;{img_path}&#39; not found or has an issue with format. Skipping.&quot;)<br>        continue<br><br>    img = cv2.imread(img_path)<br>    height, width, layers = img.shape<br>    size = (width, height)<br>    frame_array.append(img)<br><br>out = cv2.VideoWriter(output_video, cv2.VideoWriter_fourcc(*&#39;mp4v&#39;), 30, size)<br><br>for i in range(len(frame_array)):<br>    out.write(frame_array[i])<br>    time.sleep(1)  # Add a delay of 0.5 seconds between frames<br><br>out.release()<br><br>print(f&quot;Video &#39;{output_video}&#39; created successfully.&quot;)<br><br></pre><h3>The YOLO Workflow</h3><ol><li><strong>Input Image:</strong> YOLO takes an image as its input. This image can be anything from a photograph to a medical scan or a video frame.</li><li><strong>Convolutional Neural Network (CNN):</strong> The input image is processed through a deep Convolutional Neural Network (CNN). This CNN plays a crucial role in feature extraction and understanding the content of the image.</li><li><strong>Grid Division:</strong> YOLO divides the processed image into a grid. Each cell within this grid is responsible for making predictions about objects present within its boundaries. This grid-based approach enables efficient and localized object detection.</li><li><strong>Bounding Boxes:</strong> Within each grid cell, YOLO predicts bounding boxes (rectangles) that tightly enclose objects. These bounding boxes specify the position and size of the detected objects.</li><li><strong>Class Predictions:</strong> Simultaneously, YOLO predicts the class of each object contained within the bounding box. For instance, in medical imaging, it might predict whether the object is a bone fracture, joint, or normal tissue.</li><li><strong>Confidence Scores: </strong>YOLO assigns confidence scores to each predicted bounding box. These scores represent the model’s confidence in the accuracy of the prediction. High confidence scores indicate a high probability of the bounding box containing a valid object</li><li><strong>Non-Maximum Suppression:</strong> To refine the results and eliminate duplicate or overlapping predictions, YOLO employs a technique called non-maximum suppression. This step ensures that only the most accurate and relevant predictions are retained.</li><li><strong>Output:</strong> The final output of YOLO consists of bounding boxes, their associated object classes, and confidence scores. These results provide a comprehensive understanding of the objects detected within the input image.</li></ol><p><strong>Implementing Custom Pose Keypoint Detection</strong></p><p>Implementing custom pose keypoint detection for a chess-playing robot arm involves several key steps:</p><ol><li><strong>Data Collection:</strong> Gather a dataset of images or videos featuring different chessboard configurations, pieces, and robot arm positions. Annotate the keypoints of interest in the data.</li><li><strong>Model Training: </strong>Train a custom pose keypoint detection model using deep learning frameworks like PyTorch or TensorFlow. The model should learn to predict the keypoints based on the input images or video frames.</li><li><strong>Integration:</strong> Integrate the trained model into the robot’s software architecture. Ensure that it can process real-time video input from the robot’s cameras and provide accurate keypoint predictions.</li><li><strong>Testing and Refinement</strong>: Test the robot’s performance in chess games and fine-tune the keypoint detection model as needed. Continuous refinement is essential for optimal performance.</li></ol><h3>Significance of YOLO</h3><p>The YOLO object detection system has had a profound impact on various fields, including:</p><h3>1. Autonomous Vehicles</h3><p>In the realm of self-driving cars, YOLO plays a crucial role in identifying and tracking objects on the road, such as pedestrians, vehicles, traffic signs, and lane boundaries. This real-time object detection capability is fundamental to the safety and functionality of autonomous vehicles.</p><h3>2. Surveillance and Security</h3><p>Security systems rely on YOLO for detecting intruders, recognizing faces, and monitoring activities in real-time. YOLO’s ability to swiftly and accurately identify objects in surveillance footage enhances security measures.</p><h3>3. Medical Imaging</h3><p>In medical imaging, YOLO assists in the detection of anomalies and diseases by analyzing X-rays, MRIs, and CT scans. It can identify and localize specific regions of interest, enabling faster and more precise diagnoses.</p><h3>4. Industrial Automation</h3><p>Manufacturing and logistics industries use YOLO for quality control, object tracking, and inventory management. It streamlines processes by automating object detection tasks.</p><h3>5. Wildlife Conservation</h3><p>In conservation efforts, YOLO aids in wildlife monitoring and protection. It can identify and track animals in their natural habitats, contributing to research and preservation initiatives.</p><p><strong>Why is Custom Pose Keypoint Detection Necessary for Chess-Playing Robots?</strong></p><ol><li>Precise Piece Identification: Chess-playing robots must be able to precisely identify the positions of each chess piece on the board. Custom pose keypoint detection allows the robot to recognize the type and location of each piece, enabling it to make informed decisions during the game.</li><li>Planning Complex Moves: Chess is a game of complex strategies and tactics. To plan its moves effectively, a robot needs to understand the current state of the board, potential threats, and possible future positions. Custom pose keypoint detection provides the robot with the necessary data to analyze the board and strategize accordingly.</li><li>Avoiding Collisions: Robot arms can have multiple joints and a wide range of motion. To avoid collisions with the chessboard, other pieces, or even itself, the robot arm needs real-time awareness of its own configuration. Custom pose keypoint detection helps the robot monitor the positions of its joints and ensure safe and efficient movements.</li><li>Adaptability: Chess is a game with various board sizes and piece configurations. Custom pose keypoint detection can be adapted to different chessboard sizes and piece sets, making the robot versatile and capable of playing different variants of chess.</li></ol><p><strong>Conclusion</strong></p><p>Custom pose keypoint detection is a crucial component in enhancing the capabilities of a chess-playing robot arm. It enables precise piece identification, complex move planning, collision avoidance, and adaptability to different chess variants. By investing in this technology, we can create chess-playing robots that not only challenge human players but also deepen our understanding of robotics and artificial intelligence. As technology continues to advance, we can expect to see even more impressive feats from these chess-playing robots in the future.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=04be87ee915e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Revolutionizing Cricket Analytics with Computer Vision: The 3rd Umpire Wicket Tracking System]]></title>
            <link>https://medium.com/@jaykumaran2217/revolutionizing-cricket-analytics-with-computer-vision-the-3rd-umpire-wicket-tracking-system-b4fa13c2acad?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/b4fa13c2acad</guid>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[yolo]]></category>
            <category><![CDATA[sports-analytics]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Mon, 02 Oct 2023 04:43:23 GMT</pubDate>
            <atom:updated>2023-10-02T04:53:10.629Z</atom:updated>
            <content:encoded><![CDATA[<p>Author: <a href="https://www.linkedin.com/in/jayakumaran-r17/">https://www.linkedin.com/in/jayakumaran-r17/</a></p><p>Git: <a href="https://github.com/Jaykumaran/Cricket_3rd_Umpire_Stumping_Tracking">https://github.com/Jaykumaran/Cricket_3rd_Umpire_Stumping_Tracking</a></p><p><a href="https://www.linkedin.com/posts/jayakumaran-r17_cricket-stumping-out-3rd-umpire-tracking-activity-7114447007246663680-YcxZ?utm_source=share&amp;utm_medium=member_desktop">Jaykumaran R on LinkedIn: Cricket Stumping Out 3rd Umpire Tracking- Ind Vs Aus Wicket...</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UFVM0J61UqhYGoYVcuCCTQ.png" /></figure><p><strong>Introduction</strong></p><p>Cricket is a sport that has evolved significantly over the years, incorporating technology to ensure fair play and accurate decision-making. One of the most critical aspects of cricket is making decisions about wickets, and technology has played a pivotal role in this area. The advent of the 3rd umpire, aided by computer vision technology, has not only enhanced the accuracy of decision-making but also added an exciting dimension to the game.</p><p>In this blog post, we will delve into the world of 3rd umpire wicket tracking using computer vision and explore how this technology is revolutionizing the game of cricket.</p><p><strong>The Role of the 3rd Umpire</strong></p><p>Traditionally, cricket matches relied solely on the on-field umpires to make crucial decisions, including determining whether a batsman is out or not. However, with the introduction of the 3rd umpire, a new level of precision and fairness has been introduced into the game. The 3rd umpire is equipped with various technological tools, including ball-tracking and hotspot cameras, but one of the most significant advancements has been the integration of computer vision for wicket tracking.</p><p><strong>Computer Vision in Wicket Tracking</strong></p><p>Computer vision is a field of artificial intelligence that enables machines to interpret and understand visual information from the world around them. In cricket, computer vision is used to track the trajectory of the ball and its interaction with the batsman and the stumps. Here’s how it works:</p><ol><li><strong>Ball Tracking:</strong> High-speed cameras capture the entire trajectory of the ball from the moment it leaves the bowler’s hand until it reaches the batsman. Computer vision algorithms analyze these images to determine the exact path of the ball. This technology is instrumental in making LBW (Leg Before Wicket) decisions, where the ball’s trajectory in relation to the batsman’s leg is crucial.</li><li><strong>Stump Cameras:</strong> Cameras are strategically placed behind the stumps to monitor the bails and the crease. Computer vision algorithms continuously track the position of the bails and determine whether they have been dislodged by the bowler or the wicketkeeper before the batsman can ground his bat.</li><li><strong>Snicko and UltraEdge:</strong> These technologies use audio and visual data, respectively, to detect faint edges or nicks off the bat. Computer vision plays a role in analyzing the visual data and helping the 3rd umpire make accurate decisions regarding catches.</li></ol><p><strong>Benefits of Computer Vision in Cricket</strong></p><ol><li><strong>Precision:</strong> Computer vision technology provides an unprecedented level of precision in making wicket-related decisions. It can accurately track the ball’s path, the position of the bails, and any possible edges off the bat.</li><li><strong>Eliminating Human Error</strong>: While on-field umpires are highly skilled, they can still make mistakes due to the speed and complexity of the game. Computer vision helps reduce the margin of error, leading to more reliable decisions.</li><li><strong>Transparency</strong>: Fans and players alike appreciate the transparency that technology brings to the game. With replays and visual evidence, cricket enthusiasts can better understand and accept the umpire’s decisions.</li><li><strong>Fairness:</strong> Computer vision ensures that every decision is made based on concrete evidence. This fairness is crucial in maintaining the integrity of the game.</li></ol><p><strong>Challenges and Controversies</strong></p><p>Despite its many advantages, computer vision technology in cricket has not been without its challenges and controversies. Some critics argue that it can slow down the pace of the game, disrupt the flow of the match, and sometimes lead to subjective interpretations by the 3rd umpire. Additionally, there have been instances where the technology’s accuracy has been questioned.</p><p>It’s important to address the challenges and controversies surrounding computer vision technology in cricket to get a comprehensive view of its impact on the game.</p><ol><li><strong>Slow Decision-Making:</strong> Critics argue that the use of technology can sometimes lead to slow decision-making. The process of reviewing a decision, especially in high-stakes situations, can disrupt the flow of the game and test the patience of players and spectators. Striking the right balance between maintaining the pace of play and ensuring accuracy is an ongoing challenge for cricket authorities.</li><li><strong>Subjective Interpretations:</strong> Despite the use of technology, there are instances where the 3rd umpire’s interpretation of the evidence remains subjective. For example, in the case of an ultra-edge detection for a faint edge off the bat, there might be debates about whether the evidence is conclusive enough to overturn the on-field decision. This subjectivity can still lead to controversies.</li><li><strong>Technical Limitations:</strong> Like any technology, computer vision is not infallible. Weather conditions, the quality of cameras, and calibration issues can sometimes affect the accuracy of the system. While advancements continue to be made, it’s essential to recognize that technology is not immune to occasional glitches.</li></ol><p><strong>Future of Computer Vision in Cricket</strong></p><p>Despite the challenges, computer vision is expected to play an even more significant role in cricket’s future. Here are some developments to look forward to:</p><ol><li><strong>Real-Time Decision-Making:</strong> As technology continues to advance, we can expect more real-time decision-making aids for on-field umpires. Tools like “Hawk-Eye” already provide real-time ball-tracking, and this could be integrated further into the game to assist umpires.</li><li><strong>Player Performance Analysis:</strong> Beyond just decision-making, computer vision can be used for player performance analysis. It can track a player’s movements, analyze their technique, and provide valuable insights for coaching and team strategy.</li><li><strong>Fan Engagement:</strong> Technology has the potential to enhance the fan experience further. Augmented reality (AR) and virtual reality (VR) could be used to provide immersive experiences, allowing fans to virtually step into the shoes of their favorite players or explore match statistics in new ways.</li><li><strong>Refinement and Calibration:</strong> Ongoing efforts will be made to refine and calibrate the technology to minimize errors and subjectivity. This includes improving camera quality, data analysis algorithms, and the training of personnel responsible for using the technology.</li></ol><h3>The YOLO Workflow</h3><ol><li><strong>Input Image:</strong> YOLO takes an image as its input. This image can be anything from a photograph to a medical scan or a video frame.</li><li><strong>Convolutional Neural Network (CNN):</strong> The input image is processed through a deep Convolutional Neural Network (CNN). This CNN plays a crucial role in feature extraction and understanding the content of the image.</li><li><strong>Grid Division:</strong> YOLO divides the processed image into a grid. Each cell within this grid is responsible for making predictions about objects present within its boundaries. This grid-based approach enables efficient and localized object detection.</li><li><strong>Bounding Boxes:</strong> Within each grid cell, YOLO predicts bounding boxes (rectangles) that tightly enclose objects. These bounding boxes specify the position and size of the detected objects.</li><li><strong>Class Predictions:</strong> Simultaneously, YOLO predicts the class of each object contained within the bounding box. For instance, in medical imaging, it might predict whether the object is a bone fracture, joint, or normal tissue.</li><li><strong>Confidence Scores: </strong>YOLO assigns confidence scores to each predicted bounding box. These scores represent the model’s confidence in the accuracy of the prediction. High confidence scores indicate a high probability of the bounding box containing a valid object</li><li><strong>Non-Maximum Suppression:</strong> To refine the results and eliminate duplicate or overlapping predictions, YOLO employs a technique called non-maximum suppression. This step ensures that only the most accurate and relevant predictions are retained.</li><li><strong>Output:</strong> The final output of YOLO consists of bounding boxes, their associated object classes, and confidence scores. These results provide a comprehensive understanding of the objects detected within the input image.</li></ol><h3>Significance of YOLO</h3><p>The YOLO object detection system has had a profound impact on various fields, including:</p><h3>1. Autonomous Vehicles</h3><p>In the realm of self-driving cars, YOLO plays a crucial role in identifying and tracking objects on the road, such as pedestrians, vehicles, traffic signs, and lane boundaries. This real-time object detection capability is fundamental to the safety and functionality of autonomous vehicles.</p><h3>2. Surveillance and Security</h3><p>Security systems rely on YOLO for detecting intruders, recognizing faces, and monitoring activities in real-time. YOLO’s ability to swiftly and accurately identify objects in surveillance footage enhances security measures.</p><h3>3. Medical Imaging</h3><p>In medical imaging, YOLO assists in the detection of anomalies and diseases by analyzing X-rays, MRIs, and CT scans. It can identify and localize specific regions of interest, enabling faster and more precise diagnoses.</p><h3>4. Industrial Automation</h3><p>Manufacturing and logistics industries use YOLO for quality control, object tracking, and inventory management. It streamlines processes by automating object detection tasks.</p><h3>5. Wildlife Conservation</h3><p>In conservation efforts, YOLO aids in wildlife monitoring and protection. It can identify and track animals in their natural habitats, contributing to research and preservation initiatives.</p><p><strong>Conclusion</strong></p><p>The integration of computer vision into cricket, particularly in the realm of 3rd umpire wicket tracking, has marked a significant step forward in the sport’s evolution. It has not only enhanced the accuracy of decision-making but also added a layer of excitement for fans as they eagerly await the outcome of reviews. While there are challenges to be addressed, it is clear that technology is here to stay in cricket, ensuring a fairer and more transparent game for players and spectators alike. As technology continues to advance, it is likely that cricket will see even more innovations in the future, further cementing its status as a modern and progressive sport.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b4fa13c2acad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Powerful is YoloV8 Human Pose Detection]]></title>
            <link>https://medium.com/@jaykumaran2217/how-powerful-is-yolov8-human-pose-detection-71e8b149559a?source=rss-959983f4da26------2</link>
            <guid isPermaLink="false">https://medium.com/p/71e8b149559a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[yolo]]></category>
            <category><![CDATA[computer-vision]]></category>
            <dc:creator><![CDATA[Jaykumaran R]]></dc:creator>
            <pubDate>Sat, 30 Sep 2023 10:00:15 GMT</pubDate>
            <atom:updated>2023-09-30T10:00:15.950Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fig647G-3nx8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dig647G-3nx8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fig647G-3nx8%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/121658e8047947359086381cde994340/href">https://medium.com/media/121658e8047947359086381cde994340/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fdbt4rIBNx0Q%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Ddbt4rIBNx0Q&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fdbt4rIBNx0Q%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/cd29ac859054fa9492bcb6469a1b5b0d/href">https://medium.com/media/cd29ac859054fa9492bcb6469a1b5b0d/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMdIVkms4vQg%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMdIVkms4vQg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMdIVkms4vQg%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/48c642be0b13b46c934ff260fa9ec20c/href">https://medium.com/media/48c642be0b13b46c934ff260fa9ec20c/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FEQO9s5fcziU%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DEQO9s5fcziU&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FEQO9s5fcziU%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/4eee06c44ee75570d05aab47b06fbb03/href">https://medium.com/media/4eee06c44ee75570d05aab47b06fbb03/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F3CBNYmbLP18%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D3CBNYmbLP18&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F3CBNYmbLP18%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0ab121b8bd12b1e852a73a38c0b6018a/href">https://medium.com/media/0ab121b8bd12b1e852a73a38c0b6018a/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FWwQFJsOIp3U%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DWwQFJsOIp3U&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FWwQFJsOIp3U%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ea0b6eec4a291934b20c639eb6b4604a/href">https://medium.com/media/ea0b6eec4a291934b20c639eb6b4604a/href</a></iframe><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fz4bsFpIlFT0&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dz4bsFpIlFT0&amp;image=http%3A%2F%2Fi.ytimg.com%2Fvi%2Fz4bsFpIlFT0%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0e9913afcdfd24b038efcf54208f8904/href">https://medium.com/media/0e9913afcdfd24b038efcf54208f8904/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71e8b149559a" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>