<?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[ModelDepot - Medium]]></title>
        <description><![CDATA[ModelDepot is a place where you can find and share optimized, pretrained ML models that are perfect for your development needs. - Medium]]></description>
        <link>https://medium.com/modeldepot?source=rss----99b2a8be730d---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>ModelDepot - Medium</title>
            <link>https://medium.com/modeldepot?source=rss----99b2a8be730d---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 18:53:50 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/modeldepot" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Deploy your own Cutting Edge Image Classification onto AWS]]></title>
            <link>https://medium.com/modeldepot/deploy-your-own-cutting-edge-image-classification-onto-aws-e46b92f681b2?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/e46b92f681b2</guid>
            <category><![CDATA[api]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[docker]]></category>
            <dc:creator><![CDATA[Harsh Sikka]]></dc:creator>
            <pubDate>Tue, 12 Jun 2018 06:38:29 GMT</pubDate>
            <atom:updated>2018-06-12T06:38:28.766Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zkA-cLrlU76nW2pGtCufUQ.png" /><figcaption>Percept and AWS? A match made in heaven ;)</figcaption></figure><p>Machine Learning forms the backbone of many interesting and powerful features in an amazing array of technology, and part of our mission over at ModelDepot is to help anybody leverage cutting edge ML with just a few lines of code.</p><p>Here, we’ll be deploying Percept, a state of the art Image Classification model that’s packaged into a handy docker container you can deploy on any infrastructure. The Percept container is located on Docker hub, and has the registry id <em>modeldepot/percept</em>. Hold onto that, because it’s going to come in handy when we deploy our task!</p><h4><strong>Deploying to AWS</strong></h4><p><strong>Creating a Task</strong></p><p>The first step necessary in order to deploy our model is the creation of a task that defines it. A task definition functions almost like a blueprint for our model deployment. In specifying a task definition, we inform ECS as to which Docker image to use for its containers, how many containers are going to be used in the task, and the specific resources that are to be allocated to each of those containers.</p><p>To set up a task, navigate over to your ECS console. On the left side you’ll see a menu that has Clusters, Task Definitions, and Repositories. Go ahead and click Task Definitions, we’ll come back to Clusters later. Now, you’re on the Task Definition Management page! Click on the big blue <em>Create New Task Definition</em> button, and lets start setting up our task.</p><p>First, we have to select our lunch type compatibility. You should see the following screen:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kpklfiXG_-uNYo73" /></figure><p>Go ahead and select EC2, because it allows us to manage our own infrastructure using AWS EC2, and it keeps costs low!</p><p>Now, after clicking next step, we get to the Configuration screen! Give your task a cool name, and keep the Task Role and Network Mode at their default values. Your form should look like this so far:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YINcOaTwoQ6uTT9k" /></figure><p>As you move downward, you can leave the Task Execution and Task Size sections on their defaults. Now, when you reach Container Definitions, click the <em>Add Container </em>button.</p><p>A popup out menu should jump out of the right side. Go ahead and give your container a name, and supply that registry id, <em>modeldepot/percept</em>, to the Image section. Map your port from 8000 to 8000, since that’s how our Percept Docker Container is configured, and set the protocol to tcp. Your screen should look something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gS7-UpzjWhF8pNuu" /></figure><p>Awesome, lets scroll down and set up our HealthChecks and Environment. In the HealthChecks section, add the following command:</p><pre>CMD-SHELL curl -f <a href="http://localhost:8000/">http://localhost:8000/</a> || exit 1</pre><p>Set Interval, Timeout, Start Period, and Retries to 30,30,10, and 5 respectively.</p><p>In Environment, set an environmental variable MD_KEY. MD_KEY should be set to whatever key you were given when you registered for <a href="https://modeldepot.io/percept/manage">Percept</a>.</p><p>Your HealthCheck and Environment screens should look as follows:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*C8NvDtO4s0MKEn3G" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uajPD3tB6lwehAhA" /></figure><p>Great, we’re almost done with our task definition! Lets move over to Storage and Logging, and just check the box next to <em>Auto-configure Cloudwatch Logs</em>. Leave everything else as is, and hit the blue Add button!</p><p>Back on your task definition configuration screen, you should now see the following Container Definition:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jUOdb8cCUjCyuhbq" /></figure><p>Awesome, we’re just about done! Go head and click the <em>Create </em>button. If you see a green box on the next page that says Created Task Definition Successfully, you’re all set! Now lets set up a cluster to run our task on!</p><p><strong>Configuring a Cluster:</strong></p><p>Onto the home stretch! Now, we just have to configure the cluster that our ModelDepot tasks are going to run on. A cluster is really just a set of EC2 instances.</p><p>We’ll have to configure the name of the cluster, as well as the EC2 instance type. Keep the number of instances to 1 as well, so that we launch just a single instance for our cluster to sit on.</p><p>On the left side, click the Clusters tab. On the Clusters page, go ahead and click create Cluster.</p><p>Select EC2 Linux + Networking as our cluster template, and lets get this party started!</p><p>On the following screen, set a cluster name for your new cluster.</p><p>NOTE: If you’re really just playing around with deploying your cluster, you may want to opt for Spot pricing, where you can automatically bid for whatever the spare EC2 compute capacity is.</p><p>When selecting your instance type, go for at least c5.4xlarge, as it gives you more than enough resources to run your Percept container. ML can be resource intensive! The minimum requirements are 1gb of ram and 2 CPUs, if you’d like to look into your possible configuration options. Pick just 1 instance and leave everything else as is. On the bottom, configure your Security Group inbound rules to have a port range of 8000. Create a new instance IAM role if you don’t already have one.</p><p>Your screen should be looking similar to this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*cwYxoW08G5V6OIep" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vIgrRzzylBkIzjRj" /></figure><p>Give your Cluster some time to spin up! After a few minutes, it should have 3/3 complete and 3 green boxes.</p><p><strong>Final Step: Setup ECS Instances</strong></p><p>Onto the last step of setup! We just have to set up some ECS instances for our cluster to run on, and configure scaling. Navigate to the same Clusters page that you went to create a new cluster initially. You should now see your new Cluster! Move into it and click the ECS instances tab. Go ahead and click the <em>Scale ECS Instances </em>button.</p><p>Now, wait for your instance to spin up. You can check its status by navigating to the EC2 console.</p><p><strong>Run your Percept Task</strong></p><p>Once everything is ready, navigate back to your cluster, click on tasks, and hit run new task! You should be presented with a form where you should specify EC2 as the launch type. Also select the task definition and cluster you just set up!</p><p>Your screen should look something like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*B8kCYpqT-Y2fzrBk" /></figure><p><strong>Check out your newly deployed ModelDepot Image Classifier!</strong></p><p>Once running, if you navigate back to the tasks section of your cluster, you should see a new task running. Awesome! Click into it and expand the percept task definition by clicking the little arrow next to it.</p><p>Grab the External Link, and navigate to it using your web browser! Congratulations, you’ve just seen your newly deployed Percept container. What you’re looking at is the deployed documentation and API explorer, which lets you sample and play with the REST api end points that Percept comes with out of the box. If you’re interested in learning how to train your own custom classes on Percept, head over here!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*fM3nc1cUJrKX9Edtn8YisQ.png" /></figure><p>Interested in trying ModelDepot Percept? Get started for free <a href="https://modeldepot.io/percept">here</a>! If you have any questions, don’t hesitate to reach us via the in-site chat on the bottom right or via email at <a href="/service/mailto:hi@modeldepot.io">hi@modeldepot.io</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e46b92f681b2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/deploy-your-own-cutting-edge-image-classification-onto-aws-e46b92f681b2">Deploy your own Cutting Edge Image Classification onto AWS</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Apples & Oranges: A Machine Learning Classifier]]></title>
            <link>https://medium.com/modeldepot/apples-oranges-a-machine-learning-classifier-baf549451502?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/baf549451502</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mike Shi]]></dc:creator>
            <pubDate>Tue, 12 Jun 2018 06:36:52 GMT</pubDate>
            <atom:updated>2018-06-12T06:36:51.497Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MKnfk5ot2nCPFMxupOYfqw.png" /><figcaption>by DavidRockDesign (<a href="https://pixabay.com/en/pattern-fruit-fruits-orange-green-3177667/">https://pixabay.com/en/pattern-fruit-fruits-orange-green-3177667/</a>)</figcaption></figure><p>Ever wanted to use image recognition, but didn’t want to be limited to just the fixed classes that come out of the box with pretrained models or cloud APIs? Today we’ll walk through how to train a ML model to recognize some fruits it’s never seen before with 95% accuracy!</p><h4><strong>A Brief Intro on Image Classification and Retraining</strong></h4><p>The ability for a computer to be able to analyze an image and tell you what’s in it (image classification), whether it’s a banana, apple, or hotdog is one of the most visible achievements in deep learning today.</p><p>Today we’ll be using <a href="https://modeldepot.io/percept">ModelDepot Percept</a>: a Docker image that allows you to quickly predict, retrain and deploy ML state of the art image classification into production for free, with a REST API. You can find the complete code <a href="https://colab.research.google.com/drive/1DF1YlLVIJqRtPe3kujvg2BDhaUSZCqx-">here</a>.</p><h4><strong>Step 1: Acquiring Data</strong></h4><p>In this guide, I’ll be using <a href="https://github.com/Horea94/Fruit-Images-Dataset">Fruits 360</a>, a dataset of 32,000+ images of 65 different types of fruit, though we’ll just be using a small fraction of it to show how accurate we can be with a few examples!</p><p>Once we download the dataset, we’ll see that the Training data is already neatly packaged into a folder per class. We’ll simply use a short snippet of Python to organize our training data per class, and convert the data to Base64 to send it to our Percept instance.</p><h4><strong>Step 2: Exploring the /train and /predict Endpoint</strong></h4><p>Now that we have our data all set up, let’s take a look at the API docs around training and predicting. We’ll first want to queue up the data for training, using the /batch_queue endpoint. After that we can train the model by calling /train and then start making predictions with /predict!</p><p><strong>/v1/batch_queue/{label_name}</strong></p><p>The batch_train endpoint is fairly straightforward, it takes in the desired label name (ex. ‘banana’) in the URL and it takes in an array of images in the body of the post request.</p><p>The value defined in the image property of each object can be either a image URL accessible to the Docker container (gif, jpg or png), or a Base64 encoded string.</p><p>You can use the API explorer to try queueing a sample as follows:</p><p>However, we’ll be using a Python script to programmatically train the model for us, instead of through the API explorer.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1017/1*1YrsNRXi2uaHgkcZn9rr4Q.png" /></figure><p><strong>/v1/train</strong></p><p>Once we’ve queued up all of our data using batch_queue, we can tell the classifier to start training! Hitting this endpoint will kick off a training session and it’ll return with training information after it’s finished.</p><p><strong>/v1/predict</strong></p><p>The predict method only takes in an object with the property ‘image’. It’s exactly the same as a single image object being passed into /v1/batch_queue so we won’t repeat that here. We’ll see this endpoint in action as well in the following code section.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*P0vdFV0sAIv8QPH43ZaiiA.png" /></figure><h4><strong>Step 3: Writing The Code</strong></h4><p>Now that we have the data and an understanding of the API, we’ll put together the code we’ll need to train and test our custom fruits classifier. In this example, we’ll be using Python, but the basic idea still holds across any language. See the <a href="https://colab.research.google.com/drive/1DF1YlLVIJqRtPe3kujvg2BDhaUSZCqx-">final code here</a>. We’ll walk through the major parts of the code here.</p><p><strong>Loading Train/Validation Data</strong></p><p>With the data loader helper function already defined from the first part of this guide, we’ll want to load training and validation data. The validation data is held out data not used to train the model so we can accurately assess its performance.</p><pre>TRAIN_DIR = &#39;Fruit-Images-Dataset-master/Training&#39;<br>VALIDATION_DIR = &#39;Fruit-Images-Dataset-master/Validation&#39;</pre><pre>NUM_CLASSES = 15<br>TRAIN_SAMPLES = 20<br>VAL_SAMPLES = 20</pre><pre>_, all_label_names, _ = next(os.walk(TRAIN_DIR))</pre><pre>label_names = all_label_names[:NUM_CLASSES]</pre><pre>train = get_data(TRAIN_DIR, label_names, TRAIN_SAMPLES)<br>val = get_data(VALIDATION_DIR, label_names, VAL_SAMPLES)</pre><p>Here we’ll be loading 20 training samples for each fruit, and 20 validation samples for each fruit, for 15 different types of fruit. Normally the train/val split would be closer to 80%/20% of the data, but since we’re demoing how few samples you’d need to train a model, we’re just doing a 50/50 split.</p><p><strong>Training Our Model</strong></p><p>Now with our training data split up, we can go ahead and train our model. We’ll write a helper method here to send the request.</p><pre>def queue(label, samples):<br>  return requests.post(<br>    &#39;http://localhost:8000/v1/batch_queue/%s&#39; % label, <br>    json={<br>      &#39;batch_images&#39;: [{&#39;image&#39;: image} for image in samples]<br>    })</pre><p>This method simply makes a post request to the batch queue endpoint with a given label name, and then passes in the JSON encoded parameters for the image URLs.</p><p>With the helper method done, we just need 3 more lines of code to actually run the training:</p><pre>for label in train:<br>  print(queue(label, train[label]).json())</pre><pre>print(requests.post(‘http://localhost:8000/v1/train’).json())</pre><p>The first part queues up all of our data samples, and the second part kicks off the training on the server. This can take a while as it learns and optimizes the model for 300 samples.</p><p>We can further just double check that our model has everything learned:</p><pre>requests.get(‘http://localhost:8000/v1/train/status&#39;).json()</pre><p>This should tell us we have 15 classes with trained on 300 images.</p><p><strong>Testing the Model</strong></p><p>Now that we’ve trained a model, we want to understand how well the model will do on totally new data that it’s never seen before. To do this we’re going to ask it to predict the images in our held out validation set and see what it’ll tell us.</p><p>Using scikit-learn’s accuracy metric, we can see our classifier does really well!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*g6T21LjYbOUQw383R0nJAA.png" /><figcaption>95% Validation Accuracy!</figcaption></figure><p>We can even examine deeper and look at the confusion matrix to see how often it misclassified certain types of fruits for other fruits. The strong diagonal line and sparse output elsewhere shows how well our classifier does.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/686/1*sCJuzm2MW1VYVnEYoPeUsg.png" /></figure><h4>Done!</h4><p>We’ve successfully gone from nothing, to having our very own (and quite accurate) fruit classifier! If you’re interested in how all these internals work, we’ve talked about the underlying ML technology and some of our roadmap in this article here. We hope you’ve seen how easy it can be to leverage machine learning for a custom class, and hope that <a href="https://modeldepot.io/">ModelDepot</a> can help you use ML in your product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VCG-6XSe_WnsngY-BWmGYg.png" /></figure><p>Interested in trying ModelDepot Percept? Get started for free <a href="https://modeldepot.io/percept">here</a>! If you have any questions, don’t hesitate to reach us via the in-site chat on the bottom right or via email at <a href="/service/mailto:hi@modeldepot.io">hi@modeldepot.io</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=baf549451502" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/apples-oranges-a-machine-learning-classifier-baf549451502">Apples &amp; Oranges: A Machine Learning Classifier</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Percept — Hello World]]></title>
            <link>https://medium.com/modeldepot/percept-hello-world-adf984c44ce7?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/adf984c44ce7</guid>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Mike Shi]]></dc:creator>
            <pubDate>Tue, 12 Jun 2018 06:35:13 GMT</pubDate>
            <atom:updated>2018-06-12T06:35:12.662Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2FB_3rwSZQRwZ47oVkGWIg.png" /></figure><p>ModelDepot Percept is an easy to use REST API for a pre-trained image classification ML model deployable on your own infrastructure via Docker.</p><p>Percept exists to help developers that might have never considered leveraging ML, to quickly get started and go to production on their own terms in ways that might be more familiar than starting up a Jupyter Notebook. Our user interviews showed that REST APIs are infinitely more approachable than Jupyter Notebooks, and a lot easier to deploy into production as well!</p><p><strong>Our Value Proposition</strong></p><p>Our product should help you get from 0 to production ready ML without you ever having to dive deep into a machine learning course. However, we still want to provide knobs and controls that allow you to continue to fine-tuning your performance. You can deploy this product on your own infrastructure, cutting out the large premiums that providers like Amazon or Google might be charging you on infrastructure, while also allowing you to quickly train on your own datasets (something you can’t do on their platforms).</p><p>We’ve bundled together a robust set of technologies with rigorous user testing to ensure that you have the best experience while using our product (and let us know if you aren’t!). We’re confident that you’ll have a good experience, which is why we’re confident in exposing the inner workings of our product.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VCG-6XSe_WnsngY-BWmGYg.png" /></figure><p>Interested in trying ModelDepot Percept? Get started for free <a href="https://modeldepot.io/percept">here</a>! If you have any questions, don’t hesitate to reach us via the in-site chat on the bottom right or via email at <a href="/service/mailto:hi@modeldepot.io">hi@modeldepot.io</a>.</p><p>We hope that if you find it useful, you will help support our efforts of development. In addition, you can find more easy to use, open source, state-of-the-art ML models to deploy or integrate into your product on the <a href="https://modeldepot.io/browse">ModelDepot platform</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=adf984c44ce7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/percept-hello-world-adf984c44ce7">Percept — Hello World</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Percept — What’s Inside the ML Container]]></title>
            <link>https://medium.com/modeldepot/percept-whats-inside-the-ml-container-1e71f5ee2747?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/1e71f5ee2747</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[developer-tools]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[api]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mike Shi]]></dc:creator>
            <pubDate>Tue, 12 Jun 2018 06:33:39 GMT</pubDate>
            <atom:updated>2018-06-12T06:33:39.349Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mzpViZDTqJmqSwEEr8ikVg.png" /><figcaption>Ship Icon made by <a href="http://www.freepik.com/">Freepik</a> from <a href="http://www.flaticon.com/">www.flaticon.com</a></figcaption></figure><p>As part of delivering on transparent machine learning, we wanted to walk through the technology that underlies ModelDepot Percept, so that you can understand the expected behavior, performance, advantages, and potential pitfalls of our system. We’re not a fan of black boxes, and this product is no different.</p><p>We’ve bundled together a robust set of technologies together with rigorous user testing to ensure that you have the best experience using our product (and let us know if you aren’t!). We’re confident that you’ll have a good experience, which is why we’re confident in laying out the details in the open for you to see.</p><p>With that, let’s dive into the technicals.</p><p><strong>Deep Learning Backbone</strong></p><p>Today, the ML workhorse being packaged in every Docker image is a ResNet v1 101 architecture trained on Google’s <a href="https://storage.googleapis.com/openimages/web/index.html">Open Images</a> v2 Dataset. The model comes out of the box with 5,000 classes, which is more diverse than <a href="http://www.image-net.org/">ImageNet</a>. The pre-trained nature of this model allows you to quickly get started on classifying images with no training data on your side. The diversity of the training data also allows you to quickly extend the model to transfer-learn on a breadth of new classes it does not know. We’ll cover the technique being used in the next section.</p><p>In the future, we hope to provide a variety of model architectures trained on both Open Images as well as ImageNet to allow you to choose between speed, and classes that are more applicable to you out of the box. Let us know if you have any preferences, and we’ll try to get it done.</p><p><strong>Few Shot Learning Technique</strong></p><p>Of course one of the main advantages of using ModelDepot Image Classification over a typical cloud provider is the ability to train on novel image classes with your own datasets. We have a tutorial over here showing you how to do it if you haven’t explored that feature yet.</p><p>We currently use features extracted out of the backbone model and use a Support Vector Machine to find the confidence values of new classes that you’ve trained. While this is a simple method, it serves as a powerful baseline and performs favorably to state of the art methods.</p><p>Over time, we hope to include more techniques that provide even more accurate few-shot learning results. We hope to also make it easy to fine-tune the DL model itself, and potentially re-train the model from scratch if you own a big enough dataset.</p><p>With that, we hope we’ve shed some light on some of the magic, that isn’t so magic, behind our ML containers. Armed with this knowledge, you’ll hopefully be able to do further research on how good or bad our product might be for your use case, and let us know of any feedback on how we can do better.</p><p>I hope to see you explore how ML can improve your product or user’s experience, and even better if ModelDepot is able to help you along the way.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VCG-6XSe_WnsngY-BWmGYg.png" /></figure><p>Interested in trying ModelDepot Percept? Get started for free <a href="https://modeldepot.io/percept">here</a>! If you have any questions, don’t hesitate to reach us via the in-site chat on the bottom right or via email at <a href="/service/mailto:hi@modeldepot.io">hi@modeldepot.io</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1e71f5ee2747" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/percept-whats-inside-the-ml-container-1e71f5ee2747">Percept — What’s Inside the ML Container</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Need ML to solve your business problems?]]></title>
            <link>https://medium.com/modeldepot/need-ml-to-solve-your-business-problems-c8b942b5d86b?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/c8b942b5d86b</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[business]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Harsh Sikka]]></dc:creator>
            <pubDate>Thu, 19 Apr 2018 12:01:01 GMT</pubDate>
            <atom:updated>2018-04-19T12:01:00.992Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IhSmzjv5tyKSlHvSFKze0Q.jpeg" /></figure><p>At <a href="https://modeldepot.io/">ModelDepot</a>, our mission is to democratize AI by empowering every developer to leverage cutting edge pre-trained models in hours, not months.</p><p>We’re offering <a href="https://modeldepot.io/consulting">free 30 minute sessions</a> to help you identify, understand, and apply the right ML techniques and tools to successfully solve your business problems. There’s no upselling or catch. Helping teams use ML is critical to our mission, and we’re keen to learn from their experiences along the way.</p><p>ModelDepot has helped thousands of developers find the right ML model, and we’re growing faster than ever. We have more than 5 years of experience in AI and applied ML, and have worked as data scientists, software engineers, and Product Managers at some incredible places like Microsoft and Jawbone.</p><p><a href="https://modeldepot.io/consulting"><strong><em>Sign up here</em></strong></a><strong><em> if you’re interested! We’ll get back to as many requests as possible. Questions? Shoot me an </em></strong><a href="/service/mailto:harsh@modeldepot.io"><strong><em>email</em></strong></a><strong><em>!</em></strong></p><figure><a href="https://modeldepot.io/consulting"><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WrXA_rxCpOlJVhl4u7l2ww.png" /></a></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c8b942b5d86b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/need-ml-to-solve-your-business-problems-c8b942b5d86b">Need ML to solve your business problems?</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Democratizing Machine Learning for Software 2.0]]></title>
            <link>https://medium.com/modeldepot/we-previously-talked-about-how-transfer-learning-will-radically-change-ml-c4fed21a65a2?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/c4fed21a65a2</guid>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Harsh Sikka]]></dc:creator>
            <pubDate>Mon, 12 Feb 2018 08:32:55 GMT</pubDate>
            <atom:updated>2018-05-12T05:43:51.773Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*quEy6e3xCnX-ak4p0ztSnQ.png" /><figcaption>Turns out matrices may be more relevant to our future than we thought</figcaption></figure><p><em>We previously talked about how </em><a href="https://medium.com/modeldepot/transfer-learning-will-radically-change-machine-learning-for-engineers-78732b2bb415"><em>Transfer Learning will radically change ML. </em></a><em>People seemed to love it, so we thought we’d follow up with some thoughts on democratizing AI and Software 2.0.</em></p><p>Machine Learning has had significant impact on the technology industry in the past few years, but many engineers still view it as little more than a highly specialized tool with a complex, esoteric setup. This is for good reason, since ML hasn’t reached the plug and play accessibility of most modern engineering paradigms.</p><p>As it turns out, when solving problems in the real world it is significantly easier to collect data than to write explicit programs. This contrast is the driving force behind what <a href="https://medium.com/@karpathy/software-2-0-a64152b37c35">some brilliant folks</a> refer to as Software 2.0</p><h4><strong>Machine Learning as the Basis of Software 2.0</strong></h4><p>Software 2.0 differs from the current standard in one important way: it will center on neural network weights, not explicit algorithms. The engineers of the future will be significantly more productive because of this, as their role will revolve around finding the constraints of the problem they need to solve and narrowing down on the optimal ML program to satisfy those constraints. These programmers won’t be occupied with writing intricate, explicit programs and maintaining them, but instead collecting data and picking the right ML paradigm to match.</p><p>The transition from explicit programming to Software 2.0 is already underway, and can be observed in fields like Robotics, Image Classification, Translation, and Gaming. For example, Speech Recognition used to be extremely time intensive to build, involving Hidden Markov Models and preprocessing but today is primarily supported by neural networks.</p><p>Of course, this isn’t to say that that Software 1.0 will disappear, but that it will move into the role of infrastructure, supporting the paradigms proposed above.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TK3nD3QUueAo7VwuMBRAIQ.png" /></figure><h4><strong>The First Step: Accessible Machine Learning</strong></h4><p>The first step to this incredibly productive future is to make Machine Learning more accessible to the everyday engineer. Many programmers still view neural networks and other models as little more than specific tools built through a complex, arduous process.</p><p>But paradigms like Transfer Learning are changing this, and ML is more applicable to real world problems than ever before. Pretrained models are becoming increasingly prevalent and making use of them promises to supercharge most general development needs. In tapping into a trained model for a related purpose to its original design, teams can leapfrog the setup and training required to bring a model up to par for the task.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FNQK4ZY_gwKI&amp;src_secure=1&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DNQK4ZY_gwKI&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FNQK4ZY_gwKI%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="1280" height="720" frameborder="0" scrolling="no"><a href="https://medium.com/media/b1d75f6a971d6c9631e318d8fb0736a1/href">https://medium.com/media/b1d75f6a971d6c9631e318d8fb0736a1/href</a></iframe><p><em>The shift to Software 2.0 is already underway, but we still need tools to facilitate the sharing and discovery of these models to benefit the everyday engineer and supercharge their productivity.</em></p><p><em>That’s why we’re building </em><a href="http://modeldepot.io/"><em>ModelDepot</em></a><em>, to decrease the friction associated with model access and contribute to the democratization of AI in the 21st century.</em></p><p><em>Join the conversation on </em><a href="https://gitter.im/ModelDepot/Lobby"><em>Gitter</em></a><em>!</em>👋</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c4fed21a65a2" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/we-previously-talked-about-how-transfer-learning-will-radically-change-ml-c4fed21a65a2">Democratizing Machine Learning for Software 2.0</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Transfer Learning Will Radically Change Machine Learning for Engineers]]></title>
            <link>https://medium.com/modeldepot/transfer-learning-will-radically-change-machine-learning-for-engineers-78732b2bb415?source=rss----99b2a8be730d---4</link>
            <guid isPermaLink="false">https://medium.com/p/78732b2bb415</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Harsh Sikka]]></dc:creator>
            <pubDate>Thu, 01 Feb 2018 19:03:03 GMT</pubDate>
            <atom:updated>2018-05-12T05:45:21.673Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/888/1*mmnxXocQ36q3mqYgyxKz7g.jpeg" /></figure><h4>The Problem</h4><p>In traditional supervised machine learning, we teach a model to become more successful and efficient at a task, by providing it with example data. Generally, once the model begins to perform well on the training data for the domain or problem it is tasked with, we expect a reasonable performance for new data. But, if you think about it, there are a few issues with this traditional supervised learning process</p><p>As engineers, we’re forced to construct specific models that only excel at a specific problem. This costs us valuable engineering time to create, train and tune models from scratch for every new problem we want to tackle, even if it’s a problem that has been solved in industry.</p><p>From a product perspective, this is incredibly detrimental to progress, and could really hamper feature releases and engineering productivity. Transfer Learning offers an interesting solution to this problem.</p><p>Transfer learning as a paradigm can solve this problem, by allowing us to leverage existing knowledge and data from a certain related domain to the new one we’re trying to train for. In 2016, Andrew Ng posited that Transfer Learning would be essential to commercial and industry success.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FF1ka6a13S9I%3Fstart%3D12%26feature%3Doembed%26start%3D12&amp;url=http%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DF1ka6a13S9I&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FF1ka6a13S9I%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/9b29e20ff4c1bb3a141e54cc06c89162/href">https://medium.com/media/9b29e20ff4c1bb3a141e54cc06c89162/href</a></iframe><h4><em>Applications of Transfer Learning</em></h4><p>Making use of pre-trained models and related domain data promises to supercharge most general development for machine learning. By tapping into a pre-trained model for a related purpose to its original design, your team can leapfrog the data cleaning, setup and training required to bring a model up to par for the task.</p><p>Two common areas where Transfer Learning has already been demonstrated to great success are Images and Text .</p><p>Transfer Learning has been particularly effective with Image Data, and it is common to leverage a deep learning model trained on some large image data set, like ImageNet. These pretrained models can be directly included in other new models that expect some form of Image Input.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WyTv5S6b8fmOG6aBrZyiHw.jpeg" /><figcaption>Mask R-CNN: a ML model from Facebook built on top of a pre-trained ResNet, another image detection model.</figcaption></figure><p>With Textual data, words are mapped to vectors where different words with a similar meaning have a similar vector representation. Pretrained models exist to learn these representations, and are widely available. These can then be incorporated into deep learning language models, both at the input or output stage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YjO-agjV5Q5nv7BaAGzBsQ.png" /></figure><p><strong>Transfer Learning and pretrained models are the future of machine learning applications in general development, and as such need to be made more accessible and discoverable for everyone.</strong></p><p><strong><em>That’s why we’re building </em></strong><a href="https://modeldepot.io/"><strong><em>ModelDepot</em></strong></a><strong><em>, to decrease the friction associated with model access and contribute to the democratization of AI in the 21st century.</em></strong></p><p><strong><em>Join the conversation on </em></strong><a href="https://gitter.im/ModelDepot/Lobby"><strong><em>Gitter</em></strong></a><strong><em>! </em></strong>👋</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78732b2bb415" width="1" height="1" alt=""><hr><p><a href="https://medium.com/modeldepot/transfer-learning-will-radically-change-machine-learning-for-engineers-78732b2bb415">Transfer Learning Will Radically Change Machine Learning for Engineers</a> was originally published in <a href="https://medium.com/modeldepot">ModelDepot</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>