<?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 Harald Carlens on Medium]]></title>
        <description><![CDATA[Stories by Harald Carlens on Medium]]></description>
        <link>https://medium.com/@hcarlens?source=rss-711d325d27b4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*0Ac1pQUU5YyL3Sdf2udVPg.jpeg</url>
            <title>Stories by Harald Carlens on Medium</title>
            <link>https://medium.com/@hcarlens?source=rss-711d325d27b4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 11:46:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@hcarlens/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[Fixing Jupyter: ERR_ACCESS_DENIED error using Chromium on Ubuntu]]></title>
            <link>https://hcarlens.medium.com/fixing-jupyter-err-access-denied-error-using-chromium-on-ubuntu-d4268c76fc1a?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/d4268c76fc1a</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[jupyter-notebook]]></category>
            <category><![CDATA[error]]></category>
            <category><![CDATA[ubuntu]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Mon, 19 Oct 2020 09:55:40 GMT</pubDate>
            <atom:updated>2020-10-29T16:32:47.625Z</atom:updated>
            <content:encoded><![CDATA[<h4>A simple two-step fix to a problem that was driving me crazy</h4><p><em>This story was originally published at </em><a href="https://harald.co/2020/10/19/fixing-jupyter-err_access_denied-error-using-chromium-on-ubuntu/"><em>http://harald.co</em></a><em> on October 19, 2020.</em></p><p>I just discovered how to fix a really annoying issue with Jupyter on Ubuntu I’ve had for a while. Whenever I run jupyter notebook, instead of my browser opening right to the notebook page, I would get this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/885/0*oCb6nWXAoHAW9D2h" /></figure><p>Then I’d have to go back into the terminal, and click the link there.</p><p>Here’s an easy fix*:</p><ol><li>Generate a jupyter config file: jupyter notebook --generate-config</li><li>Set c.NotebookApp.use_redirect_file = False</li></ol><p>And you’re done! Next time you run jupyter notebook, it should open in your browser without any issues:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/998/0*XBk9kcetmwcQL5UL" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/462/0*AYDe7Qps-5tqbj7e.jpeg" /></figure><p>*credit for this goes to <a href="https://stackoverflow.com/questions/59096651/change-startup-url-in-jupyter-notebook/59578946#59578946">kosii on StackOverflow</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d4268c76fc1a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Five Tools to Supercharge Your Cloud Deep Learning Workflow]]></title>
            <link>https://medium.com/genesis-cloud-blog/five-tools-to-supercharge-your-cloud-deep-learning-workflow-3afb2e25be9f?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3afb2e25be9f</guid>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[machine-learning-tools]]></category>
            <category><![CDATA[cloud]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Wed, 12 Aug 2020 05:00:00 GMT</pubDate>
            <atom:updated>2020-09-03T08:51:04.772Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YAaNNaYDrSri1PKA.jpg" /></figure><p>Being a productive deep learning researcher takes more than just an understanding of theory. Spending a little time setting up the right tools can pay off in just a few hours, and save you an enormous amount of time!</p><p>If you feel like using cloud GPUs adds overhead to your training or development workflow, these tools are for you. Once you get used to them, developing and training models in the cloud will feel just like doing it on your local computer — except with unlimited resources!</p><p>We’ll talk about five key tools here: remote debugging, agent forwarding, port forwarding, WSL, and Cmder.</p><h3>1: Remote Debugging</h3><p>This might sound familiar to you: I used to prototype and debug new models on my local pc, then push my code to a remote VM and start a training run there. If I encountered an issue, I’d have to try to replicate it locally if I wanted to step through it in the debugger — but this wasn’t always an option, since some issues were specific to certain types of hardware. So I ended up having to add lots of extra logging to diagnose issues that occurred on remote VMs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gU8-WYPz5nwOr5zu.png" /></figure><p>This is where remote debugging saves the day. <a href="https://code.visualstudio.com/blogs/2019/07/25/remote-ssh">Visual Studio Code</a> added this last summer, and the <a href="https://www.jetbrains.com/help/pycharm/remote-debugging-with-product.html">paid version of PyCharm</a> has had it for a while. We’ll talk about the VS Code version here since it’s free. This installs a version of VS Code on the cloud VM and connects to that, so that you can edit remote code in VS Code as if it was local code, with all your extensions and everything else. You can even run code on your remote VM, and step through it in the debugger, all from a local VS Code window!</p><h3>2: SSH Keys and Agent Forwarding</h3><p>Finding out about agent forwarding was a huge game changer for me. I used to dread setting up new cloud VMs — I’d have to clone my repository, and type in my GitHub password every time I wanted to pull or push code.</p><p>Turns out I was doing it all wrong! If you <a href="https://help.github.com/en/github/authenticating-to-github/connecting-to-github-with-ssh">set up an SSH key</a> on your GitHub account, you can pull or push code from a computer with the SSH key on it.</p><p>And you can go one better than that — rather than setting up a new SSH key on every VM you set up and adding that to your GitHub account, you can make your local SSH keys available to any VM you SSH into using something called <a href="https://docs.github.com/en/developers/overview/using-ssh-agent-forwarding">SSH Agent Forwarding</a>.</p><p>It’s as simple as replacing ssh user@hostname with ssh -A user@hostname!</p><p>No more typing passwords or copying SSH keys…</p><h3>3: Using Cloud Jupyter Notebooks: SSH Port Forwarding</h3><p><a href="https://jupyter.org/">Jupyter</a> is unmatched for data exploration, and with a little extra setup a cloud Jupyter server can work just like a Jupyter server running locally.</p><p>When you run a Jupyter server locally, it will give you a link to open in your browser. When you do the same on a remote VM, the link doesn’t automatically work since it refers to the remote VM as ‘localhost’. The trick here is to use something called SSH Port Forwarding, which can forward traffic for a specific port on your local computer to the remote VM — in this case, the traffic to the notebook server.</p><p><em>Note: Genesis cloud Preconfigured ML instances come with Jupyter Lab pre-installed. If you’re using another provider and haven’t already installed Jupyter on your VM, see </em><a href="https://jupyter.readthedocs.io/en/latest/install/notebook-classic.html"><em>this link</em></a><em> for installation instructions</em>.</p><ol><li>SSH into your remote VM, forwarding port 8888 (the default port used by Jupyter) ssh user@host -L 8888:localhost:8888</li><li>Start a Jupyter notebook server on the remote VM: jupyter notebook --no-browser</li><li>Click on the notebook link to open the remote notebook in your browser. You’re now editing your cloud notebooks from your local computer!</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*LK4AlbYh-5Cipf0A.png" /><figcaption>The command line on the remote VM</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/918/0*R_RgXKIo5O3lEKgI.png" /><figcaption>A notebook session in your local browser</figcaption></figure><h3>4 &amp; 5: Windows Subsystem for Linux (WSL) and Cmder</h3><p>If you use Linux or macOS on your desktop or laptop, opening a console on a cloud Linux VM is super easy. But if you’re on Windows… it turns out it’s now pretty easy too! You’ll just need to run <em>WSL</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*uM7D264sfHAvLxMm.png" /></figure><p>WSL. Image source: <a href="https://docs.microsoft.com/en-us/windows/wsl/install-win10">https://docs.microsoft.com/en-us/windows/wsl/install-win10</a></p><p>WSL is the <a href="https://docs.microsoft.com/en-us/windows/wsl/about">Windows Subsystem for Linux</a> — think of it as running a mini Linux system alongside Windows — and it integrates really nicely into Windows 10, allow you to open Linux terminals just like you would open a command prompt! Installing the base WSL takes <a href="https://docs.microsoft.com/en-us/windows/wsl/install-win10">just a few steps</a>, and then you can go ahead and install your favourite distro on top of it (if you’re not sure which to go for, Ubuntu 18.04 LTS is a good choice!).</p><p>Alongside WSL, you might like to use <a href="https://cmder.net/">Cmder</a> to manage your terminal windows. It’s highly configurable, so you can set it up for example to open one windows command prompt and one Linux shell when it starts up, or several different shells with different virtual environments initialised. It makes it easier to manage multiple connections to remote VMs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YYScKXb8EeEWEAap.png" /><figcaption>cmder. Source image: <a href="https://cmder.net/">https://cmder.net/</a></figcaption></figure><h3>Summary</h3><p>Go forth and be productive!</p><p>Here’s a quick summary of the tools and how to use them:</p><ol><li>Use an IDE which has <a href="https://code.visualstudio.com/blogs/2019/07/25/remote-ssh">Remote Debugging</a>. More detailed instructions for how to set this up on Genesis Cloud can be found <a href="https://support.genesiscloud.com/support/solutions/articles/47001117271-develop-with-vs-code-and-ssh-directly-on-your-genesis-cloud-instance">here.</a></li><li>Specify the -A flag when using SSH to avoid having to type passwords or copy SSH keys, e.g. ssh -A user@hostname</li><li>Use port forwarding to access a remote Jupyter server from your local browser — e.g. ssh user@host -L 8888:localhost:8888</li><li>For Windows users: install <a href="https://docs.microsoft.com/en-us/windows/wsl/install-win10">Windows Subsystem for Linux</a> (WSL)</li><li>For Windows users: install <a href="https://cmder.net/">Cmder</a></li></ol><p><em>Originally published at </em><a href="https://blog.genesiscloud.com/2020/Five-Tools-to-Supercharge-Your-Cloud-Deep-Learning-Workflow"><em>https://blog.genesiscloud.com</em></a><em> on August 12, 2020.</em></p><p>Related:</p><ul><li><a href="https://towardsdatascience.com/better-data-loading-20x-pytorch-speed-up-for-tabular-data-e264b9e34352">Better Data Loading: 20x PyTorch Speed-Up for Tabular Data</a></li><li><a href="https://medium.com/@hcarlens/reproducibility-issues-using-openai-gym-8ca605071d6a">How to Properly Set Random Seeds with OpenAI Gym</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3afb2e25be9f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/genesis-cloud-blog/five-tools-to-supercharge-your-cloud-deep-learning-workflow-3afb2e25be9f">Five Tools to Supercharge Your Cloud Deep Learning Workflow</a> was originally published in <a href="https://medium.com/genesis-cloud-blog">Genesis Cloud Blog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Better Data Loading: 20x PyTorch Speed-Up for Tabular Data]]></title>
            <link>https://medium.com/data-science/better-data-loading-20x-pytorch-speed-up-for-tabular-data-e264b9e34352?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/e264b9e34352</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[deep-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Tue, 28 Apr 2020 18:52:48 GMT</pubDate>
            <atom:updated>2026-03-30T15:35:57.718Z</atom:updated>
            <content:encoded><![CDATA[<h4>A simple change to speed up your deep learning training massively</h4><h3>Deep Learning: Need For Speed</h3><p>When training deep learning models, performance is crucial. Datasets can be huge, and inefficient training means slower research iterations, less time for hyperparameter optimisation, longer deployment cycles, and higher compute cost.</p><p>Despite all that, it can be hard to justify investing too much time speeding things up, as there are many potential dead ends to explore. But fortunately, there are some quick wins available!</p><p>I’m going to show you how a simple change I made to my dataloaders in PyTorch for tabular data sped up training by <strong>over 20x </strong>— without any change to the training loop! Just a simple drop-in replacement for PyTorch’s standard dataloader. For the model I was looking at, that’s a sixteen minute iteration time reduced to forty seconds!</p><p>And all without installing any new packages, making any low-level code changes, or changing any hyperparameters.</p><h3>Research/Industry Disconnect</h3><p>In supervised learning, a quick look at <a href="https://www.arxiv-sanity.com/top">Arxiv-Sanity</a> tells us that the top research papers at the moment are all either about images (whether classification or GANs for generation), or text (mostly variations on BERT). These are great in areas where traditional machine learning just doesn’t stand a chance — but require expertise and a significant research budget to execute well.</p><p>On the other hand, much of the data that many companies hold already resides in databases, in a nice tabular format. Some examples include customer details for lifetime value estimation, click-through optimisation, and financial time-series data.</p><h3>What’s Special About Tabular Data?</h3><p>So why is this rift between research and industry a problem for us? Well, the needs of state-of-the-art text/vision researchers are very different of those doing supervised learning on tabular data sets.</p><p>Having data in tabular form (i.e. a database table, Pandas DataFrame, NumPy Array, or PyTorch Tensor) makes things easier in several ways:</p><ol><li>Training batches can be taken from contiguous chunks of memory by <em>slicing</em>.</li><li>No per-sample preprocessing cost, allowing us to make full use of <a href="https://arxiv.org/pdf/1803.09820.pdf">large-batch training</a> for additional speed (remembering to increase the learning rate so we don’t overfit!).</li><li>If your dataset is small enough, it can be loaded on to the GPU all in one go. (while this is technically also possible with text/vision data, datasets there tend to be larger and some preprocessing steps are more easily done on CPU).</li></ol><p>These optimisations are possible for tabular data and not for text/vision data because of two main areas of difference: models and data.</p><p><strong>Models:</strong> vision research tends to use large deep convolutional neural nets (CNNs); text tends to use large recurrent neural nets (RNNs) or Transformers; but on tabular data plain fully connected deep neural nets (FCDNN) can do fine. While not always the case, in general vision and text models require more parameters to learn more nuanced representations than interactions between variables in tabular data, and so forward and backward passes can take longer.</p><p><strong>Data</strong>: vision data tends to be saved as nested folders full of images, which can require significant pre-processing (cropping, scaling, rotating, etc). Text data can be large files or other text streams. Both of these will generally be saved on disk, and loaded from disk in batches. This isn’t an issue because the disk read/write speeds aren’t the bottleneck here — the preprocessing or backward passes are. Tabular data, on the other hand, has the nice property of being easily loaded into contiguous chunks of memory in the form of an array or tensor. Preprocessing on tabular data tends to be done separately in advance, either in a database, or as a vectorised operation on a dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*J414ZbNFsog5EwP1NCuXFw.png" /><figcaption>Comparison of different types of Supervised Learning research</figcaption></figure><h3>PyTorch &amp; DataLoaders</h3><p>As we’ve seen, loading tabular data can be really easy and fast! So of course PyTorch works great by default for tabular data… <em>right?</em></p><p>It turns out it doesn’t! 😩</p><p>Just last week I was training a PyTorch model on some tabular data, and wondering it was taking so long to train. I couldn’t see any obvious bottlenecks, but for some reason, the GPU usage was much lower than expected. When I dug into it with some profiling I found the culprit… <em>the DataLoader</em>.</p><p><strong>What is a DataLoader? </strong>DataLoaders do exactly what you might think they do: they load your data from wherever it is (on disk, in the cloud, in memory) to wherever it needs to be for your model to use it (in RAM or GPU memory). In addition to this, they take care of splitting your data into batches, shuffling it, and pre-processing individual samples if necessary. Wrapping this code in a DataLoader is nicer than having it scattered throughout, as it allows you to keep your main training code clean. <a href="https://pytorch.org/tutorials/beginner/data_loading_tutorial.html">The official PyTorch tutorial also recommends using DataLoaders</a>.</p><p><strong>How do you use them? </strong>It depends on the type of data you have. For tabular data, PyTorch’s default DataLoader can take a TensorDataset. This is a lightweight wrapper around the tensors required for training — usually an X (or features) and Y (or labels) tensor.</p><pre>data_set = TensorDataset(train_x, train_y)<br>train_batches = DataLoader(data_set, batch_size=1024, shuffle=False)</pre><p>You can then use this in your training loop:</p><pre>for x_batch, y_batch in train_batches:</pre><pre>    optimizer.zero_grad()</pre><pre>    loss = loss_fn(model(x_batch), y_batch)</pre><pre>    loss.backward()</pre><pre>    optimizer.step()</pre><pre>    ...</pre><p><strong>Why is this bad? </strong>This looks pretty good, and certainly very clean! The problem is that, each time a batch is loaded, PyTorch’s DataLoader calls the __getitem__()function on the DataSet <em>once per example </em>and concatenates them, rather than reading a batch in one go as a big chunk! So we don’t end up making use of the advantages of our tabular data set. This is especially bad when we use large batch sizes.</p><p><strong>How can we fix this?</strong> Easy — replace the first two lines above with the two lines below, and copy the definition of FastTensorDataLoader from <a href="https://github.com/hcarlens/pytorch-tabular/blob/master/fast_tensor_data_loader.py">this file</a> (credit for this goes to <a href="https://cs.stanford.edu/~muj/">Jesse Mu</a>, for <a href="https://discuss.pytorch.org/t/dataloader-much-slower-than-manual-batching/27014/6">this answer</a> on the PyTorch forums):</p><pre>train_batches = FastTensorDataLoader(train_x, train_y, batch_size=1024, shuffle=False)</pre><p>FastTensorDataLoader is just a small custom class with no dependencies other than PyTorch —and using it doesn’t require any changes to your training code! It supports shuffling too, though the benchmarks below are for non-shuffled data.</p><p><strong>What difference does this make? </strong>On the benchmark set I used, the custom tabular DataLoader ends up being <strong>over 20x faster.</strong> In this case, that means that instead of a 10-epoch run taking over 15 minutes, it takes less than 40 seconds— a huge difference in iteration speed!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/743/1*xbCdMc0ItlYlftPz5VjNXA.png" /><figcaption>Two near-identical runs — except one takes over 15 minutes, and the other takes less than a minute!</figcaption></figure><p>This benchmark was run on the <a href="http://archive.ics.uci.edu/ml/datasets/HIGGS">Higgs dataset</a> used in <a href="https://www.nature.com/articles/ncomms5308">this Nature paper</a>. With 11m examples, it makes for a more realistic deep learning benchmark than most public tabular ML datasets (which can be tiny!). It’s a binary classification problem, with 21 real-valued features. It’s nice to see that we can get to over 0.77 ROC AUC on the test set within just 40s of training, before any hyperparameter optimisation! Though we’re still a while off from the 0.88 reached in the paper.</p><p>I hope this has been helpful, and that you’re able to see similar speed increases in your own training code! After implementing this I found some further optimisations which resulted in a total speedup of closer to 100x! Leave a comment if you’d like to see more, and we can cover these in a follow-up article.</p><p>See the Appendix for how you can run the benchmark code yourself. The example includes code for running the default PyTorch DataLoader, the faster custom one, as well as timing the results and logging to TensorBoard.</p><p><em>About the author: Harald worked as an equities quant for over eight years, and is currently an independent researcher and writer. He writes at </em><a href="http://www.harald.co./"><em>harald.co</em></a><em> and </em><a href="https://joltml.com/"><em>joltml.com</em></a><em>. He runs </em><a href="http://mlcontests.com/"><em>ML Contests</em></a><em> — a directory of ongoing machine learning competitions — as well as a </em><a href="https://cloud-gpus.com/"><em>cloud GPU comparison</em></a><em> site.</em></p><p><em>By the same author:</em></p><ul><li><a href="https://medium.com/@hcarlens/five-tools-to-supercharge-your-cloud-deep-learning-workflow-3afb2e25be9f"><em>Five Tools to Supercharge Your Cloud Deep Learning Workflow</em></a></li><li><a href="https://medium.com/@hcarlens/reproducibility-issues-using-openai-gym-8ca605071d6a"><em>How to Properly Set Random Seeds with OpenAI Gym</em></a></li><li><a href="https://medium.com/machine-learning-insights/top-cloud-gpu-providers-for-machine-learning-in-2022-894f6828da08">Top Cloud GPU Providers for Machine Learning in 2022</a></li></ul><h3>Appendix: Running The Benchmark Test</h3><p>So you can see the results for yourself, here are the instructions for replicating the experiment. If you have a local GPU and PyTorch already installed, you can skip the first two steps!</p><ol><li>Create a new Ubuntu 18.04 instance with your favourite GPU cloud provider <a href="https://cloud-gpus.com/?ref=tds">take your pick of over 30 cloud GPU providers here</a>).</li></ol><p>2. Install CUDA and PyTorch in one go by using <a href="https://lambdalabs.com/lambda-stack-deep-learning-software">Lambda Stack</a>: (don’t forget to reboot after doing this!)</p><pre>LAMBDA_REPO=$(mktemp) &amp;&amp; \<br>wget -O${LAMBDA_REPO} <a href="https://lambdalabs.com/static/misc/lambda-stack-repo.deb">https://lambdalabs.com/static/misc/lambda-stack-repo.deb</a> &amp;&amp; \<br>sudo dpkg -i ${LAMBDA_REPO} &amp;&amp; rm -f ${LAMBDA_REPO} &amp;&amp; \<br>sudo apt-get update &amp;&amp; \<br>sudo apt-get — yes upgrade &amp;&amp; \<br>sudo apt-get install — yes — no-install-recommends lambda-server &amp;&amp; \<br>sudo apt-get install — yes — no-install-recommends nvidia-headless-440 nvidia-utils-440 &amp;&amp; \<br>sudo apt-get install — yes — no-install-recommends lambda-stack-cuda</pre><p>3. Download the data set:</p><pre>wget <a href="http://archive.ics.uci.edu/ml/machine-learning-databases/00280/HIGGS.csv.gz">http://archive.ics.uci.edu/ml/machine-learning-databases/00280/HIGGS.csv.gz</a></pre><p>4. Clone the repository:</p><pre>git clone <a href="/service/mailto:git@github.com">git@github.com</a>:hcarlens/pytorch-tabular.git</pre><p>5. Run the benchmark script:</p><pre>python3 pytorch-tabular/higgs_benchmark.py</pre><p>If you’re running on an instance with a GTX 1080Ti like the Genesis Cloud one I used, you should get results like these:</p><pre>ubuntu@genesis:~$ python3 pytorch-tabular/higgs_benchmark.py<br>2020-04-12 15:05:55.961134: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.0<br>Epoch 0 done.<br>Epoch 1 done.<br>Epoch 2 done.<br>Epoch 3 done.<br>Epoch 4 done.<br>Epoch 5 done.<br>Epoch 6 done.<br>Epoch 7 done.<br>Epoch 8 done.<br>Epoch 9 done.<br>Epoch 0 done.<br>Epoch 1 done.<br>Epoch 2 done.<br>Epoch 3 done.<br>Epoch 4 done.<br>Epoch 5 done.<br>Epoch 6 done.<br>Epoch 7 done.<br>Epoch 8 done.<br>Epoch 9 done.<br>Standard dataloader: 124.55s/epoch.<br>Custom dataloader: 5.24s/epoch.</pre><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e264b9e34352" width="1" height="1" alt=""><hr><p><a href="https://medium.com/data-science/better-data-loading-20x-pytorch-speed-up-for-tabular-data-e264b9e34352">Better Data Loading: 20x PyTorch Speed-Up for Tabular Data</a> was originally published in <a href="https://medium.com/data-science">TDS Archive</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Reproducibility issues using OpenAI Gym]]></title>
            <link>https://hcarlens.medium.com/reproducibility-issues-using-openai-gym-8ca605071d6a?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/8ca605071d6a</guid>
            <category><![CDATA[reinforcement-learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[reproducibility]]></category>
            <category><![CDATA[openai-gym]]></category>
            <category><![CDATA[pytorch]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Tue, 30 Jul 2019 22:43:05 GMT</pubDate>
            <atom:updated>2020-04-17T13:53:25.311Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This post was first published on </em><a href="https://harald.co/"><em>www.harald.co</em></a><em>.</em></p><p>Reproducibility is hard.</p><p>Last week I wrote a simple <a href="http://karpathy.github.io/2016/05/31/rl/">Reinforcement Learning</a> agent, and I ran into some reproducibility problems while testing it on <a href="https://gym.openai.com/envs/CartPole-v0/">CartPole</a>. This should be one of the simplest tests of an RL agent, and even here I found it took me a while to get repeatable results.</p><p>I was trying to follow <a href="http://karpathy.github.io/2019/04/25/recipe/">Andrej Karpathy</a> and <a href="http://amid.fish/reproducing-deep-rl">Matthew Rahtz</a> ‘s recommendations to focus on reproducibility and set up random seeds early, but this was taking me much longer than expected — despite adding seeds everywhere I thought necessary, sometimes my agent would learn a perfect policy in a few hundred episodes, whereas other times it didn’t find a useful policy even after a thousand episodes.</p><p>I checked the obvious — setting seeds for PyTorch, NumPy, and the OpenAI gym environment I was using. I even added a seed for Python’s <em>random</em> module, even though I was pretty sure I didn’t use that anywhere.</p><pre>RANDOM_SEED = 0 torch.manual_seed(RANDOM_SEED) env.seed(RANDOM_SEED) np.random.seed(RANDOM_SEED) random.seed(RANDOM_SEED)</pre><p>Still I got different results on each run. I found a few resources pointing me to other things to check:</p><ul><li>Consistency in data preparation and processing (not really relevant here- all the data I’m processing comes from the gym environment)</li><li>CuDNN specific seeding in PyTorch (my network is small enough to run quickly on CPU, so I’m not using CuDNN)</li></ul><p>Out of ideas, I returned to debugging. My initial policy and target network weights were the same each run. Good. The first environment observation was the same too. Also good. But then, when I came to selecting a random action, I noticed env.action_space.sample() sometimes gave different results. Bad.</p><p>I looked through the OpenAI gym code for random seeds, and couldn’t find any seeding being done on the action space, even when the environment is passed a specific seed! I then found <a href="https://github.com/openai/gym/commit/58e6aa95e5af2c738557431f812abb81c505a7cf#diff-a5b46f3fbe92ca4207f3dfb21bd66f8b">this commit</a>, where Greg Brockman and others discuss how seeding should be done in OpenAI Gym environments. It looks like they initially wanted to seed action spaces as well as environments, but decided not to because they see action space sampling as belonging to the agent rather than the environment.</p><p>So here’s the solution, in one extra line:</p><pre>env.action_space.seed(RANDOM_SEED)</pre><p>I’d love to know why this isn’t called from env.seed()!</p><p>Anyway, now I’m getting reproducible results. To get an idea of how significant the difference is between different seeds even on a problem as simple of CartPole, here are five runs with different seeds:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/513/0*1SpNSpk4I46tm-7v.png" /></figure><p>Some resources with other useful reproducibility suggestions:</p><p><em>Originally published at </em><a href="https://harald.co/2019/07/30/reproducibility-issues-using-openai-gym/"><em>http://harald.co</em></a><em> on July 30, 2019.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8ca605071d6a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[NeurIPS 2018 Day 3: Reproducibility, Robustness, and Robot Racism]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://hcarlens.medium.com/neurips-2018-day-3-reproducibility-robustness-and-robot-racism-82b0d52c8715?source=rss-711d325d27b4------2"><img src="https://cdn-images-1.medium.com/max/700/0*oyNJJk8pYPbJX9g7" width="700"></a></p><p class="medium-feed-snippet">This post was first published on www.harald.co.</p><p class="medium-feed-link"><a href="https://hcarlens.medium.com/neurips-2018-day-3-reproducibility-robustness-and-robot-racism-82b0d52c8715?source=rss-711d325d27b4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://hcarlens.medium.com/neurips-2018-day-3-reproducibility-robustness-and-robot-racism-82b0d52c8715?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/82b0d52c8715</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[reproducibility]]></category>
            <category><![CDATA[neurips-2018]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Thu, 06 Dec 2018 16:10:01 GMT</pubDate>
            <atom:updated>2020-04-17T14:00:18.360Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[NeurIPS 2018 Day 2: Cronenbergs]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://hcarlens.medium.com/neurips-2018-day-2-cronenbergs-9b1476e6bd89?source=rss-711d325d27b4------2"><img src="https://cdn-images-1.medium.com/max/700/0*HDTz-qLqkMfU3TSG" width="700"></a></p><p class="medium-feed-snippet">This story was first posted on www.harald.co.</p><p class="medium-feed-link"><a href="https://hcarlens.medium.com/neurips-2018-day-2-cronenbergs-9b1476e6bd89?source=rss-711d325d27b4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://hcarlens.medium.com/neurips-2018-day-2-cronenbergs-9b1476e6bd89?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/9b1476e6bd89</guid>
            <category><![CDATA[neurips-2018]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[biology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Wed, 05 Dec 2018 15:41:17 GMT</pubDate>
            <atom:updated>2020-04-17T13:58:54.598Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[NeurIPS 2018 Day 1: Tutorials]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://hcarlens.medium.com/neurips-2018-day-1-tutorials-1d33fbb0e59d?source=rss-711d325d27b4------2"><img src="https://cdn-images-1.medium.com/max/700/0*TARzVhkqs2bujzr1" width="700"></a></p><p class="medium-feed-snippet">This story was first published on www.harald.co.</p><p class="medium-feed-link"><a href="https://hcarlens.medium.com/neurips-2018-day-1-tutorials-1d33fbb0e59d?source=rss-711d325d27b4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://hcarlens.medium.com/neurips-2018-day-1-tutorials-1d33fbb0e59d?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/1d33fbb0e59d</guid>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neurips-2018]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Wed, 05 Dec 2018 00:10:24 GMT</pubDate>
            <atom:updated>2020-04-17T13:57:47.899Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[NeurIPS 2018 Day 0: Expo]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://hcarlens.medium.com/neurips-2018-day-0-expo-580e3f7e592f?source=rss-711d325d27b4------2"><img src="https://cdn-images-1.medium.com/max/700/0*qO_9Ef-gbEU5O8Zu" width="700"></a></p><p class="medium-feed-snippet">This story was first published on www.harald.co</p><p class="medium-feed-link"><a href="https://hcarlens.medium.com/neurips-2018-day-0-expo-580e3f7e592f?source=rss-711d325d27b4------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://hcarlens.medium.com/neurips-2018-day-0-expo-580e3f7e592f?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/580e3f7e592f</guid>
            <category><![CDATA[neurips]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[neurips-2018]]></category>
            <category><![CDATA[neural-networks]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Mon, 03 Dec 2018 03:26:58 GMT</pubDate>
            <atom:updated>2020-04-17T13:56:29.675Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[Reading and remembering]]></title>
            <link>https://hcarlens.medium.com/reading-and-remembering-dc7c56562c68?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/dc7c56562c68</guid>
            <category><![CDATA[books]]></category>
            <category><![CDATA[reading]]></category>
            <category><![CDATA[nonfiction]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Fri, 21 Sep 2018 16:26:05 GMT</pubDate>
            <atom:updated>2018-09-22T07:17:04.826Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This post was first published on </em><a href="https://harald.co/2018/09/21/reading-and-remembering/"><em>www.harald.co</em></a><em>.</em></p><p>It’s easier to read than to remember what you’ve read. I used to struggle to remember what a book was about, even just a few years after reading it.</p><p>I don’t have this problem anymore. In a few recent conversations about books, people have asked me how I manage to remember so much about books I read a long time ago. I’ve also been surprised at how clear the structure of certain books is in my mind.</p><p>I don’t think it’s because my brain has got better at remembering things. I think I’ve picked up habits from various places that put the information from books into my brain so that I remember it better. I now have a simple but powerful approach I use when reading most non-fiction books (i.e. books I want to remember rather than just enjoy).</p><p>The basic structure of this approach is:</p><ol><li>Get a broad overview of the book.</li><li>Read the book, slowly, noting key ideas and passages.</li><li>Summarise the book.</li><li>Occasionally review the summary.</li></ol><h3>Getting an overview</h3><p>This first step happens almost automatically for most people before they decide to read a book. It involves some meta-reading, and answering a few questions:</p><ul><li>What is this book about?</li><li>Who is the author? What is their background? Why did they write this book?</li><li>When was this book written?</li><li>Which related books are out there? Why read this one rather than the others?</li><li>What is the structure of the book? What are the key ideas?</li></ul><p>As part of this process I usually skim through the book, reading the introduction, the contents, the first few lines in each chapter, and the conclusion.</p><p>The nature of a book means that an author writing about lots of interconnected ideas is forced them to write about them sequentially. If you know what’s coming later in the book, you can make links to that material while reading the first few chapters. This is why some people like to read books multiple times. (it’s also why <a href="https://recurringdevelopments.com/">some TV shows</a> are better the second time you watch them)</p><h3>Reading</h3><p>I read non-fiction on Kindle, unless there are important visual elements (e.g. <a href="https://www.amazon.co.uk/Douglas-Hofstadter-Eternal-anniversary-Anniversary/dp/B00I60P54M/ref=sr_1_6?ie=UTF8&amp;qid=1537542712&amp;sr=8-6&amp;keywords=godel+escher+bach_">Godel, Escher, Bach</a>) or the book is only available in print. The neat thing on Kindle is that you can easily highlight pieces of text, add notes, and look up words or concepts.</p><p>If I’m reading a physical book I’ll either mark passages with a pencil and <a href="https://tim.blog/2007/12/05/how-to-take-notes-like-an-alpha-geek-plus-my-2600-date-challenge/">keep an index at the front of the book</a> or write down key ideas and pages on a separate piece of paper, or take pictures of passages.</p><p>I read slowly. This is intentional. With many books, it’s possible to read them and get the key ideas in just a few hours. <a href="https://www.quora.com/How-do-you-read-3-books-a-week">Some people</a> advocate spending time increasing your reading speed so you can easily read multiple books per week. (though if reading more is your main focus, you could consider <a href="https://www.blinkist.com/">reading book summaries instead</a>)</p><p>It’s true that reading faster allows you to read more. But if you want to remember what you read, and generate useful knowledge and incremental understanding, it’s important to integrate new ideas with your existing knowledge. For the special case of self-help books, it’s crucial that you spend time thinking about how to apply the contents of the book to your life, and to set up processes for doing so. Ideas alone won’t change anything.</p><p>Reading slowly allows you to stop when you encounter new ideas, and ask:</p><ul><li>How does this relate to other ideas in this book?</li><li>How does this relate to ideas in other books? (e.g. on reading The Righteous Mind, you might think “wow, this Elephant/Rider analogy sounds a lot like the System 1/System 2 concept in Thinking, Fast and Slow)</li><li>Does this idea conflict with anything else I’ve read, or anything I already believe? Is there a way to resolve this conflict (e.g. the two ideas apply in different situations)? If not, which idea do I think is most likely to be true, and why? What might I want to read next to help me figure out which of these ideas is most likely to be right?</li><li>How does this change my view of the world? (or my <a href="https://www.lesswrong.com/posts/a7n8GdKiAZRX86T5A/making-beliefs-pay-rent-in-anticipated-experiences">expectations of future experiences</a>)</li></ul><p>This process of reading slowly integrates the new ideas in the book solidly within your existing knowledge base, by building lots of links with things you’re already familiar with.</p><h3>Summarising</h3><p>When I read a great book with lots of new ideas, I try to write my own brief summary of it. For a book I read on Kindle, this is really straightforward — I’ll <a href="https://www.clippings.io/">export</a> my highlights and notes into <a href="https://www.evernote.com/referral/Registration.action?sig=9f29f5d3e56cd8f760d7bb3694fc89b6151353dc8faaf0fd56f013a716bce7bf&amp;uid=107212089">Evernote</a>, and shape those into a short summary of the key ideas. I’ll usually wait at least a few weeks after finishing a book to do this, so I’ve had a chance to digest the ideas properly. Sometimes the summary will be written specifically for myself, with lots of shorthand. Other times it’s more general, like <a href="https://harald.co/2016/02/07/influence-by-robert-cialdini/">this summary of Influence by Robert Cialdini</a>. More examples on <a href="https://sivers.org/book">Derek Sivers’ website</a>.</p><p>Sometimes I don’t get around to editing my notes into a summary, but having the highlights in Evernote is still useful for easily searching through the key bits of the book later. If I read a paper version of a book, I might just put all the pictures I took of the pages into one Evernote file.</p><h3>Reviewing</h3><p>The last part of this process is to occasionally review or re-read your own summary. Some triggers for doing this:</p><ul><li>When you read a related book. (e.g. a book with a different view on the same topic, or a book that builds on the topic)</li><li>When you want to apply the knowledge (e.g. re-reading notes on a negotiation book before you talk to your boss about your salary)</li><li>Thinking back to a book and realising you can’t remember it as clearly as you’d like</li></ul><p>Skimming over your summary is usually sufficient. I just finished reading <a href="https://www.amazon.co.uk/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111">Superintelligence</a>, and there were so many new concepts that I’ve started toying with the idea of adding key concepts into a <a href="https://en.wikipedia.org/wiki/Spaced_repetition">spaced repetition</a> tool like an <a href="https://apps.ankiweb.net/">Anki deck</a>, and reviewing those regularly. This can be taken further by using a system like like Piotr Woźniak’s <a href="https://www.supermemo.com/help/read.htm">incremental reading</a>.</p><p><em>*I’ve long wanted to publish a series of self help books with titles along the lines of “the simple two-step guide to improved productivity”. Step one, on page one, would be “stop reading self-help books”. Step two, on page two, would be “go and do the things you want to do”. The following two hundred pages would be blank, but there to legitimise the book and help it blend in along the other self-help books containing similar amounts of useful information.</em></p><p><em>Originally published at </em><a href="http://harald.co/2018/09/21/reading-and-remembering/"><em>harald.co</em></a><em> on September 21, 2018.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dc7c56562c68" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Segmenting communication]]></title>
            <link>https://hcarlens.medium.com/segmenting-communication-121219725152?source=rss-711d325d27b4------2</link>
            <guid isPermaLink="false">https://medium.com/p/121219725152</guid>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[work]]></category>
            <category><![CDATA[productivity]]></category>
            <category><![CDATA[messaging]]></category>
            <category><![CDATA[deep-work]]></category>
            <dc:creator><![CDATA[Harald Carlens]]></dc:creator>
            <pubDate>Thu, 30 Aug 2018 16:02:06 GMT</pubDate>
            <atom:updated>2018-09-09T09:37:20.404Z</atom:updated>
            <content:encoded><![CDATA[<p><em>This post was first published on </em><a href="https://harald.co/2018/08/30/segmenting-communication/"><em>www.harald.co</em></a><em>.</em></p><p>I’ve been called old-fashioned when it comes to communication. Some people don’t understand why I like email so much. (I love email. It’s the best. Please email me.)</p><p>One reason is that it’s the only good way to send messages which are clearly non-urgent, and can be easily tracked. Why is this important? I think it saves everyone a lot of time and attention, by removing unnecessary interruptions.</p><h3>Message urgency</h3><p>Communications can be segmented in various ways, one of which is by urgency. Most people would agree that there’s a relatively clear one-dimensional spectrum of message urgency, from “I found a cool cat picture” to “I’m outside your door” to “My house is on fire”. Broadly we could refer to these as low-urgency (whenever), medium-urgency (as soon as convenient), and high-urgency (right now).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/371/0*I3FMZQZuR2r6KjtC.gif" /></figure><p>A hundred years ago, when our choices of communication were limited, things were pretty straightforward. Low urgency? Send a letter. Medium urgency? A telegram might do. High urgency? Phone call.</p><p>Twenty years ago, things were similar. You might send a letter or email for a low-urgency message. Medium urgency could be a text, or a message through an IM platform such as MSN, AIM, or IRC. High urgency would be a phone call, ideally to someone’s mobile if they had one.</p><p>Now we’re in a world where a significant amount of communication goes through platforms like Facebook Messenger, WhatsApp, WeChat, Snapchat, and Instagram, where messages occupy an usually large portion of the urgency spectrum. While there are <a href="https://www.economist.com/graphic-detail/2018/05/18/how-heavy-use-of-social-media-is-linked-to-mental-illness">other reasons</a> these platforms may not be ideal, I think there’s a high productivity cost associated with the lack of urgency segmentation in messages sent through these platforms.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/0*LSVXZqDggxDHunNd.jpg" /></figure><h3>Cost of interruptions</h3><p>There might be benefits to multi-tasking in some situations, but generally it’s considered <a href="https://www.joelonsoftware.com/2001/02/12/human-task-switches-considered-harmful/">bad for productivity</a>. Taking this idea further, the concept of <a href="http://calnewport.com/blog/2015/11/20/deep-work-rules-for-focused-success-in-a-distracted-world/">Deep Work</a> — long periods of uninterrupted focus — has become popular in recent years, and associated with success in challenging fields.</p><p>The interruption cost is worth paying in the case of an urgent message. But because of the lack of segmentation, a large number of non-urgent messages also incur this cost. <a href="https://www.statista.com/chart/13762/whatsapp-messages-sent-per-day/">The average WhatsApp user receives around 50 messages per day</a>. How many of those are urgent to the extent of needing to be seen within a few minutes? Almost certainly no more than a handful. Yet for most people those will all trigger a sound or vibration alert, distracting them from what they’re doing at the time.</p><p>Tools for dealing with this are fairly limited. Some options:</p><p>Two things I’d really like to see (please let me know if you know of a way to do this!):</p><ul><li><em>Opt-in, rather than opt-out, to notifications</em>: this should really be a standard feature on mobile operating systems. Currently whenever I install a new app I then have to go and disable notifications separately.</li><li><em>Prod</em>: disable notifications by default for a messaging app, but give users the option to notify others of urgent messages on a per-message basis.</li></ul><h3>Productivity</h3><p>It might seem a bit misguided to talk about social networking apps in terms of productivity, since productivity might not be something you particularly care about. However, increased productivity will tend to mean you spend less time doing things you don’t like and more time doing things you enjoy, so even if you’re not interested in increasing your output, you’ll probably gain from investing in increased productivity.</p><p>Equally, interruptions detract from experiences where productivity isn’t involved. Watching a film, going for a walk, or just having dinner with someone are all generally better without unnecessary notifications.</p><p>In certain areas productivity can be a final goal — such as at work. Now that IM (Skype, Slack, Bloomberg…) is a staple in the office, it’s a huge attention drain. I think in the vast majority of cases it probably does more harm than good, especially in a culture that overly values quick responses. Note that here, email is only a good solution if it’s used properly. As always, <a href="http://calnewport.com/blog/2016/04/19/write-longer-emails/">Cal Newport has some suggestions</a>.</p><p><em>Originally published at </em><a href="https://harald.co/2018/08/30/segmenting-communication/"><em>harald.co</em></a><em> on August 30, 2018.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=121219725152" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>