This article originated from personal public account: TechFlow, original is not easy, for attention


After I recommended the free Spark cluster last time, many friends asked me if I had a good cloud GPU platform recommendation. I have not recommended it to you, the main reason is that I use Mac all the time, so I don’t have a deep understanding of GPU configuration. However, I have used several cloud GPU platforms, so TODAY I would like to talk with you briefly.

Colab

Let’s start with free. The most famous free platform is Colab. Colab is a free cloud service provided by Google and also supports GPU, so we can use it to do deep learning.

Colab is embedded in Google Drive. We first open Google Drive, then go to New -> More -> Associate more apps.


Then type Colab in the box and click Install.


After the installation is complete, we can click New and select Colab files from more. When the file is opened, a Notebook opens.


After opening it, we go to Edit -> Notebook Settings, and then we can select hardware, we select GPU. Now we have a TPU, I think it’s a Tensor unit, but I haven’t used it before, if you’re interested.


Colab integrates TensorFlow, Pytorch, Keras, OpenCV and other mainstream frameworks. Pytorch can be used to verify whether the GPU has been mounted.


The usage of this Notebook is basically the same as that of the local Notebook, so we can use magic commands to install packages that are not available in the environment. For example, if you find pandas does not have one, we can install it like this:

! pip install pandasCopy the code

Colab also allows you to upload files, clone code from Github, and read files from Google Drive. Considering it’s a completely free platform, that’s pretty good.

However, the disadvantages are also obvious. One is that we cannot visit China. Another is that the overall process is not very stable. And Notebook itself is only good for small, experimental tasks, so if you want to seriously train a model, I don’t feel very good about it.

Kaggle

Kaggle is a popular machine learning algorithm contest platform with a wide variety of machine learning problems and data sets. Like Google’s Colab, Kaggle also has a free cloud GPU Notebook to use.

We went to Kaggle’s home page and hit the New Notebook from the Notebook Notebooks.


In the box we select Python and Notebook.

In Settings on the right, Accelerator selects GPU. Note that Kaggle requires you to bind your phone if you are using it for the first time. When we select GPU, a pop-up box will pop up telling you that each account can only use GPU for 30 hours per week.


Let’s run PyTorch and see that the GPU is already mounted.


If you want to see the details of the GPU can still run! Nvidia-smi command to check.


You can see that the GPU is Telsa P100, and you can also see the VERSION of CUDA and the information about memory. I looked it up and found it was a good setup, though it was only available for 30 hours a week, but it was free and enough for experimentation.

Drops of cloud

As you can see, both Colab and Kaggle are good, but they are only good for making small models and doing experimental things. When you want to really “make a big story,” it’s a little short. It seems to want to spend money or not, money can become stronger (fog). Therefore, I would like to recommend some platforms that require payment. Comparatively speaking, although it costs money, the experience is much better. Moreover, these platforms are carefully selected by me and are not very expensive, so even students can certainly afford them.

First, let’s introduce didi Cloud, which is my recently discovered treasure. It’s only recently launched, so probably not many people have heard of it.

To buy its Notebook service, we went to app.didiyun.com to register and verify our real name. Notebook provides a GPU. Let’s look at the configuration:


Notebook is only available on a P4 GPU, and the combination of a quad-core CPU and a 120GB cloud drive takes 3.2 hours. That’s more than half the price of at least eight of the same specs in the Geek Cloud. And, for the record, I got an internal coupon code. Enter my master code 2323 and get 10% off the original price, just over two pieces.


If Notebook doesn’t cut it for you, you can also choose its cloud GPU server, which has more configurations to choose from:


Also enter the didi cloud master code 2323, the same can enjoy 10% discount.

Geek cloud

Before the emergence of Didi Cloud, Geekcloud was also the GPU cloud platform I used a lot. It was made by a domestic manufacturer. Its biggest advantage is that it can pay by wechat/Alipay, which is very convenient XD. In addition, it supports both Notebook and SSH connections, and supports resource sharing and provides a large amount of shared data. All major public data sets can be found in its shared data sets, so you don’t need to download them.

Another cool thing about geek Cloud is that it supports exclusive and shared modes, where gpus can’t be used, but they can be used very cheaply. In the exclusive mode, we can use the entire GPU by ourselves to ensure efficiency. This model has to be said to be very good, sometimes we want to keep our machine running, but do not want to be too expensive, this time can turn on the sharing model to save money. And it also supports monthly and weekly packages, relatively cost-effective.


While some instances are cheaper, the geek cloud as a whole is expensive. For example, the Tesla P4 costs $8 an hour exclusively, and the Tesla P100 is free on Kaggle at $17 an hour. Not that it is dark, after all, it is a small platform, not many people use it, the maintenance cost is relatively high, so it is understandable that the fee is a little bit higher.


My previous kernels cost about 3 yuan an hour, which is about the same as Internet cafes. The experience is obviously better than that of Colab and Kaggle, but the platform is a bit unstable, and often the next time you restart an instance you will find that the same model is gone and need to create another instance. Although we can put the data in the shared folder to avoid loss, after all, some customized configurations will be lost, so I do not have a good experience in this regard.

Vast.ai

Rewind. Ai is a foreign cloud service rental site that supports both Notebook and cloud servers, but we can only choose one, not both.

Ai is very easy to remember. When we open the page of Creating Instance, we can see that there are many GPU configurations to choose from, and the price is very cheap. The cheapest cost $0.10 an hour, which works out to less than a dollar. In Edit Image & Config on the left, we can set our requirements, such as whether we want to use it in the form of Jupyter Notebook or SSH remote login.


The last time I used it, I was able to configure Notebook remote access after logging in using SSH. However, I used it again this time. It seems that there is something wrong with the Notebook address opened after remote login, and I cannot access it. Maybe they have upgraded the system and prohibited relevant operations.

The experience isn’t any better than Colab and Kaggle, and the documentation isn’t as thorough. The only advantage should be that it is cheap. In addition, this platform supports UnionPay, so there should be no problem to use it in China.

conclusion

Others would recommend Google’s cloud platform, which also offers GPU server rentals and a $300 free voucher for signing up. But it also has a problem, is the GPU resources are very limited, often half a day to get up, because the resources are fully occupied. For example, choose the European address of the server, basically have to wait for the European sleep time to open the machine, or very inconvenient. So Google’s cloud platform is not recommended in this article.

I hope you can find a satisfactory GPU platform through this article and complete your small goal of deep learning.

If you like this article, if you can, please click the following, give me a little encouragement, but also convenient access to more articles.

This article is formatted using MDNICE