Forem: satish

Pytorch MNIST Dataset Error

satish — Tue, 27 Apr 2021 14:55:39 +0000

There seems to be some problem with the Hosting, and the Pytorch team is solving the following issue.

Use this code to solve the issue. It provides a new mirror to download the dataset.

from torchvision import datasets,transforms
new_mirror = 'https://ossci-datasets.s3.amazonaws.com/mnist'
datasets.MNIST.resources = [
   ('/'.join([new_mirror, url.split('/')[-1]]), md5)
   for url, md5 in datasets.MNIST.resources
]
train_dataset = datasets.MNIST(
   "../data", train=True, download=True, transform=transforms.Totensor()
)

Courtesy: https://stackoverflow.com/questions/66577151/http-error-when-trying-to-download-mnist-data

NATS for .net developers - Part I

satish — Sat, 10 Oct 2020 13:08:28 +0000

Nats is a high-performance message system developed by Derek Collison. It's built right from the ground up as a cloud-native service. And it's been graduated to a Mature project with the CNCF community, Alongside with kubernetes, Prometheus, Grpc.

https://www.cncf.io/blog/2018/03/15/cncf-to-host-nats/

If you are on the path to move towards microservice and looking for a high reliability and performance messaging System, this should be your choice.

Nats, at its core, is a Publish/Subscribe system. A PubSub-Model essentially means the System can communicate with each other without knowing a great deal about other endpoints. Clients become subscribers by registering their interest in the subject, and whenever the publisher sends the message system will deliver to all the available clients.

Here the transmitting of the message from producer to consumer is done Via Nat's server component.

Delivery Guarantees or QoS

Nats by default provide At-Most once delivery guarantee and with the help Nats streaming At least once

At-Most once - Message will be delivered to the client if the client is alive;
At-least once - Message will be delivered to the client at least come. In this config, the Nats streaming server maintains a commit log of messages.

Nats don't support exactly-once delivery like Kafka. These are
some of the choices to make NATS a reliable and straightforward
messaging platform.

Performance

Nats designed for ultralight and performance-minded modern messaging middleware. In my laptop with other programs like Visual studio and chrome with 9 tabs(Yikes), I was able to bench 18 million messages per second.

Installation

Installing Nats is a Breeze, and it supports multiple platforms with different variants. In this case, I want to use docker.

docker run -d --name nats-main -p 4222:4222 -p 6222:6222 -p 8222:8222 nats

Will look at the usage from .net library in the next part.

C# Value Task

satish — Sat, 15 Aug 2020 18:07:42 +0000

System.Threading.Task is one of the de facto best practices for .net programming for async programming for all these years. C# 7.0 has added a cherry on top of it with Value Task for optimizing performance and abstraction.

Performance benefit

To illustrate the benefit in performance, Let's call an API to return the restaurant data for a particular city and cache the result. I have used caching here to show that there is a mix of synchronous and asynchronous code.

    [MemoryDiagnoser]
    public class ZomatoCitySearch
    {
        private static IDictionary<string, SearchSuggestion> CacheList = new Dictionary<string, SearchSuggestion>();

        [Benchmark]
        public async Task<SearchSuggestion> SearchCity()
        {
            if (CacheList.ContainsKey("ben"))
            {
                return CacheList["ben"];
            }
            HttpClient client = new HttpClient();
            client.DefaultRequestHeaders.Add("user-key","*****");
            var response = await client.GetAsync(
                "https://developers.zomato.com/api/v2.1/cities?q=ben");
            var content = await response.Content.ReadFromJsonAsync<SearchSuggestion>();
            CacheList["ben"] = content;
            return content;
        }

        [Benchmark]
        public async ValueTask<SearchSuggestion> SearchCityWithValueTask()
        {
            if (CacheList.ContainsKey("new"))
            {
                return CacheList["new"];
            }
            HttpClient client = new HttpClient();
            client.DefaultRequestHeaders.Add("user-key","***");
            var response = await client.GetAsync(
                "https://developers.zomato.com/api/v2.1/cities?q=new");
            var content = await response.Content.ReadFromJsonAsync<SearchSuggestion>();
            CacheList["new"] = content;
            return content; 
        }
    }

I ran the benchmark test with the Value Task and Task, Here is what the benefit

As per the result, we are using half the memory of Task and in a high-performance environment that can a lifesaver.

Value Task being a struct helps avoid the creation of the object also the internal state machine when the code is executing the synchronous way.

Abstraction

One of the problems with using Task in the abstraction is we will force the implementation to use the asynchronous code where it may not be required.

For example, If we create an ICacheprovier interface with the async Get method, We will force the same to Implemented by MemoryCacheprovider too. With the advent of ValueTask, this can be avoided and memory cache doesn't need to cost of Task creation with this model