Today, thousands of open-source models are available to you, and many are released every week on the Hugging Face Hub. Hugging Face Hub is an open platform that hosts models, datasets, and machine learning demos that are called Hugging Face Spaces. How do you find a model that you need for your project? Let's head over to the Hugging Face Hub to find out. You'll find models suitable for many tasks on the Models page. The number of models here may seem overwhelming. It's a good idea to begin your search by identifying what task you're working on in machine learning terms. In this course, you'll see plenty examples of tasks. Let's say I want to do automatic speech recognition. Let's choose it from the left side panel. There are still many models to pick from, but you can narrow your search down further. Let's say you want a model to transcribe speech in French. You can choose your language here. And let's say you want a model with a permissive license. By permissive, I mean a license that allows you to use the model for most kinds of applications, including commercial use. This leaves you with much fewer options. You can sort by downloads. If you want to find models that are commonly used for this task, Or if you'd like to try a recent model that the community is excited about, you can sort by trending. Before picking one model or the other, check out their model cards. A well-written model card is like a readme file for a model. It contains a lot of useful information, such as model's architecture, how it was trained, what limitations it has, and so on. As you can see here, models can have checkpoints with varying number of parameters. So we say that this type of model comes in different sizes. Checkpoint refers to the saved model, including the pre-trained weights and all the necessary configurations. We often say we load a model, but technically speaking, we load a model checkpoint. Some checkpoints have dozens of millions of parameters. Others have a billion or a few billions of parameters. Depending on your hardware, you may not be able to run the largest checkpoints. So let me show you a rule of thumb that I use to estimate how much memory I will need for a model. We'll go to Files and Versions. Here you can find a file called PyTorchModelBin. This file stores the trained weights of the model, and you can easily see its size. Multiply that size by 1.2. In other words, add 20% on top. And this is approximately how much memory you'll need to run this model. Now let me quickly show you an alternative way to find a model for a task, a dataset, or a demo. Let's go to the Tasks page. On this page, you can learn about different machine learning tasks. Let's choose a task that we're interested in. Again, let's go with automatic speech recognition. On this page, you can learn about the task itself. So this is a great way to discover machine learning tasks that you have not worked with yet. You can also find suggestions for models that will work with this task. Datasets you can use. And find some demos where you can play with models that perform this task. Note that Whisper by OpenAI is suggested as the top choice here. Let's go back to this Models page. To load this model from the Hugging Face Hub, you can use the Transformers library. Notice the Using Transformers button. If you click it, you'll find two helpful code snippets showing how to load the model checkpoint. In this course, you'll be working with models using the Pipeline object. as in the first example. The Pipeline object offers a high level abstraction to solve tasks. It also takes care of complex preprocessing of inputs to match the model's expectations. For example, some audio models expect the input audio to come in the shape of a "logmel" spectrogram Text typically needs to be converted into so-called "tokens" And images often need to be properly resized and normalized. With the pipeline, you won't need to do any of these pre-processing steps by hand. Now that you know how to find models for your tasks and where to find the pipeline code snippet, let's build your first application. Let's go on to the next lesson.
Open Source Models with Hugging Face
- Introduction
- Selecting models
- Natural Language Processing (NLP)
- Translation and Summarization
- Sentence Embeddings
- Zero-Shot Audio Classification
- Automatic Speech Recognition
- Text to Speech
- Object Detection
- Image Segmentation
- Image Retrieval
- Image Captioning
- Multimodal Visual Question Answering
- Zero-Shot Image Classification
- Deployment
- Conclusion