Quickly create beautiful user interfaces around your machine learning models. Gradio (pronounced GRAY-dee-oh) makes it easy for you to demo your model in your browser or let people "try it out" by dragging-and-dropping in their own images, pasting text, recording their own voice, etc. and seeing what the model outputs.
<span>2.</span> Run the code below as a Python script or in a Python notebook (or in a [colab notebook](https://colab.research.google.com/drive/18ODkJvyxHutTN0P5APWyGFO_xwNcgHDZ?usp=sharing)).
<span>3.</span> The interface below will appear automatically within the Python notebook, or pop in a browser on [http://localhost:7860](http://localhost:7860/) if running from a script.
Gradio can wrap almost any Python function with an easy-to-use user interface. In the example above, we saw a simple text-based function. But the function could be anything from image enhancer to a tax calculator to (most commonly) the prediction function of a pretrained machine learning model.
Let's say we want to customize the input text field - for example, we wanted it to be larger and have a text hint. If we use the actual input class for `Textbox` instead of using the string shortcut, we have access to much more customizability. To see a list of all the components we support and how you can customize them, check out the [Docs](https://gradio.app/docs).
Let's say we had a much more complex function, with multiple inputs and outputs. In the example below, we have a function that takes a string, boolean, and number, and returns a string and number. Take a look how we pass a list of input and output components.
We simply wrap the components in a list. Each component in the `inputs` list corresponds to one of the parameters of the function, in order. Each component in the `outputs` list corresponds to one of the values returned by the function, again in order.
Let's try an image-to-image function. When using the `Image` component, your function will receive a numpy array of your specified size, with the shape `(width, height, 3)`, where the last dimension represents the RGB values. We'll return an image as well in the form of a numpy array.
Additionally, our `Image` input interface comes with an 'edit' button which opens tools for cropping, flipping, rotating, drawing over, and applying filters to images. We've found that manipulating images in this way will often reveal hidden flaws in a model.
In addition to images, Gradio supports other media input types, such as audio or video uploads, as well as many output components. Read about these in the [Docs](https://gradio.app/docs).
You can use Gradio to support inputs and outputs from your typical data libraries, such as numpy arrays, pandas dataframes, and plotly graphs. Take a look at the demo below (ignore the complicated data manipulation in the function!)
You can provide example data that a user can easily load into the model. This can be helpful to demonstrate the types of inputs the model expects, as well as to provide a way to explore your dataset in conjunction with your model. To load example data, you provide a **nested list** to the `examples=` keyword argument of the Interface constructor. Each sublist within the outer list represents a data sample, and each element within the sublist represents an input for each input component. The format of example data for each component is specified in the [Docs](https://gradio.app/docs).
You can load a large dataset into the examples to browse and interact with the dataset through Gradio. The examples will be automatically paginated (you can configure this through the `examples_per_page` argument of Interface) and you can use CTRL + arrow keys to navigate through the examples quickly.
You can make interfaces automatically refresh by setting `live=True` in the interface. Now the interface will recalculate as soon as the user input changes.
Your function may use data that persists beyond a single function call. If the data is something accessible to all function calls and all users, you can create a global variable outside the function call and access it inside the function. For example, you may load a large model outside the function and use it inside the function so that every function call does not need to reload the model.
Another type of data persistence Gradio supports is session **state**, where data persists across multiple submits within a page load. However, data is *not* shared between different users of your model. To store data in a session state, you need to do three things: (1) Pass in an extra parameter into your function, which represents the state of the interface. (2) At the end of the function, return the updated value of the state as an extra return value (3) Add the `'state'` input and `'state'` output components when creating your `Interface`. See the chatbot example below:
Notice how the state persists across submits within each page, but the state is not shared between the two pages. Some more points to note: you can pass in a default value to the state parameter, which is used as the initial value of the state. The state must be a something that can be serialized to a JSON format (e.g. a dictionary, a list, or a single value. Typically, objects will not work).
Underneath the output interfaces, there is a button marked "Flag". When a user testing your model sees input with interesting output, such as erroneous or unexpected model behaviour, they can flag the input for the interface creator to review. Within the directory provided by the `flagging_dir=` argument to the Interface constructor, a CSV file will log the flagged inputs. If the interface involves file data, such as for Image and Audio components, folders will be created to store those flagged data as well.
You can review these flagged inputs by manually exploring the flagging directory, or load them into the examples of the Gradio interface by pointing the `examples=` argument to the flagged directory. If you wish for the user to provide a reason for flagging, you can pass a list of strings to the `flagging_options` argument of Interface. Users will have to select one of the strings when flagging, which will be saved as an additional column to the CSV.
This generates a public, shareable link that you can send to anybody! When you send this link, the user on the other side can try out the model in their browser. Because the processing happens on your device (as long as your device stays on!), you don't have to worry about any packaging any dependencies. If you're working out of colab notebook, a share link is always automatically created. It usually looks something like this: **XXXXX.gradio.app**. Although the link is served through a gradio link, we are only a proxy for your local server, and do not store any data sent through the interfaces.
Keep in mind, however, that these links are publicly accessible, meaning that anyone can use your model for prediction! Therefore, make sure not to expose any sensitive information through the functions you write, or allow any critical changes to occur on your device. If you set `share=False` (the default), only a local link is created, which can be shared by [port-forwarding](https://www.ssh.com/ssh/tunneling/example) with specific users.
Huggingface provides the infrastructure to permanently host your Gradio model on the internet, for free! You can either drag and drop a folder containing your Gradio model and all related files, or you can point HF Spaces to your Git repository and HP Spaces will pull the Gradio interface from there. See [Huggingface Spaces](http://huggingface.co/spaces/) for more information.
You may wish to put an authentication page in front of your interface to limit who can open your interface. With the `auth=` keyword argument in the `launch()` method, you can pass a list of acceptable username/password tuples; or, for more complex authentication handling, you can even pass a function that takes a username and password as arguments, and returns True to allow authentication, False otherwise. Here's an example that provides password-based authentication for a single user named "admin":
Most models are black boxes such that the internal logic of the function is hidden from the end user. To encourage transparency, we've made it very easy to add interpretation to your model by simply setting the `interpretation` keyword in the `Interface` class to `default`. This allows your users to understand what parts of the input are responsible for the output. Take a look at the simple interface below which shows an image classifier that also includes interpretation:
In addition to `default`, Gradio also includes [Shapley-based interpretation](https://christophm.github.io/interpretable-ml-book/shap.html), which provides more accurate interpretations, albeit usually with a slower runtime. To use this, simply set the `interpretation` parameter to `"shap"` (note: also make sure the python package `shap` is installed). Optionally, you can modify the the `num_shap` parameter, which controls the tradeoff between accuracy and runtime (increasing this value generally increases accuracy). Here is an example:
This will work for any function, even if internally, the model is a complex neural network or some other black box. If you use Gradio's `default` or `shap` interpretation, the output component must be a `Label`. All common input components are supported. Here is an example with text input.
So what is happening under the hood? With these interpretation methods, Gradio runs the prediction multiple times with modified versions of the input. Based on the results, you'll see that the interface automatically highlights the parts of the text (or image, etc.) that contributed increased the likelihood of the class as red. The intensity of color corresponds to the importance of that part of the input. The parts that decrease the class confidence are highlighted blue.
You can also write your own interpretation function. The demo below adds custom interpretation to the previous demo. This function will take the same inputs as the main wrapped function. The output of this interpretation function will be used to highlight the input of each input interface - therefore the number of outputs here corresponds to the number of input interfaces. To see the format for interpretation for each input interface, check the Docs.
Here are the themes we currently support: `"default"`, `"huggingface"`, `"grass"`, `"peach"`, and the dark themes corresponding to each of these: `"darkdefault"`, `"darkhuggingface"`, `"darkgrass"`, `"darkpeach"`.
If you'd like to have more fine-grained control over any aspect of the app, you can also write your own css or pass in a css file, with the `css` parameter of the `Interface` class.
In some cases, you might like to provide your users or testers with *more* than just a binary option to flag a sample. You can provide `flagging_options` that they select from a dropdown each time they click the flag button. This lets them provide additional feedback every time they flag a sample.
Gradio integrates nicely with the Hugging Face Hub, allowing you to load models and Spaces with just one line of code. To use this, simply use the `load()` method in the `Interface` class. So:
- To load any model from the Hugging Face Hub and create an interface around it, you pass `"model/"` or `"huggingface/"` followed by the model name, like these examples:
- To load any Space from the Hugging Face Hub and recreate it locally (so that you can customize the inputs and outputs for example), you pass `"spaces/"` followed by the model name:
```python
gr.Interface.load("spaces/eugenesiow/remove-bg", inputs="webcam", title="Remove your webcam background!").launch()
```
One of the great things about loading Hugging Face models or spaces using Gradio is that you can then immediately use the resulting `Interface` object just like function in your Python code (this works for every type of model/space: text, images, audio, video, and even multimodal models):
Gradio also lets you mix interfaces very easily using the `gradio.Parallel` and `gradio.Series` classes. `Parallel` lets you put two similar models (if they have the same input type) in parallel to compare model predictions:
If many people are using your interface or if the inference time of your function is long (> 1min), simply set the `enable_queue` parameter in the `launch` method to `True` to prevent timeouts.
This sets up a queue of workers to handle the predictions and return the response to the front end. This is strongly recommended if you are planning on uploading your demo to Hugging Face Spaces (as described above) so that you can manage a large number of users simultaneously using your demo.
Gradio requires Python `3.7+` and has been tested on the latest versions of Windows, MacOS, and various common Linux distributions (e.g. Ubuntu). For Python package requirements, please see the `setup.py` file.
If you would like to contribute and your contribution is small, you can directly open a pull request (PR). If you would like to contribute a larger feature, we recommend first creating an issue with a proposed design for discussion. Please see our [contributing guidelines](https://github.com/gradio-app/gradio/blob/master/CONTRIBUTING.md) for more info.
See, also, the accompanying paper: ["Gradio: Hassle-Free Sharing and Testing of ML Models in the Wild"](https://arxiv.org/pdf/1906.02569.pdf), *ICML HILL 2019*, and please use the citation below.
```
@article{abid2019gradio,
title={Gradio: Hassle-Free Sharing and Testing of ML Models in the Wild},
author={Abid, Abubakar and Abdalla, Ali and Abid, Ali and Khan, Dawood and Alfozan, Abdulrahman and Zou, James},
