Spaces:
Runtime error
Runtime error
| # -*- coding: utf-8 -*- | |
| """4_3-gradio-and-huggingface-spaces.ipynb | |
| Automatically generated by Colaboratory. | |
| Original file is located at | |
| https://colab.research.google.com/drive/1ML3Jf1UwkDRuEPK7NoVr1Uel9tWa_oP7 | |
| # Gradio Interfaces and HuggingFace Spaces | |
| Huggingface [Spaces](https://huggingface.co/spaces) provide an easy-to-use way to explore and demo models. The platform is highly accessible, free to use, and allows you to share models without the need for the user to run any code. | |
| The best part - you can insert your own model from huggingface, build your app with [gradio](https://gradio.app/docs/), and deploy in no time! | |
| Let's use the model that we generated in the `4_1-text-classification-finetune-solns.ipynb` notebook and create a gradio space to demonstrate it! | |
| ## Install and Import Packages | |
| """ | |
| # Commented out IPython magic to ensure Python compatibility. | |
| # %%capture | |
| # !pip install gradio transformers | |
| # import necessary libraries | |
| import gradio as gr | |
| import numpy as np | |
| from transformers import AutoModelForSequenceClassification, AutoTokenizer | |
| from huggingface_hub import notebook_login | |
| #!git config --global credential.helper store | |
| #notebook_login() | |
| """## Load in Your Model | |
| Next, we'll load in our model from huggingface. This should be in a HF repo under your name, probably formatted `your-username/model-name`. | |
| We'll use the `Auto` classes to load in this model. The `Auto` classes in the Hugging Face transformers library are designed to automatically infer the correct model architecture or tokenizer based on the model checkpoint provided. | |
| For example, below, AutoModelForSequenceClassification is specifically designed for sequence classification tasks, such as text classification or sentiment analysis (which is what bert-emotion was). If you've fine-tuned a model for a different type of task, like question answering or named entity recognition, you would need to use a different auto model class that corresponds to that task. For example, for question answering, you might use AutoModelForQuestionAnswering. | |
| To ensure the right model class is used, you should use the appropriate auto model class based on the task your model was fine-tuned for. You can look at the config.json file associated with a model checkpoint to see the type of model. (You can also use this model name directly - but the `Auto` classes will give you more flexibility!) | |
| [ See more about Auto classes [here](https://huggingface.co/docs/transformers/model_doc/auto#auto-classes). ] | |
| """ | |
| # specify the model name | |
| # replace 'your-username/model-name' with the name of your custom trained model | |
| model_name = 'apetulante/bert-emotion' | |
| # initialize the model and tokenizer | |
| model = AutoModelForSequenceClassification.from_pretrained(model_name) | |
| tokenizer = AutoTokenizer.from_pretrained(model_name) | |
| """Let's also define our labels so we know how to interpret the output from the model.""" | |
| labels = {0: 'anger', 1: 'joy', 2: 'optimism', 3: 'sadness'} | |
| """## Define and Create the Gradio Interface | |
| Next, we'll define a function that will do the sentiment analysis task for us. A lot of this should look very similar to how we did basic inferencing with Huggingface, because now that we've pushed our model there, we can grab it just like any other model! | |
| """ | |
| # Define the prediction function | |
| def predict_sentiment(text): | |
| # Tokenize the input tweet using the tokenizer | |
| inputs = tokenizer.encode_plus( | |
| text, | |
| add_special_tokens=True, # Add special tokens for BERT | |
| truncation=True, # Truncate the input if it exceeds the maximum sequence length | |
| padding='longest', # Pad the input sequences to the length of the longest sequence | |
| return_tensors='pt' # Return PyTorch tensors | |
| ) | |
| # Pass the tokenized inputs to the model | |
| outputs = model(**inputs) | |
| # Get the predicted class by finding the index of the highest logit score | |
| logits = outputs.logits.detach().numpy() | |
| predicted_class = np.argmax(logits, axis=1).item() | |
| # Map the predicted class index to the corresponding sentiment label using the labels dictionary | |
| sentiment_label = labels[predicted_class] | |
| # Return the predicted sentiment label | |
| return sentiment_label | |
| predict_sentiment("okay,let's go!") | |
| """Let's define the Gradio interface with `sentiment_analysis` as the function that takes user inputs and generates outputs. The `inputs` argument specifies the input component, in this case a textbox where users can enter text. The `outputs` argument specifies the type of the output, in this case a simple text.""" | |
| # Define the Gradio interface | |
| iface = gr.Interface( | |
| fn=predict_sentiment, | |
| inputs="text", | |
| outputs="text", | |
| title="Sentiment Analysis", | |
| description="Enter a tweet and get its sentiment prediction.", | |
| examples=[ | |
| ["I'm furious right now."], | |
| ["I have been feeling amazing lately!"], | |
| ["I think that everything is going to turn out okay."], | |
| ["Feeling really down today."], | |
| ] | |
| ) | |
| # Run the Gradio interface | |
| iface.launch() | |
| """You may notice a "flag" option here. The flag functionality is a default feature in Gradio. When you launch a Gradio interface, you'll notice a "Flag" button alongside each input-output pair. Clicking this button allows you to flag examples where the model's output may not be correct or as expected. | |
| We can view these flagged examples in the `log.csv` file that will be saved in the `flagged` folder to the left. | |
| ## Turn it into a Huggingface Space! | |
| Simply turn this code into a app.py file, and create a huggingface space. Since the model is already hosted on huggingface, you should be up and running in no time! | |
| """ | |
| """## Optional Homework | |
| We've just touched the surface of what gradio can do here, but there are a TON of other options of cool features to add or things to do with gradio. Try out a few on your own! | |
| The code to create the gradio space is also fairly short. You can try giving the code to make this space to ChatGPT, and ask it to help you come up with additional features. | |
| """ | |
| #@title Add Confidence Information | |
| #@markdown With each of these predictions, the model has some confidence | |
| #@markdown that the given prediction is correct. | |
| #@markdown It can be useful to display the relative prediction confidence | |
| #@markdown for *all* classes, so we can know if the model was less sure of | |
| #@markdown an answer | |
| #@title Predict in Batch | |
| #@markdown Often, it's convenient to use a gradio space to allow | |
| #@markdown users to predict on a batch of inputs. | |
| #@markdown Imagine you have a text file with a new tweet to determine the sentiment | |
| #@markdown of on each line. How can you edit this gradio space to accept | |
| #@markdown and return a .txt file? | |
| #@title Try Visualizations | |
| #@markdown With a batch prediction, there's an opportunity | |
| #@markdown to try visualizations with the data. | |
| #@markdown Try to show a pie or bar chart of the sentiments of a batch. |