Create app.py

app.py

@@ -0,0 +1,41 @@
+# Import gradio and transformers libraries
+import gradio as gr
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+
+# Load the small deberta models for hate and offensive speech detection
+hate_model = AutoModelForSequenceClassification.from_pretrained("KoalaAI/HateSpeechDetector")
+hate_tokenizer = AutoTokenizer.from_pretrained("KoalaAI/HateSpeechDetector")
+
+offensive_model = AutoModelForSequenceClassification.from_pretrained("KoalaAI/OffensiveSpeechDetector")
+offensive_tokenizer = AutoTokenizer.from_pretrained("KoalaAI/OffensiveSpeechDetector")
+
+# Define a function that takes an input text and returns the scores from the models
+def get_scores(text):
+  # Tokenize and encode the input text
+  hate_input = hate_tokenizer(text, return_tensors="pt")
+  offensive_input = offensive_tokenizer(text, return_tensors="pt")
+
+  # Get the logits from the models
+  hate_logits = hate_model(**hate_input).logits
+  offensive_logits = offensive_model(**offensive_input).logits
+
+  # Apply softmax to get probabilities
+  hate_probs = hate_logits.softmax(dim=1)
+  offensive_probs = offensive_logits.softmax(dim=1)
+
+  # Get the labels from the models
+  hate_labels = hate_model.config.id2label
+  offensive_labels = offensive_model.config.id2label
+
+  # Format the output as a dictionary of scores
+  output = {}
+  output["Hate speech"] = {hate_labels[i]: round(p.item(), 4) for i, p in enumerate(hate_probs[0])}
+  output["Offensive speech"] = {offensive_labels[i]: round(p.item(), 4) for i, p in enumerate(offensive_probs[0])}
+
+  return output
+
+# Create a gradio interface with a text input and a json output
+iface = gr.Interface(fn=get_scores, inputs="text", outputs="json")
+
+# Launch the interface
+iface.launch()