import streamlit as st import tensorflow as tf from transformers import AutoTokenizer, TFT5ForConditionalGeneration from datasets import load_dataset from pyarabic import araby import os # Load Tokenizer and Model model_name = "google/byt5-small" tokenizer = AutoTokenizer.from_pretrained(model_name) model = TFT5ForConditionalGeneration.from_pretrained(model_name) # Define custom classification head class BYT5ForClassification(tf.keras.Model): def __init__(self, base_model, num_labels): super(BYT5ForClassification, self).__init__() self.base_model = base_model self.classifier = tf.keras.layers.Dense(num_labels, activation='softmax') def call(self, inputs): # Use the encoder part of the BYT5 model only encoder_outputs = self.base_model.encoder( input_ids=inputs['input_ids'], attention_mask=inputs['attention_mask'] ) # Get the output for the [CLS] token (first token in the sequence) last_hidden_state = encoder_outputs[0] cls_token_output = last_hidden_state[:, 0, :] # Pass through classifier logits = self.classifier(cls_token_output) return logits # Initialize the model with the classification head num_labels = 14 model = BYT5ForClassification(base_model=model, num_labels=num_labels) # Label Mapping (from Hugging Face dataset or local file) label_file = os.path.join('./final_baits', 'labels.txt') with open(label_file, 'r') as f: label2name = [line.strip() for line in f.readlines()] # Preprocessing function for Arabic text def preprocess_text(text): text = araby.strip_tashkeel(text) # Strip Tashkeel excluded_chars = '!()*-ـ.:=o[]«»;؛,،~?؟\u200f\ufeffـ' # Define excluded characters cleaned_text = ''.join([char for char in text if char not in excluded_chars]) return cleaned_text.strip() # Tokenization for BYT5 def tokenize_input(texts): return tokenizer(texts, truncation=True, padding=True, max_length=512, return_tensors='tf') # Model output processing for classification def classify_with_byt5(sentence): sentence = preprocess_text(sentence) # Preprocess the sentence inputs = tokenize_input([sentence]) # Tokenize input # Get the model's output logits = model(inputs) # Apply softmax to logits to get probabilities probabilities = tf.nn.softmax(logits, axis=-1) # Convert probabilities to predicted label index predicted_label = tf.argmax(probabilities, axis=-1).numpy()[0] # Debugging output st.write(f"Logits: {logits}") st.write(f"Predicted Label Index: {predicted_label}") st.write(f"Number of labels: {num_labels}") # Ensure the predicted label is within the range of the label2name list if predicted_label < 0 or predicted_label >= num_labels: st.error("Predicted label is out of range.") return st.write(f"Predicted label: {label2name[predicted_label]}") # Example usage in Streamlit st.title("Arabic Poetry Classifier") text_input = st.text_input("Enter Arabic Poetry:") if text_input: classify_with_byt5(text_input)