Upload nlpapp_4900.py

nlpapp_4900.py

@@ -0,0 +1,216 @@
+# -*- coding: utf-8 -*-
+"""nlpapp_4900.ipynb
+
+Automatically generated by Colaboratory.
+
+Original file is located at
+    https://colab.research.google.com/drive/1QE4TN1ipB5SwIVy2a7D6_bTPujHMBoDu
+"""
+
+# Import required modules for auto-reconnecting to the Colab runtime
+# Set up auto-reconnect
+import IPython
+from google.colab import output
+
+# Configure the JS code to automatically reconnect
+def restart_runtime():
+  output.clear()
+  print('Restarting runtime...')
+  display(IPython.display.Javascript('''
+    const outputArea = this;
+    const kernel = IPython.notebook.kernel;
+    const command = 'notebook_utils.restart_runtime()';
+    kernel.execute(command).then(() => {
+      outputArea.clear_output();
+      console.log('Runtime restarted. Running all cells...');
+      IPython.notebook.execute_all_cells();
+    });
+  '''))
+
+# Set the number of minutes of inactivity before auto-reconnecting
+
+minutes = 30 #@param {type: "slider", min: 1, max: 180, step: 1} # define a slider widget to set the time before auto-reconnecting
+
+# Register the auto-reconnect function
+import time
+def auto_reconnect():
+  while True:
+    print(f'Auto-reconnect in {minutes} minute(s)...')
+    time.sleep(minutes * 60 - 5) # Subtract 5 seconds to give the reconnect JS code time to run
+    restart_runtime()
+
+# Start the auto-reconnect loop in the background
+import threading
+auto_reconnect_thread = threading.Thread(target=auto_reconnect)
+auto_reconnect_thread.start()
+
+!pip install transformers torch gradio
+
+# Download and extract the Yelp review dataset: download and extract the dataset using wget and tar
+!wget https://s3.amazonaws.com/fast-ai-nlp/yelp_review_polarity_csv.tgz
+!tar -xvzf yelp_review_polarity_csv.tgz
+
+import pandas as pd
+
+train_df = pd.read_csv('yelp_review_polarity_csv/train.csv', header=None, names=['label', 'text'])
+test_df = pd.read_csv('yelp_review_polarity_csv/test.csv', header=None, names=['label', 'text'])
+
+import re
+
+# Define a function for removing HTML tags, URLs, and extra spaces, and converting the text to lowercase
+def preprocess_text(text):
+    # remove HTML tags
+    text = re.sub('<[^<]+?>', '', text)
+    # remove URLs
+    text = re.sub(r'http\S+', '', text)
+    # remove extra spaces
+    text = re.sub(r' +', ' ', text)
+    return text.strip().lower()
+
+# Apply the preprocessing function to the 'text' column of the training and test datasets    
+train_df['text'] = train_df['text'].apply(preprocess_text)
+test_df['text'] = test_df['text'].apply(preprocess_text)
+
+import torch
+import transformers
+import gradio as gr
+from transformers import TextDataset, DataCollatorForLanguageModeling, Trainer, TrainingArguments
+
+# Instantiate the tokenizer and the GPT-2 language model from the transformers library, and set the device to CUDA if available, otherwise to CPU
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+tokenizer = transformers.GPT2Tokenizer.from_pretrained('gpt2')
+model = transformers.GPT2LMHeadModel.from_pretrained('gpt2').to(device)
+
+# Fine-tune the model on Yelp review dataset
+train_dataset = TextDataset(
+    tokenizer=tokenizer,
+    file_path='yelp_review_polarity_csv/train.csv',
+    block_size=128,
+)
+test_dataset = TextDataset(
+    tokenizer=tokenizer,
+    file_path='yelp_review_polarity_csv/test.csv',
+    block_size=128,
+)
+data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
+
+# Define the training arguments, instantiate the Trainer class from the transformers library, and train the model
+training_args = TrainingArguments(
+    output_dir='./results',
+    evaluation_strategy='steps',
+    eval_steps=1000,
+    save_steps=1000,
+    save_total_limit=5,
+    logging_steps=100,
+    logging_dir='./logs',
+    num_train_epochs=3,
+    per_device_train_batch_size=16,
+    per_device_eval_batch_size=32,
+    learning_rate=1e-4,
+    weight_decay=0.01,
+    gradient_accumulation_steps=2,
+    push_to_hub=False,
+    max_steps=10000,  # set a fixed number of training steps
+    # save model checkpoints at specified intervals
+    save_strategy="steps",
+)
+
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=train_dataset,
+    data_collator=data_collator,
+    eval_dataset=test_dataset,
+)
+trainer.train()
+trainer.save_model('./gpt2_yelp_review')
+
+# Evaluate the model on the test dataset and print the perplexity score
+eval_results = trainer.evaluate(eval_dataset=test_dataset)
+print(f"Perplexity: {eval_results['eval_loss']}")
+
+import pandas as pd
+import gradio as gr
+import re
+import torch
+import transformers
+
+# Define a function for generating text based on a prompt using the fine-tuned GPT-2 model and the tokenizer
+def generate_text(prompt, length=100, theme=None, **kwargs):
+    model = transformers.GPT2LMHeadModel.from_pretrained('./gpt2_yelp_review').to(device)
+    tokenizer = transformers.GPT2Tokenizer.from_pretrained('gpt2')
+
+    # If a theme is specified, add it to the prompt as a prefix for a special token
+    if theme:
+        prompt = ' <{}> '.format(theme.strip()) + prompt.strip()
+
+    input_ids = tokenizer.encode(prompt, return_tensors='pt').to(device)
+    attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=device)
+    pad_token_id = tokenizer.eos_token_id
+
+    # Set the max length of the generated text based on the input parameter
+    max_length = length if length > 0 else 100
+
+    sample_outputs = model.generate(
+        input_ids,
+        attention_mask=attention_mask,
+        pad_token_id=pad_token_id,
+        do_sample=True,
+        max_length=max_length,
+        top_k=50,
+        top_p=0.95,
+        temperature=0.8,
+        num_return_sequences=1,
+        no_repeat_ngram_size=2,
+        repetition_penalty=1.5,
+    )
+    generated_text = tokenizer.decode(sample_outputs[0], skip_special_tokens=True)
+
+    # Post preprocessing of the generated text
+    
+    # Remove any leading and trailing quotation marks
+    generated_text = generated_text.strip('"')
+
+    # Remove leading and trailing whitespace
+    generated_text = generated_text.strip()
+
+    # Find the special token in the generated text and remove it
+    match = re.search(r'<([^>]+)>', generated_text)
+    if match:
+        generated_text = generated_text[:match.start()] + generated_text[match.end():]
+
+    # Remove any leading numeric characters and quotation marks
+    generated_text = re.sub(r'^\d+', '', generated_text)
+    generated_text = re.sub(r'^"', '', generated_text)
+
+    # Remove any newline characters from the generated text
+    generated_text = generated_text.replace('\n', '')
+
+    # Remove any other unwanted special characters
+    generated_text = re.sub(r'[^\w\s]+', '', generated_text)
+    
+    return generated_text.strip().capitalize()
+
+# Define a Gradio interface for the generate_text function, allowing users to input a prompt and generate text based on it
+iface = gr.Interface(
+    fn=generate_text,
+    inputs=['text', gr.inputs.Slider(minimum=10, maximum=100, default=50, label='Length of text'),
+            gr.inputs.Textbox(default='Food', label='Theme')],
+    outputs=[gr.outputs.Textbox(label='Generated Text')],
+    title='Yelp Review Generator',
+    description='Generate a Yelp review based on a prompt, length of text, and theme.',
+    examples=[
+        ['I had a great experience at this restaurant.', 50, 'Service'],
+        ['The service was terrible and the food was cold.', 50, 'Atmosphere'],
+        ['The food was delicious but the service was slow.', 50, 'Food'],
+        ['The ambiance was amazing and the staff was friendly.', 75, 'Service'],
+        ['The waitstaff was knowledgeable and attentive, but the noise level was a bit high.', 75, 'Atmosphere'],
+        ['The menu had a good variety of options, but the portion sizes were a bit small for the price.', 75, 'Food']
+    ],
+    allow_flagging="manual",
+    flagging_options=[("🙌", "positive"), ("😞", "negative")],
+)
+
+iface.launch(debug=False, share=True)
+