asd

app.py

@@ -1,9 +1,192 @@
+# import os
+# import json
+# import gradio as gr
+# import spaces
+# import torch
+# from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
+# from sentence_splitter import SentenceSplitter
+# from itertools import product
+
+# # Get the Hugging Face token from environment variable
+# hf_token = os.getenv('HF_TOKEN')
+
+# cuda_available = torch.cuda.is_available()
+# device = torch.device("cpu" if cuda_available else "cpu")
+# print(f"Using device: {device}")
+
+# # Initialize paraphraser model and tokenizer
+# paraphraser_model_name = "NoaiGPT/777"
+# paraphraser_tokenizer = AutoTokenizer.from_pretrained(paraphraser_model_name, use_auth_token=hf_token)
+# paraphraser_model = AutoModelForSeq2SeqLM.from_pretrained(paraphraser_model_name, use_auth_token=hf_token).to(device)
+
+# # Initialize classifier model and tokenizer
+# classifier_model_name = "andreas122001/roberta-mixed-detector"
+# classifier_tokenizer = AutoTokenizer.from_pretrained(classifier_model_name)
+# classifier_model = AutoModelForSequenceClassification.from_pretrained(classifier_model_name).to(device)
+
+# # Initialize sentence splitter
+# splitter = SentenceSplitter(language='en')
+
+# def classify_text(text):
+#     inputs = classifier_tokenizer(text, return_tensors="pt", truncation=True, max_length=512).to(device)
+#     with torch.no_grad():
+#         outputs = classifier_model(**inputs)
+#     probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
+#     predicted_class = torch.argmax(probabilities, dim=-1).item()
+#     main_label = classifier_model.config.id2label[predicted_class]
+#     main_score = probabilities[0][predicted_class].item()
+#     return main_label, main_score
+
+# # @spaces.GPU
+# def generate_paraphrases(text, setting, output_format):
+#     sentences = splitter.split(text)
+#     all_sentence_paraphrases = []
+    
+#     if setting == 1:
+#         num_return_sequences = 5
+#         repetition_penalty = 1.1
+#         no_repeat_ngram_size = 2
+#         temperature = 1.0
+#         max_length = 128
+#     elif setting == 2:
+#         num_return_sequences = 10
+#         repetition_penalty = 1.2
+#         no_repeat_ngram_size = 3
+#         temperature = 1.2
+#         max_length = 192
+#     elif setting == 3:
+#         num_return_sequences = 15
+#         repetition_penalty = 1.3
+#         no_repeat_ngram_size = 4
+#         temperature = 1.4
+#         max_length = 256
+#     elif setting == 4:
+#         num_return_sequences = 20
+#         repetition_penalty = 1.4
+#         no_repeat_ngram_size = 5
+#         temperature = 1.6
+#         max_length = 320
+#     else:
+#         num_return_sequences = 25
+#         repetition_penalty = 1.5
+#         no_repeat_ngram_size = 6
+#         temperature = 1.8
+#         max_length = 384
+    
+#     top_k = 50
+#     top_p = 0.95
+#     length_penalty = 1.0
+    
+#     formatted_output = "Original text:\n" + text + "\n\n"
+#     formatted_output += "Paraphrased versions:\n"
+    
+#     json_output = {
+#         "original_text": text,
+#         "paraphrased_versions": [],
+#         "combined_versions": [],
+#         "human_like_versions": []
+#     }
+    
+#     for i, sentence in enumerate(sentences):
+#         inputs = paraphraser_tokenizer(f'paraphraser: {sentence}', return_tensors="pt", padding="longest", truncation=True, max_length=max_length).to(device)
+        
+#         # Generate paraphrases using the specified parameters
+#         outputs = paraphraser_model.generate(
+#             inputs.input_ids,
+#             attention_mask=inputs.attention_mask,
+#             num_return_sequences=num_return_sequences,
+#             repetition_penalty=repetition_penalty,
+#             no_repeat_ngram_size=no_repeat_ngram_size,
+#             temperature=temperature,
+#             max_length=max_length,
+#             top_k=top_k,
+#             top_p=top_p,
+#             do_sample=True,
+#             early_stopping=False,
+#             length_penalty=length_penalty
+#         )
+        
+#         paraphrases = paraphraser_tokenizer.batch_decode(outputs, skip_special_tokens=True)
+        
+#         formatted_output += f"Original sentence {i+1}: {sentence}\n"
+#         for j, paraphrase in enumerate(paraphrases, 1):
+#             formatted_output += f"  Paraphrase {j}: {paraphrase}\n"
+        
+#         json_output["paraphrased_versions"].append({
+#             f"original_sentence_{i+1}": sentence,
+#             "paraphrases": paraphrases
+#         })
+        
+#         all_sentence_paraphrases.append(paraphrases)
+#         formatted_output += "\n"
+    
+#     all_combinations = list(product(*all_sentence_paraphrases))
+    
+#     formatted_output += "\nCombined paraphrased versions:\n"
+#     combined_versions = []
+#     for i, combination in enumerate(all_combinations[:50], 1):  # Limit to 50 combinations
+#         combined_paraphrase = " ".join(combination)
+#         combined_versions.append(combined_paraphrase)
+    
+#     json_output["combined_versions"] = combined_versions
+    
+#     # Classify combined versions
+#     human_versions = []
+#     for i, version in enumerate(combined_versions, 1):
+#         label, score = classify_text(version)
+#         formatted_output += f"Version {i}:\n{version}\n"
+#         formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
+#         if label == "human-produced" or (label == "machine-generated" and score < 0.98):
+#             human_versions.append((version, label, score))
+    
+#     formatted_output += "\nHuman-like or Less Confident Machine-generated versions:\n"
+#     for i, (version, label, score) in enumerate(human_versions, 1):
+#         formatted_output += f"Version {i}:\n{version}\n"
+#         formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
+    
+#     json_output["human_like_versions"] = [
+#         {"version": version, "label": label, "confidence_score": score}
+#         for version, label, score in human_versions
+#     ]
+    
+#     # If no human-like versions, include the top 5 least confident machine-generated versions
+#     if not human_versions:
+#         human_versions = sorted([(v, l, s) for v, l, s in zip(combined_versions, [classify_text(v)[0] for v in combined_versions], [classify_text(v)[1] for v in combined_versions])], key=lambda x: x[2])[:5]
+#         formatted_output += "\nNo human-like versions found. Showing top 5 least confident machine-generated versions:\n"
+#         for i, (version, label, score) in enumerate(human_versions, 1):
+#             formatted_output += f"Version {i}:\n{version}\n"
+#             formatted_output += f"Classification: {label} (confidence: {score:.2%})\n\n"
+    
+#     if output_format == "text":
+#         return formatted_output, "\n\n".join([v[0] for v in human_versions])
+#     else:
+#         return json.dumps(json_output, indent=2), "\n\n".join([v[0] for v in human_versions])
+
+# # Define the Gradio interface
+# iface = gr.Interface(
+#     fn=generate_paraphrases,
+#     inputs=[
+#         gr.Textbox(lines=5, label="Input Text"),
+#         gr.Slider(minimum=1, maximum=5, step=1, label="Readability to Human-like Setting"),
+#         gr.Radio(["text", "json"], label="Output Format")
+#     ],
+#     outputs=[
+#         gr.Textbox(lines=20, label="Detailed Paraphrases and Classifications"),
+#         gr.Textbox(lines=10, label="Human-like or Less Confident Machine-generated Paraphrases")
+#     ],
+#     title="Advanced Diverse Paraphraser with Human-like Filter",
+#     description="Enter a text, select a setting from readable to human-like, and choose the output format to generate diverse paraphrased versions. Combined versions are classified, and those detected as human-produced or less confidently machine-generated are presented in the final output."
+# )
+
+# # Launch the interface
+# iface.launch()
+
 import os
 import json
 import gradio as gr
 import spaces
 import torch
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
+from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
 from sentence_splitter import SentenceSplitter
 from itertools import product
 
@@ -11,13 +194,14 @@ from itertools import product
 hf_token = os.getenv('HF_TOKEN')
 
 cuda_available = torch.cuda.is_available()
-device = torch.device("cpu" if cuda_available else "cpu")
+device = torch.device("cuda" if cuda_available else "cpu")
 print(f"Using device: {device}")
 
 # Initialize paraphraser model and tokenizer
-paraphraser_model_name = "NoaiGPT/777"
-paraphraser_tokenizer = AutoTokenizer.from_pretrained(paraphraser_model_name, use_auth_token=hf_token)
-paraphraser_model = AutoModelForSeq2SeqLM.from_pretrained(paraphraser_model_name, use_auth_token=hf_token).to(device)
+paraphraser_model_name = "sharad/ParaphraseGPT"
+paraphraser_tokenizer = AutoTokenizer.from_pretrained("humarin/chatgpt_paraphraser_on_T5_base")
+paraphraser_model = AutoModelForSeq2SeqLM.from_pretrained(paraphraser_model_name).to(device)
+paraphrase_pipeline = pipeline("text2text-generation", model=paraphraser_model, tokenizer=paraphraser_tokenizer)
 
 # Initialize classifier model and tokenizer
 classifier_model_name = "andreas122001/roberta-mixed-detector"
@@ -37,7 +221,7 @@ def classify_text(text):
     main_score = probabilities[0][predicted_class].item()
     return main_label, main_score
 
-# @spaces.GPU
+@spaces.GPU
 def generate_paraphrases(text, setting, output_format):
     sentences = splitter.split(text)
     all_sentence_paraphrases = []
@@ -88,36 +272,29 @@ def generate_paraphrases(text, setting, output_format):
     }
     
     for i, sentence in enumerate(sentences):
-        inputs = paraphraser_tokenizer(f'paraphraser: {sentence}', return_tensors="pt", padding="longest", truncation=True, max_length=max_length).to(device)
-        
-        # Generate paraphrases using the specified parameters
-        outputs = paraphraser_model.generate(
-            inputs.input_ids,
-            attention_mask=inputs.attention_mask,
+        paraphrases = paraphrase_pipeline(
+            sentence,
+            num_beams=3,
+            num_beam_groups=3,
             num_return_sequences=num_return_sequences,
-            repetition_penalty=repetition_penalty,
+            diversity_penalty=2.0,
             no_repeat_ngram_size=no_repeat_ngram_size,
-            temperature=temperature,
-            max_length=max_length,
-            top_k=top_k,
-            top_p=top_p,
-            do_sample=True,
-            early_stopping=False,
-            length_penalty=length_penalty
+            repetition_penalty=repetition_penalty,
+            max_length=max_length
         )
         
-        paraphrases = paraphraser_tokenizer.batch_decode(outputs, skip_special_tokens=True)
+        paraphrases_texts = [p['generated_text'] for p in paraphrases]
         
         formatted_output += f"Original sentence {i+1}: {sentence}\n"
-        for j, paraphrase in enumerate(paraphrases, 1):
+        for j, paraphrase in enumerate(paraphrases_texts, 1):
             formatted_output += f"  Paraphrase {j}: {paraphrase}\n"
         
         json_output["paraphrased_versions"].append({
             f"original_sentence_{i+1}": sentence,
-            "paraphrases": paraphrases
+            "paraphrases": paraphrases_texts
         })
         
-        all_sentence_paraphrases.append(paraphrases)
+        all_sentence_paraphrases.append(paraphrases_texts)
         formatted_output += "\n"
     
     all_combinations = list(product(*all_sentence_paraphrases))