Upload app.py

app.py

@@ -4,17 +4,40 @@ from transformers import Trainer, TrainingArguments, AutoTokenizer, AutoModelFor
 from transformers import DataCollatorForSeq2Seq
 from datasets import load_dataset, concatenate_datasets, load_from_disk
 import traceback
-
+from sklearn.metrics import accuracy_score
+import numpy as np
+import torch
 
 import os
+from huggingface_hub import login
+from peft import get_peft_model, LoraConfig
 
+#os.environ['HF_HOME'] = '/data/.huggingface'
 
-@spaces.GPU
+@spaces.GPU(duration=120)
 def fine_tune_model(model_name, dataset_name, hub_id, api_key, num_epochs, batch_size, lr, grad):
     try:
-        #login(api_key.strip())
+        torch.cuda.empty_cache()
+        def compute_metrics(eval_pred):
+            logits, labels = eval_pred
+            predictions = np.argmax(logits, axis=1)
+            accuracy = accuracy_score(labels, predictions)
+            return {
+                'eval_accuracy': accuracy,
+                'eval_loss': eval_pred.loss,  # If you want to include loss as well
+            }        
+        login(api_key.strip())
+        lora_config = LoraConfig(
+            r=16,  # Rank of the low-rank adaptation
+            lora_alpha=32,  # Scaling factor
+            lora_dropout=0.1,  # Dropout for LoRA layers
+            bias="none"  # Bias handling
+        )        
+    
         # Load the model and tokenizer
-        model = AutoModelForSeq2SeqLM.from_pretrained('google/t5-efficient-tiny-nh8', num_labels=2)
+        model = AutoModelForSeq2SeqLM.from_pretrained(model_name.strip(), num_labels=2, force_download=True)
+        model.gradient_checkpointing_enable()
+        #model = get_peft_model(model, lora_config)
         
     
         # Set training arguments
@@ -42,48 +65,66 @@ def fine_tune_model(model_name, dataset_name, hub_id, api_key, num_epochs, batch
             save_total_limit=3, 
         )
         # Check if a checkpoint exists and load it
+        if os.path.exists(training_args.output_dir) and os.listdir(training_args.output_dir):
+            print("Loading model from checkpoint...")
+            model = AutoModelForSeq2SeqLM.from_pretrained(training_args.output_dir)        
+    
         max_length = 128
-        # Load the dataset
-        dataset = load_dataset(dataset_name.strip())
-        tokenizer = AutoTokenizer.from_pretrained('google/t5-efficient-tiny-nh8')
-        # Tokenize the dataset
-        def tokenize_function(examples):
+        try:
+            tokenized_train_dataset = load_from_disk(f'/data/{hub_id.strip()}_train_dataset')
+            tokenized_test_dataset = load_from_disk(f'/data/{hub_id.strip()}_test_dataset')
             
-            # Assuming 'text' is the input and 'target' is the expected output
-            model_inputs = tokenizer(
-                examples['text'], 
-                max_length=max_length,  # Set to None for dynamic padding
-                padding=True,     # Disable padding here, we will handle it later
-                truncation=True,
-            )
-        
-            # Setup the decoder input IDs (shifted right)
-            labels = tokenizer(
-                examples['target'], 
-                max_length=max_length,  # Set to None for dynamic padding
-                padding=True,     # Disable padding here, we will handle it later
-                truncation=True,
-                text_target=examples['target']  # Use text_target for target text
-            )
-        
-            # Add labels to the model inputs
-            model_inputs["labels"] = labels["input_ids"]
-        
+            # Create Trainer
+            trainer = Trainer(
+                model=model,
+                args=training_args,
+                train_dataset=tokenized_train_dataset,
+                eval_dataset=tokenized_test_dataset,
+                compute_metrics=compute_metrics,
+                #callbacks=[LoggingCallback()], 
+            )            
+        except:
+            # Load the dataset
+            dataset = load_dataset(dataset_name.strip())
+            tokenizer = AutoTokenizer.from_pretrained(model_name)
+            # Tokenize the dataset
+            def tokenize_function(examples):
+                
+                # Assuming 'text' is the input and 'target' is the expected output
+                model_inputs = tokenizer(
+                    examples['text'], 
+                    max_length=max_length,  # Set to None for dynamic padding
+                    padding=True,     # Disable padding here, we will handle it later
+                    truncation=True,
+                )
+            
+                # Setup the decoder input IDs (shifted right)
+                labels = tokenizer(
+                    examples['target'], 
+                    max_length=max_length,  # Set to None for dynamic padding
+                    padding=True,     # Disable padding here, we will handle it later
+                    truncation=True,
+                    text_target=examples['target']  # Use text_target for target text
+                )
+            
+                # Add labels to the model inputs
+                model_inputs["labels"] = labels["input_ids"]
+                return model_inputs
         
-        tokenized_datasets = dataset.map(tokenize_function, batched=True)
+            tokenized_datasets = dataset.map(tokenize_function, batched=True)
+            
+            tokenized_datasets['train'].save_to_disk(f'/data/{hub_id.strip()}_train_dataset')
+            tokenized_datasets['test'].save_to_disk(f'/data/{hub_id.strip()}_test_dataset')
         
-        tokenized_datasets['train'].save_to_disk(f'/data/{hub_id.strip()}_train_dataset')
-        tokenized_datasets['test'].save_to_disk(f'/data/{hub_id.strip()}_test_dataset')
-    
-        # Create Trainer
-        trainer = Trainer(
-            model=model,
-            args=training_args,
-            train_dataset=tokenized_datasets['train'],
-            eval_dataset=tokenized_datasets['test'],
-            compute_metrics=compute_metrics,
-            #callbacks=[LoggingCallback()], 
-        )            
+            # Create Trainer
+            trainer = Trainer(
+                model=model,
+                args=training_args,
+                train_dataset=tokenized_datasets['train'],
+                eval_dataset=tokenized_datasets['test'],
+                compute_metrics=compute_metrics,
+                #callbacks=[LoggingCallback()], 
+            )            
 
         # Fine-tune the model
         trainer.train()
@@ -103,6 +144,7 @@ def predict(text):
 '''
 # Create Gradio interface
 try:
+    model = AutoModelForSeq2SeqLM.from_pretrained('google/t5-efficient-tiny-nh8'.strip(), num_labels=2, force_download=True)
     iface = gr.Interface(
         fn=fine_tune_model,
         inputs=[