dev

app.py

@@ -6,7 +6,6 @@ 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
@@ -20,14 +19,13 @@ lora_config = LoraConfig(
     lora_dropout=0.1,  # Dropout for LoRA layers
     bias="none"  # Bias handling
 )
-model = AutoModelForSeq2SeqLM.from_pretrained('google/t5-efficient-tiny-nh8', num_labels=2, force_download=True)
+model = AutoModelForSeq2SeqLM.from_pretrained('google/t5-efficient-tiny', num_labels=2, force_download=True)
 model = get_peft_model(model, lora_config)
 model.gradient_checkpointing_enable()   
 
 @spaces.GPU(duration=120)
 def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size, lr, grad):
     try:
-        torch.cuda.empty_cache()
         def compute_metrics(eval_pred):
             logits, labels = eval_pred
             predictions = np.argmax(logits, axis=1)
@@ -53,7 +51,7 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
             per_device_eval_batch_size=int(batch_size), 
             num_train_epochs=int(num_epochs),
             weight_decay=0.01,
-            gradient_accumulation_steps=int(grad),
+            #gradient_accumulation_steps=int(grad),
             max_grad_norm = 1.0, 
             load_best_model_at_end=True,
             metric_for_best_model="accuracy",
@@ -84,7 +82,6 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
                 train_dataset=tokenized_train_dataset,
                 eval_dataset=tokenized_test_dataset,
                 compute_metrics=compute_metrics,
-                #callbacks=[LoggingCallback()], 
             )            
         except:
             # Load the dataset