Update app_flash1.py

app_flash1.py

@@ -12,17 +12,17 @@ from huggingface_hub import Repository
 from typing import Tuple
 
 # ============================================================
-# 🖥 CPU device setup
+# 🖥 Device setup (CPU-only)
 # ============================================================
 device = torch.device("cpu")
 torch.set_num_threads(4)
 print(f"🔧 Using device: {device} (CPU-only mode)")
 
 # ============================================================
-# 1️⃣ FlashPack MLP model (CPU-friendly)
+# 1️⃣ FlashPack model
 # ============================================================
 class GemmaTrainer(nn.Module, FlashPackMixin):
-    def __init__(self, input_dim: int, hidden_dim: int = 512, output_dim: int = 768):
+    def __init__(self, input_dim: int, hidden_dim: int = 1024, output_dim: int = 1536):
         super().__init__()
         self.fc1 = nn.Linear(input_dim, hidden_dim)
         self.relu = nn.ReLU()
@@ -38,39 +38,36 @@ class GemmaTrainer(nn.Module, FlashPackMixin):
         return x
 
 # ============================================================
-# 2️⃣ Lazy-loading GPT-2 encoder
+# 2️⃣ Encoder (mean + max pooling)
 # ============================================================
-_embed_model = None
-_tokenizer = None
+def build_encoder(model_name="gpt2", max_length: int = 128):
+    print(f"📦 Loading tokenizer and model for {model_name}...")
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
 
-def get_encoder(model_name="gpt2", max_length=64):
-    global _embed_model, _tokenizer
-    if _embed_model is None or _tokenizer is None:
-        print("⚡ Loading GPT-2 encoder model...")
-        _tokenizer = AutoTokenizer.from_pretrained(model_name)
-        if _tokenizer.pad_token is None:
-            _tokenizer.pad_token = _tokenizer.eos_token
-        _embed_model = AutoModel.from_pretrained(model_name).to(device)
-        _embed_model.eval()
-    return _tokenizer, _embed_model
+    embed_model = AutoModel.from_pretrained(model_name).to(device)
+    embed_model.eval()
+    print(f"✅ Encoder ready: {model_name}")
 
-@torch.no_grad()
-def encode_prompt(prompt: str) -> torch.Tensor:
-    tokenizer, embed_model = get_encoder()
-    inputs = tokenizer(prompt, return_tensors="pt", truncation=True,
-                       padding="max_length", max_length=64).to(device)
-    last_hidden = embed_model(**inputs).last_hidden_state
-    mean_pool = last_hidden.mean(dim=1)
-    max_pool, _ = last_hidden.max(dim=1)
-    return torch.cat([mean_pool, max_pool], dim=1).cpu()
+    @torch.no_grad()
+    def encode(prompt: str) -> torch.Tensor:
+        inputs = tokenizer(prompt, return_tensors="pt", truncation=True,
+                           padding="max_length", max_length=max_length).to(device)
+        last_hidden = embed_model(**inputs).last_hidden_state
+        mean_pool = last_hidden.mean(dim=1)
+        max_pool, _ = last_hidden.max(dim=1)
+        return torch.cat([mean_pool, max_pool], dim=1).cpu()  # double dimension
+
+    return tokenizer, embed_model, encode
 
 # ============================================================
-# 3️⃣ Push FlashPack model to Hugging Face Hub
+# 3️⃣ Push FlashPack model to HF
 # ============================================================
 def push_flashpack_model_to_hf(model, hf_repo: str):
     logs = []
     with tempfile.TemporaryDirectory() as tmp_dir:
-        logs.append(f"📂 Using temp dir: {tmp_dir}")
+        logs.append(f"📂 Temporary directory: {tmp_dir}")
         repo = Repository(local_dir=tmp_dir, clone_from=hf_repo, use_auth_token=True)
         pack_path = os.path.join(tmp_dir, "model.flashpack")
         model.save_flashpack(pack_path, target_dtype=torch.float32)
@@ -86,52 +83,60 @@ def push_flashpack_model_to_hf(model, hf_repo: str):
 # ============================================================
 def train_flashpack_model(
     dataset_name: str = "rahul7star/prompt-enhancer-dataset",
-    max_encode: int = 500,
-    hidden_dim: int = 512,
-    push_to_hub: bool = True,
+    max_encode: int = 1000,
+    hidden_dim: int = 1024,
     hf_repo: str = "rahul7star/FlashPack",
-    early_stop_threshold: float = 0.001
+    push_to_hub: bool = True
 ) -> Tuple[GemmaTrainer, object, object, object, torch.Tensor]:
-    
+
     print("📦 Loading dataset...")
     dataset = load_dataset(dataset_name, split="train")
-    dataset = dataset.select(range(min(max_encode, len(dataset))))
-    n_train = int(0.8 * len(dataset))
-    n_test = len(dataset) - n_train
-    train_dataset = dataset.select(range(n_train))
-    test_dataset = dataset.select(range(n_train, len(dataset)))
-    print(f"⚡ Train: {n_train}, Test: {n_test}")
-
-    # Encode prompts lazily
-    def batch_encode(ds):
+    limit = min(max_encode, len(dataset))
+    dataset = dataset.select(range(limit))
+    print(f"⚡ Using {len(dataset)} prompts for training")
+
+    # Split train/test
+    train_size = int(0.9 * len(dataset))
+    test_size = len(dataset) - train_size
+    dataset_train = dataset.select(range(train_size))
+    dataset_test = dataset.select(range(train_size, len(dataset)))
+    print(f"🧪 Train/Test split: {len(dataset_train)} / {len(dataset_test)}")
+
+    tokenizer, embed_model, encode_fn = build_encoder("gpt2", max_length=128)
+
+    # Encode embeddings
+    def encode_dataset(ds):
         short_list, long_list = [], []
         for i, item in enumerate(ds):
-            short_list.append(encode_prompt(item["short_prompt"]))
-            long_list.append(encode_prompt(item["long_prompt"]))
-            if (i+1) % 20 == 0 or (i+1) == len(ds):
+            short_list.append(encode_fn(item["short_prompt"]))
+            long_list.append(encode_fn(item["long_prompt"]))
+            if (i+1) % 50 == 0 or (i+1) == len(ds):
                 print(f"  → Encoded {i+1}/{len(ds)} prompts")
                 gc.collect()
         return torch.vstack(short_list), torch.vstack(long_list)
 
-    short_train, long_train = batch_encode(train_dataset)
-    short_test, long_test = batch_encode(test_dataset)
-    print(f"✅ Embeddings shapes: short_train={short_train.shape}, long_train={long_train.shape}")
+    short_train, long_train = encode_dataset(dataset_train)
+    short_test, long_test = encode_dataset(dataset_test)
+    print(f"✅ Embeddings shapes: train {short_train.shape}/{long_train.shape}, test {short_test.shape}/{long_test.shape}")
 
+    # Build model
     input_dim = short_train.shape[1]
     output_dim = long_train.shape[1]
     model = GemmaTrainer(input_dim=input_dim, hidden_dim=hidden_dim, output_dim=output_dim).to(device)
 
+    # Loss and optimizer
     criterion = nn.CosineSimilarity(dim=1)
     optimizer = optim.Adam(model.parameters(), lr=1e-3)
     max_epochs = 50
-    batch_size = 16
+    batch_size = 32
+    n = short_train.shape[0]
 
+    # Training loop
     print("🚀 Training model...")
-    n = short_train.shape[0]
     for epoch in range(max_epochs):
         model.train()
-        perm = torch.randperm(n)
         epoch_loss = 0.0
+        perm = torch.randperm(n)
         for start in range(0, n, batch_size):
             idx = perm[start:start+batch_size]
             inputs = short_train[idx].to(device)
@@ -144,68 +149,97 @@ def train_flashpack_model(
             epoch_loss += loss.item() * inputs.size(0)
         epoch_loss /= n
 
-        # Evaluate on test set
+        # Evaluate on test
         model.eval()
         with torch.no_grad():
-            outputs_test = model(short_test.to(device))
-            test_loss = 1 - criterion(outputs_test, long_test.to(device)).mean().item()
-
-        print(f"Epoch {epoch+1}/{max_epochs} | Train Loss={epoch_loss:.6f} | Test Loss={test_loss:.6f}")
+            test_outputs = model(short_test.to(device))
+            test_loss = 1 - criterion(test_outputs, long_test.to(device)).mean()
+        if epoch % 5 == 0 or epoch == max_epochs-1:
+            print(f"Epoch {epoch+1}/{max_epochs} — Train Loss: {epoch_loss:.6f}, Test Loss: {test_loss:.6f}")
 
-        # Early stop: very low test loss means model is good
-        if test_loss < early_stop_threshold:
-            print("🎯 Early stop: test loss below threshold. Model is ready!")
+        # Auto stop if test_loss is very small
+        if test_loss < 0.01:
+            print("🎯 Test loss very low — early stopping!")
             break
 
-    if push_to_hub:
+    print("✅ Training finished!")
+
+    # Push to HF if training good
+    if push_to_hub and test_loss < 0.05:
         logs = push_flashpack_model_to_hf(model, hf_repo)
         for log in logs:
             print(log)
+    else:
+        print("⚠️ Model not pushed — test loss not low enough.")
 
-    return model, dataset, None, None, long_train  # embed_model and tokenizer lazy-loaded
+    # Return
+    return model, dataset, embed_model, tokenizer, long_train
 
 # ============================================================
-# 5️⃣ Lazy load or train
+# 5️⃣ Lazy-load model
 # ============================================================
 def get_flashpack_model(hf_repo="rahul7star/FlashPack"):
     try:
-        print(f"🔁 Attempting to load FlashPack model from {hf_repo}")
+        print(f"🔁 Loading FlashPack model from {hf_repo}...")
         model = GemmaTrainer.from_flashpack(hf_repo)
         model.eval()
-        print("✅ Loaded model from HF")
-        return model
+        tokenizer, embed_model, encode_fn = build_encoder("gpt2", max_length=128)
+        return model, tokenizer, embed_model
     except Exception as e:
         print(f"⚠️ Load failed: {e}")
-        print("⏬ Training new FlashPack model locally...")
-        model, dataset, _, _, long_embeddings = train_flashpack_model()
-        return model, dataset, long_embeddings
+        print("⏬ Training new model locally...")
+        model, dataset, embed_model, tokenizer, long_embeddings = train_flashpack_model()
+        return model, tokenizer, embed_model, dataset, long_embeddings
+
+# ============================================================
+# 6️⃣ Initialize model
+# ============================================================
+model, tokenizer, embed_model, dataset, long_embeddings = get_flashpack_model()
 
 # ============================================================
-# 6️⃣ Inference helpers
+# 7️⃣ Inference helpers (closure for Gradio)
 # ============================================================
 @torch.no_grad()
-def enhance_prompt(user_prompt: str, chat_history, model, long_embeddings, dataset):
-    chat_history = chat_history or []
-    short_emb = encode_prompt(user_prompt)
-    mapped = model(short_emb.to(device)).cpu()
-    sims = (long_embeddings @ mapped.t()).squeeze(1)
-    long_norms = long_embeddings.norm(dim=1)
-    mapped_norm = mapped.norm()
-    sims = sims / (long_norms * (mapped_norm + 1e-12))
-    best_idx = int(sims.argmax().item())
-    enhanced_prompt = dataset[best_idx]["long_prompt"]
+def encode_prompt(prompt: str) -> torch.Tensor:
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True,
+                       padding="max_length", max_length=128).to(device)
+    last_hidden = embed_model(**inputs).last_hidden_state
+    mean_pool = last_hidden.mean(dim=1)
+    max_pool, _ = last_hidden.max(dim=1)
+    return torch.cat([mean_pool, max_pool], dim=1).cpu()
+
+def make_enhance_fn(model, long_embeddings, dataset):
+    @torch.no_grad()
+    def fn(user_prompt, chat_history):
+        chat_history = chat_history or []
+        short_emb = encode_prompt(user_prompt)
+        mapped = model(short_emb.to(device)).cpu()
+        sims = (long_embeddings @ mapped.t()).squeeze(1)
+        long_norms = long_embeddings.norm(dim=1)
+        mapped_norm = mapped.norm()
+        sims = sims / (long_norms * (mapped_norm + 1e-12))
+        best_idx = int(sims.argmax().item())
+        enhanced_prompt = dataset[best_idx]["long_prompt"]
+
+        chat_history.append({"role": "user", "content": user_prompt})
+        chat_history.append({"role": "assistant", "content": enhanced_prompt})
+        return chat_history
+    return fn
 
-    chat_history.append({"role": "user", "content": user_prompt})
-    chat_history.append({"role": "assistant", "content": enhanced_prompt})
-    return chat_history
+enhance_fn = make_enhance_fn(model, long_embeddings, dataset)
 
 # ============================================================
-# 7️⃣ Launch Gradio app
+# 8️⃣ Gradio UI
 # ============================================================
-model, dataset, long_embeddings = get_flashpack_model()
-
 with gr.Blocks(title="Prompt Enhancer – FlashPack (CPU)", theme=gr.themes.Soft()) as demo:
-    gr.Markdown("# ✨ Prompt Enhancer (FlashPack mapper)\nEnter a short prompt, and it will expand it.")
+    gr.Markdown(
+        """
+        # ✨ Prompt Enhancer (FlashPack mapper)
+        Enter a short prompt, and the model will **expand it with details and creative context**.
+        (CPU-only mode.)
+        """
+    )
+
     with gr.Row():
         chatbot = gr.Chatbot(height=400, label="Enhanced Prompts", type="messages")
         with gr.Column(scale=1):
@@ -213,9 +247,12 @@ with gr.Blocks(title="Prompt Enhancer – FlashPack (CPU)", theme=gr.themes.Soft
             send_btn = gr.Button("🚀 Enhance Prompt", variant="primary")
             clear_btn = gr.Button("🧹 Clear Chat")
 
-    send_btn.click(enhance_prompt, [user_prompt, chatbot, model, long_embeddings, dataset], chatbot)
-    user_prompt.submit(enhance_prompt, [user_prompt, chatbot, model, long_embeddings, dataset], chatbot)
+    send_btn.click(enhance_fn, [user_prompt, chatbot], chatbot)
+    user_prompt.submit(enhance_fn, [user_prompt, chatbot], chatbot)
     clear_btn.click(lambda: [], None, chatbot)
 
+# ============================================================
+# 9️⃣ Launch
+# ============================================================
 if __name__ == "__main__":
     demo.launch(show_error=True)