sweep9.py

import transformers
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
import numpy as np
import matplotlib.pyplot as plt
import json
from tqdm.auto import tqdm
import random
from scipy.signal import savgol_filter
import wandb
from transformers import AutoModelForCausalLM, AutoTokenizer
import os
from itertools import product
import pandas as pd
import multiprocessing as mp
from functools import partial
import logging

logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

class Config:
    def __init__(self):
        self.model_name = "Qwen/Qwen2-0.5B"
        self.data_dir = "dataset_chunks"
        self.max_length = 1024
        self.batch_size = 32
        self.num_seeds = 10000
        self.num_lr_steps = 10000
        self.min_lr = 1e-8
        self.max_lr = 10
        self.hidden_dim_ratio = 0.5
        self.dropout = 0.1
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.num_workers = mp.cpu_count()

class ImprovedAutoencoder(nn.Module):
    def __init__(self, input_dim, hidden_dim, num_layers, dropout):
        super().__init__()
        self.encoder = nn.ModuleList([
            nn.Linear(input_dim if i == 0 else hidden_dim, hidden_dim, dtype=torch.bfloat16)
            for i in range(num_layers)
        ])
        self.decoder = nn.ModuleList([
            nn.Linear(hidden_dim, hidden_dim if i < num_layers - 1 else input_dim, dtype=torch.bfloat16)
            for i in range(num_layers)
        ])
        self.layer_norms = nn.ModuleList([
            nn.LayerNorm(hidden_dim, dtype=torch.bfloat16)
            for _ in range(num_layers * 2 - 1)
        ])
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        for enc, norm in zip(self.encoder, self.layer_norms[:len(self.encoder)]):
            x = F.relu(norm(enc(x)))
            x = self.dropout(x)
        for dec, norm in zip(self.decoder[:-1], self.layer_norms[len(self.encoder):]):
            x = F.relu(norm(dec(x)))
            x = self.dropout(x)
        x = self.decoder[-1](x)
        return x

class TokenizedDataset(torch.utils.data.Dataset):
    def __init__(self, file_paths):
        self.data = []
        for file_path in tqdm(file_paths, desc="Loading data chunks"):
            chunk_data = torch.load(file_path)
            logger.info(f"Loaded data from {file_path}")
            logger.info(f"Type of loaded data: {type(chunk_data)}")

            if isinstance(chunk_data, dict):  # Handle dictionary format (if present)
                logger.info(f"Keys in the dictionary: {chunk_data.keys()}")
                logger.info(f"Shape of input_ids: {chunk_data['input_ids'].shape}")
                self.data.append(chunk_data)
            elif isinstance(chunk_data, transformers.tokenization_utils_base.BatchEncoding): 
                logger.info(f"Keys in the BatchEncoding: {chunk_data.keys()}")
                logger.info(f"Shape of input_ids: {chunk_data['input_ids'].shape}")
                self.data.append(chunk_data) # Handle BatchEncoding format
            else:
                logger.warning(f"Unexpected data type: {type(chunk_data)}")

        logger.info(f"Loaded {len(self.data)} chunks of data")

    def __len__(self):
        return sum(len(chunk['input_ids']) for chunk in self.data)

    def __getitem__(self, idx):
        for chunk in self.data:
            if idx < len(chunk['input_ids']):
                return {k: v[idx] for k, v in chunk.items()}
            idx -= len(chunk['input_ids'])
        raise IndexError("Index out of range")

def set_seed(seed):
    random.seed(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)

def load_data(config):
    logger.info(f"Looking for data in directory: {config.data_dir}")
    chunk_files = [f for f in os.listdir(config.data_dir) if f.endswith('_tokenized.pt')]
    logger.info(f"Found {len(chunk_files)} chunk files: {chunk_files}")
    
    if not chunk_files:
        raise ValueError(f"No tokenized data files found in {config.data_dir}")
    
    chunk_files.sort(key=lambda x: int(x.split('_')[1]))
    chunk_files = [os.path.join(config.data_dir, f) for f in chunk_files]
    
    dataset = TokenizedDataset(chunk_files[:1])  # Load only the first chunk for now
    logger.info(f"Created dataset with {len(dataset)} samples")
    
    return DataLoader(dataset, batch_size=config.batch_size, shuffle=True, num_workers=config.num_workers)

def extract_hidden_states(batch, model):
    with torch.no_grad():
        outputs = model(**batch, output_hidden_states=True)
    return outputs.hidden_states[0], outputs.hidden_states[-1]

class KLDivergenceLoss(nn.Module):
    def forward(self, pred, target):
        pred = F.log_softmax(pred, dim=-1)
        target = F.softmax(target, dim=-1)
        return F.kl_div(pred, target, reduction='batchmean', log_target=False)

def lr_finder(model, autoencoder, loss_fn, optimizer, train_loader, config):
    model.eval()
    autoencoder.train()
    log_lrs, losses = [], []
    best_loss, best_lr = float('inf'), None

    pbar = tqdm(total=config.num_lr_steps, desc="LR Finder")
    for batch_idx, batch in enumerate(train_loader):
        if batch_idx >= config.num_lr_steps:
            break

        lr = config.min_lr * (config.max_lr / config.min_lr) ** (batch_idx / (config.num_lr_steps - 1))
        optimizer.param_groups[0]['lr'] = lr

        batch = {k: v.to(config.device) for k, v in batch.items()}
        first_states, last_states = extract_hidden_states(batch, model)

        optimizer.zero_grad()
        reconstructed = autoencoder(first_states)
        loss = loss_fn(reconstructed, last_states)
        loss.backward()
        optimizer.step()

        if loss < best_loss:
            best_loss = loss.item()
            best_lr = lr

        log_lrs.append(lr)
        losses.append(loss.item())

        pbar.update(1)
        pbar.set_postfix({"Loss": f"{loss.item():.4f}", "LR": f"{lr:.2e}"})

    pbar.close()
    return log_lrs, losses, best_lr, best_loss

def run_experiment(config, model, train_loader, num_layers, seed):
    set_seed(seed)
    input_dim = model.config.hidden_size
    hidden_dim = int(input_dim * config.hidden_dim_ratio)
    autoencoder = ImprovedAutoencoder(input_dim, hidden_dim, num_layers, config.dropout).to(config.device)

    loss_fn = KLDivergenceLoss()
    optimizer = optim.AdamW(autoencoder.parameters(), lr=config.min_lr)

    log_lrs, losses = [], []
    best_loss, best_lr = float('inf'), None

    pbar = tqdm(total=config.num_lr_steps, desc=f"LR Finder (Layers: {num_layers}, Seed: {seed})")
    for batch_idx, batch in enumerate(train_loader):
        if batch_idx >= config.num_lr_steps:
            break

        lr = config.min_lr * (config.max_lr / config.min_lr) ** (batch_idx / (config.num_lr_steps - 1))
        optimizer.param_groups[0]['lr'] = lr

        batch = {k: v.to(config.device) for k, v in batch.items()}
        first_states, last_states = extract_hidden_states(batch, model)

        optimizer.zero_grad()
        reconstructed = autoencoder(first_states)
        loss = loss_fn(reconstructed, last_states)
        loss.backward()
        optimizer.step()

        if loss < best_loss:
            best_loss = loss.item()
            best_lr = lr

        log_lrs.append(lr)
        losses.append(loss.item())

        # Log to wandb at every step
        wandb.log({
            "loss": loss.item(),
            "lr": lr,
            "batch_idx": batch_idx,
            "num_layers": num_layers,
            "seed": seed,
            "best_loss": best_loss,
            "best_lr": best_lr
        })

        pbar.update(1)
        pbar.set_postfix({"Loss": f"{loss.item():.4f}", "LR": f"{lr:.2e}"})

    pbar.close()

    result = {
        'seed': seed,
        'num_layers': num_layers,
        'hidden_dim_ratio': config.hidden_dim_ratio,
        'dropout': config.dropout,
        'final_loss': losses[-1],
        'final_lr': log_lrs[-1],
        'best_lr': best_lr,
        'best_loss': best_loss
    }

    logger.info(f"Experiment completed: {result}")
    return result

def main():
    config = Config()
    wandb.init(project="qwen-autoencoder-lr-finder", config=config.__dict__)

    logger.info("Loading Qwen model and tokenizer...")
    model = AutoModelForCausalLM.from_pretrained(config.model_name, torch_dtype=torch.bfloat16).to(config.device)
    tokenizer = AutoTokenizer.from_pretrained(config.model_name)

    logger.info("Loading data...")
    train_loader = load_data(config)

    logger.info("Starting experiments...")
    results = []
    for num_layers in range(4, 9):
        for seed in range(1, config.num_seeds + 1):
            # Start a new wandb run for each experiment
            with wandb.init(project="qwen-autoencoder-lr-finder", 
                            config=config.__dict__, 
                            group=f"layers_{num_layers}",
                            name=f"seed_{seed}",
                            job_type="experiment",
                            reinit=True):
                
                result = run_experiment(config, model, train_loader, num_layers, seed)
                results.append(result)
                
                # Save results after each experiment
                with open('lr_finder_results.jsonl', 'a') as f:
                    json.dump(result, f)
                    f.write('\n')
                
                # Log final results to wandb
                wandb.log(result)

    logger.info("Creating visualizations...")
    plot_results(results)
    create_heatmap(results)
    create_parallel_coordinates_plot(results)
    create_3d_scatter(results)

    logger.info("Experiment completed. Check WandB for detailed results and visualizations.")

if __name__ == "__main__":
    main()
def run_experiments_sequential(config, model, train_loader):
    results = []
    for num_layers in tqdm(range(4, 9), desc="Number of Layers"):
        for seed in tqdm(range(1, config.num_seeds + 1), desc="Seeds", leave=False):
            result = run_experiment(config, model, train_loader, num_layers, seed)
            results.append(result)
    return results

def plot_results(results):
    fig, axs = plt.subplots(3, 2, figsize=(20, 30))
    fig.suptitle('Learning Rate Finder Results')

    for i, num_layers in enumerate(range(4, 9)):
        layer_results = [r for r in results if r['num_layers'] == num_layers]
        best_lrs = [r['Best'] for r in layer_results]
        best_losses = [r['best_loss'] for r in layer_results]

        axs[i // 2, i % 2].scatter(best_lrs, best_losses, alpha=0.5)
        axs[i // 2, i % 2].set_xlabel('Best Learning Rate')
        axs[i // 2, i % 2].set_ylabel('Best Loss')
        axs[i // 2, i % 2].set_title(f'{num_layers} Layers')
        axs[i // 2, i % 2].set_xscale('log')
        axs[i // 2, i % 2].set_yscale('log')

    plt.tight_layout()
    wandb.log({"lr_loss_relationships": wandb.Image(plt)})
    plt.close()

def create_heatmap(results):
    layer_counts = len(set(r['num_layers'] for r in results))
    seed_counts = len(set(r['seed'] for r in results))
    
    heatmap_data = np.zeros((layer_counts, seed_counts))
    for r in results:
        layer_idx = r['num_layers'] - 4
        seed_idx = r['seed'] - 1
        heatmap_data[layer_idx, seed_idx] = r['best_loss']
    
    plt.figure(figsize=(20, 10))
    plt.imshow(heatmap_data, aspect='auto', cmap='viridis')
    plt.colorbar(label='Best Loss')
    plt.xlabel('Seed')
    plt.ylabel('Number of Layers')
    plt.title('Heatmap of Best Loss across Layers and Seeds')
    plt.tight_layout()
    wandb.log({"loss_heatmap": wandb.Image(plt)})
    plt.close()

def create_parallel_coordinates_plot(results):
    df = pd.DataFrame(results)
    
    plt.figure(figsize=(20, 10))
    pd.plotting.parallel_coordinates(df, 'num_layers', colormap='viridis')
    plt.title('Parallel Coordinates Plot of Hyperparameters')
    plt.tight_layout()
    wandb.log({"parallel_coordinates": wandb.Image(plt)})
    plt.close()

def create_3d_scatter(results):
    fig = plt.figure(figsize=(15, 15))
    ax = fig.add_subplot(111, projection='3d')
    
    for num_layers in range(4, 9):
        layer_results = [r for r in results if r['num_layers'] == num_layers]
        x = [r['Best'] for r in layer_results]
        y = [r['best_loss'] for r in layer_results]
        z = [r['seed'] for r in layer_results]
        ax.scatter(x, y, z, label=f'{num_layers} Layers')
    
    ax.set_xlabel('Best Learning Rate')
    ax.set_ylabel('Best Loss')
    ax.set_zlabel('Seed')
    ax.set_xscale('log')
    ax.set_yscale('log')
    ax.legend()
    plt.title('3D Scatter Plot of Best LR, Loss, and Seed')
    plt.tight_layout()
    wandb.log({"3d_scatter": wandb.Image(plt)})
    plt.close()

def main():
    mp.set_start_method('spawn')
    config = Config()
    wandb.init(project="qwen-autoencoder-lr-finder", config=config.__dict__)

    logger.info("Loading Qwen model and tokenizer...")
    model = AutoModelForCausalLM.from_pretrained(config.model_name, torch_dtype=torch.bfloat16).to(config.device)
    tokenizer = AutoTokenizer.from_pretrained(config.model_name)

    logger.info("Loading data...")
    train_loader = load_data(config)

    logger.info("Starting experiments...")
    results = run_experiments_sequential(config, model, train_loader)

    logger.info("Saving results...")
    with open('lr_finder_results.jsonl', 'w') as f:
        for result in results:
            json.dump(result, f)
            f.write('\n')

    logger.info("Creating visualizations...")
    plot_results(results)
    create_heatmap(results)
    create_parallel_coordinates_plot(results)
    create_3d_scatter(results)

    logger.info("Experiment completed. Check WandB for detailed results and visualizations.")
    wandb.finish()

if __name__ == "__main__":
    main()