train.py

"""
To run the program execute the below command -
$ python train.py
"""

import os
import time
import math
import pickle

import numpy as np
import torch
import wandb

from model import GPT
from config import config

master_process = True
seed_offset = 0

for key, value in config.items():
    globals()[key] = value

tokens_per_iter = gradient_accumulation_steps * batch_size * block_size
print(f"tokens per iteration will be: {tokens_per_iter:,}")

os.makedirs(out_dir, exist_ok=True)

torch.manual_seed(1337 + seed_offset)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True

device = 'cuda'
config['device'] = device
dtype = 'bfloat16' if torch.cuda.is_available() and torch.cuda.is_bf16_supported() else 'float16'
config['dtype'] = dtype

device_type = 'cuda' if 'cuda' in device else 'cpu'

# note: float16 data type will automatically use a GradScaler
ptdtype = {'float32': torch.float32, 'bfloat16': torch.bfloat16, 'float16': torch.float16}[dtype]
ctx = torch.amp.autocast(device_type=device_type, dtype=ptdtype)

data_dir = 'data/openwebtext'
train_data = np.memmap(os.path.join(data_dir, 'train.bin'), dtype=np.uint16, mode='r')
# Use only 70% of the training data.
train_data = train_data[:int(0.7*len(train_data))]
val_data = np.memmap(os.path.join(data_dir, 'val.bin'), dtype=np.uint16, mode='r')

# TODO; Implement a DataLoader for this
def get_batch(split):
    data = train_data if split == 'train' else val_data
    ix = torch.randint(len(data) - block_size, (batch_size,))
    x = torch.stack([torch.from_numpy((data[i:i+block_size]).astype(np.int64)) for i in ix])
    y = torch.stack([torch.from_numpy((data[i+1:i+1+block_size]).astype(np.int64)) for i in ix])
    if device_type == 'cuda':
        x, y = x.pin_memory().to(device, non_blocking=True), y.pin_memory().to(device, non_blocking=True)
    else:
        x, y = x.to(device), y.to(device)
    return x, y

iter_num = 0
best_val_loss = 1e9

model = GPT(n_layer=n_layer, n_head=n_head, n_embd=n_embd, block_size=block_size,
                  bias=bias, dropout=dropout)

model.to(device)

# initialize a GradScaler. If enabled=False scaler is a no-op
scaler = torch.cuda.amp.GradScaler(enabled=(dtype == 'float16'))

# optimizer
optimizer = model.configure_optimizers(weight_decay, learning_rate, (beta1, beta2), device_type)
checkpoint = None

unoptimized_model = model

@torch.no_grad()
def estimate_loss():
    out = {}
    model.eval()
    for split in ['train', 'val']:
        losses = torch.zeros(eval_iters)
        for k in range(eval_iters):
            X, Y = get_batch(split)
            with ctx:
                logits, loss = model(X, Y)
            losses[k] = loss.item()
        out[split] = losses.mean()
    model.train()
    return out

# learning rate scheduler with decay and linear warmup according to the GPT paper.
def get_lr(it):
    if it < warmup_iters:
        return learning_rate * it / warmup_iters
    if it > lr_decay_iters:
        return min_lr
    decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters)
    assert 0 <= decay_ratio <= 1
    coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio))
    return min_lr + coeff * (learning_rate - min_lr)

# logging
if wandb_log and master_process:
    import wandb
    wandb.login(key="42926c014f7382a7434aad694030d03d52f66025")
    wandb.init(project=wandb_project, name=wandb_run_name, config=config)

# training loop
X, Y = get_batch('train') # fetch the very first batch
t0 = time.time()
local_iter_num = 0 # number of iterations in the lifetime of this process
raw_model = model

while True:
    # determine and set the learning rate for this iteration
    lr = get_lr(iter_num) if decay_lr else learning_rate
    for param_group in optimizer.param_groups:
        param_group['lr'] = lr

    # evaluate the loss on train/val sets and write checkpoints
    if iter_num % eval_interval == 0 and master_process:
        losses = estimate_loss()
        print(f"step {iter_num}: train loss {losses['train']:.4f}, val loss {losses['val']:.4f}")
        # Wandb logging
        if wandb_log:
            wandb.log({
                "iter": iter_num,
                "train/loss": losses['train'],
                "val/loss": losses['val'],
                "lr": lr,
            })
    if iter_num == 0 and eval_only:
        break

    for micro_step in range(gradient_accumulation_steps):
        with ctx:
            logits, loss = model(X, Y)
            loss = loss / gradient_accumulation_steps 

        X, Y = get_batch('train')
        scaler.scale(loss).backward()
    
    # Gradient Clipping
    if grad_clip != 0.0:
        scaler.unscale_(optimizer)
        torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)

    scaler.step(optimizer)
    scaler.update()
    optimizer.zero_grad(set_to_none=True)

    # timing and logging
    t1 = time.time()
    dt = t1 - t0
    t0 = t1
    if iter_num % log_interval == 0 and master_process:
        lossf = loss.item() * gradient_accumulation_steps
        print(f"iter {iter_num}: loss {lossf:.4f}, time {dt*1000:.2f}ms")
    iter_num += 1
    local_iter_num += 1

    if iter_num > max_iters:
        break