train_ftp.py

import sys
import torch
sys.path.append("..")

from transformers import AutoTokenizer, AutoModelForCausalLM, Trainer, TrainingArguments, DataCollatorForLanguageModeling
from utils_llama import PERTURBATIONS, BABYLM_SPLITS, BABYLM_DATA_PATH, \
    GENRES, MARKER_TOKEN_IDS, marker_sg_token, marker_pl_token, marker_rev_token, write_file
from datasets import load_dataset
from FTP import AdamP

import wandb
import argparse
import copy
import math
import os

os.environ["TOKENIZERS_PARALLELISM"] = "false"

ftp_k = 1
class TrainerAdamP(Trainer):

    def create_optimizer(self):
        optimizer_params = {
            "lr": 5e-6,
            "weight_decay": 0.0,
            "k": ftp_k,  # Example parameter for AdamP
            "exclude_set": set()  # Use empty set if you don't want exclusion
        }

        # Cache pre-trained model weights 
        params_to_opt = [x[1] for x in self.model.named_parameters() if x[1].requires_grad]
        params_to_opt_name = [x[0] for x in self.model.named_parameters() if x[1].requires_grad]
        params_anchor = copy.deepcopy(params_to_opt)
        param_group = [{'params': params_to_opt, 'pre': params_anchor, 'name': params_to_opt_name}]

        # Initialize the AdamP optimizer
        self.optimizer = AdamP(param_group, **optimizer_params)



if __name__ == "__main__":

    # === CONFIGURATION SETTINGS ===
    parser = argparse.ArgumentParser(description="Training configuration.")

    parser.add_argument('--perturbation', type=str, default='hop_tokens4', help='Type of perturbation to use.')
    parser.add_argument('--train_set', type=str, default='10M', help='Dataset size for training.')
    parser.add_argument('--batch_size', type=int, default=3, help='Batch size for training.')
    parser.add_argument('--epoch', type=int, default=3, help='train epoch')
    parser.add_argument('--seed', type=int, default=0, help='Random seed.')
    parser.add_argument('--lr', type=float, default=5e-6, help='Learning rate.')

    args = parser.parse_args()

    # no_pos_encodings_underscore = ""  # Ex: "_nopos" if needed
    ckpt_path = "./checkpoints"
    # effective_bsz = 512

    model_name = "meta-llama/Llama-3.2-3B"
    model_save_name = "Llama-3.2-3B-FTP"

    # === FILE PATHS BASED ON CONFIGURATION ===

    
    wandb_id = f"{model_save_name}_{args.perturbation}_train_set_{args.train_set}_epoch_{args.epoch}_batch_size_{args.batch_size}_seed_{args.seed}_lr_{args.lr}_wandb_ftp_{ftp_k}"
    wandb.init(project="exp-impo-shuffle", group="ftp-1", name=wandb_id)
    wandb.config.update(args) 

    run_id = f"babylm_{args.perturbation}_{args.train_set}_seed{args.seed}"
    cache_dir = os.path.join(ckpt_path, f"{model_save_name}", run_id, "artifacts")
    run_dir = os.path.join(ckpt_path, f"{model_save_name}", run_id, "runs")
    os.makedirs(cache_dir, exist_ok=True)
    os.makedirs(run_dir, exist_ok=True)

    # === DATASET LOADING === 
    dataset_name = f"babylm_{args.perturbation}_{args.train_set}_seed{args.seed}"
    dataset = load_dataset('babylm_dataset_test.py', name=dataset_name, trust_remote_code=True)
    train_dataset = dataset['train']
    valid_dataset = dataset['validation']

    # === TOKENIZER & MODEL LOADING ===
    # model_name = f"gpt2{'' if no_pos_encodings_underscore == '' else '-no-pos'}-small-{perturbation}-{paren_model}"
 
    # tokenizer = AutoTokenizer.from_pretrained(model_name, cache_dir=cache_dir)
    tokenizer = PERTURBATIONS[args.perturbation]['llama_tokenizer']
    model = AutoModelForCausalLM.from_pretrained(model_name,  
                                                # device_map="auto",  # deepspeed needs to delete this setting
                                                cache_dir=cache_dir)

    # print("model:", model)
    # === TOKENIZATION ===
    def tokenize_function(examples): 
        return tokenizer(examples['text'], padding="max_length", truncation=True, max_length=1024)
    tokenized_train = train_dataset.map(tokenize_function, batched=True, remove_columns=["text"])
    tokenized_valid = valid_dataset.map(tokenize_function, batched=True, remove_columns=["text"])

    shuffled_valid = tokenized_valid.shuffle()
    tokenized_valid = shuffled_valid.select(range(1000))
    print("tokenized_valid:", tokenized_valid)
    # print(train_dataset)
    # === DATA COLLATOR ===2
    data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)

    # === TRAINING ARGUMENTS ===
    training_args = TrainingArguments(
        output_dir=run_dir,
        evaluation_strategy="steps",
        eval_steps=10,
        per_device_train_batch_size=args.batch_size,  # set "auto" in deepspeed config, adjust it in trainer
        logging_dir='./logs',
        logging_steps=1,
        save_steps=100,
        learning_rate=args.lr, # align with deepspeed
        num_train_epochs=args.epoch,
        seed=args.seed,
        gradient_accumulation_steps=2, # # set "auto" in deepspeed config, adjust it in trainer
        fp16=True, # align with deepspeed
        report_to="wandb",
        warmup_ratio=0.1, 
        deepspeed="deepspeed_config/train_dp_config.json"
    )

    # === TRAINER ===
    trainer = TrainerAdamP(
        model=model,
        args=training_args,
        train_dataset=tokenized_train,
        eval_dataset=tokenized_valid,
        tokenizer=tokenizer,
        data_collator=data_collator
    )
    
    # === TRAIN MODEL ===
    trainer.train()
    # End logging
    wandb.finish()