runner/train.py

# Copyright 2024 ByteDance and/or its affiliates.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import datetime
import hashlib
import logging
import os
import time
from contextlib import nullcontext

import torch
import torch.distributed as dist
import wandb
from ml_collections.config_dict import ConfigDict
from torch.nn.parallel import DistributedDataParallel as DDP
from tqdm import tqdm

from configs.configs_base import configs as configs_base
from configs.configs_data import data_configs
from configs.configs_model_type import model_configs
from protenix.config import parse_configs, parse_sys_args
from protenix.config.config import save_config
from protenix.data.dataloader import get_dataloaders
from protenix.metrics.lddt_metrics import LDDTMetrics
from protenix.model.loss import ProtenixLoss
from protenix.model.protenix import Protenix
from protenix.utils.distributed import DIST_WRAPPER
from protenix.utils.lr_scheduler import FinetuneLRScheduler, get_lr_scheduler
from protenix.utils.metrics import SimpleMetricAggregator
from protenix.utils.permutation.permutation import SymmetricPermutation
from protenix.utils.seed import seed_everything
from protenix.utils.torch_utils import autocasting_disable_decorator, to_device
from protenix.utils.training import get_optimizer, is_loss_nan_check
from runner.ema import EMAWrapper

# Disable WANDB's console output capture to reduce unnecessary logging
os.environ["WANDB_CONSOLE"] = "off"


class AF3Trainer(object):
    def __init__(self, configs):
        self.configs = configs
        self.init_env()
        self.init_basics()
        self.init_log()
        self.init_model()
        self.init_loss()
        self.init_data()
        self.try_load_checkpoint()

    def init_basics(self):
        # Step means effective step considering accumulation
        self.step = 0
        # Global_step equals to self.step * self.iters_to_accumulate
        self.global_step = 0
        self.start_step = 0
        # Add for grad accumulation, it can increase real batch size
        self.iters_to_accumulate = self.configs.iters_to_accumulate

        self.run_name = self.configs.run_name + "_" + time.strftime("%Y%m%d_%H%M%S")
        run_names = DIST_WRAPPER.all_gather_object(
            self.run_name if DIST_WRAPPER.rank == 0 else None
        )
        self.run_name = [name for name in run_names if name is not None][0]
        self.run_dir = f"{self.configs.base_dir}/{self.run_name}"
        self.checkpoint_dir = f"{self.run_dir}/checkpoints"
        self.prediction_dir = f"{self.run_dir}/predictions"
        self.structure_dir = f"{self.run_dir}/structures"
        self.dump_dir = f"{self.run_dir}/dumps"
        self.error_dir = f"{self.run_dir}/errors"

        if DIST_WRAPPER.rank == 0:
            os.makedirs(self.run_dir)
            os.makedirs(self.checkpoint_dir)
            os.makedirs(self.prediction_dir)
            os.makedirs(self.structure_dir)
            os.makedirs(self.dump_dir)
            os.makedirs(self.error_dir)
            save_config(
                self.configs,
                os.path.join(self.configs.base_dir, self.run_name, "config.yaml"),
            )

        self.print(
            f"Using run name: {self.run_name}, run dir: {self.run_dir}, checkpoint_dir: "
            + f"{self.checkpoint_dir}, prediction_dir: {self.prediction_dir}, structure_dir: "
            + f"{self.structure_dir}, error_dir: {self.error_dir}"
        )

    def init_log(self):
        if self.configs.use_wandb and DIST_WRAPPER.rank == 0:
            wandb.init(
                project=self.configs.project,
                name=self.run_name,
                config=vars(self.configs),
                id=self.configs.wandb_id or None,
            )
        self.train_metric_wrapper = SimpleMetricAggregator(["avg"])

    def init_env(self):
        """Init pytorch/cuda envs."""
        logging.info(
            f"Distributed environment: world size: {DIST_WRAPPER.world_size}, "
            + f"global rank: {DIST_WRAPPER.rank}, local rank: {DIST_WRAPPER.local_rank}"
        )
        self.use_cuda = torch.cuda.device_count() > 0
        if self.use_cuda:
            self.device = torch.device("cuda:{}".format(DIST_WRAPPER.local_rank))
            os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
            all_gpu_ids = ",".join(str(x) for x in range(torch.cuda.device_count()))
            devices = os.getenv("CUDA_VISIBLE_DEVICES", all_gpu_ids)
            logging.info(
                f"LOCAL_RANK: {DIST_WRAPPER.local_rank} - CUDA_VISIBLE_DEVICES: [{devices}]"
            )
            torch.cuda.set_device(self.device)
        else:
            self.device = torch.device("cpu")
        if DIST_WRAPPER.world_size > 1:
            timeout_seconds = int(os.environ.get("NCCL_TIMEOUT_SECOND", 600))
            dist.init_process_group(
                backend="nccl", timeout=datetime.timedelta(seconds=timeout_seconds)
            )
        if not self.configs.deterministic_seed:
            # use rank-specific seed
            hash_string = f"({self.configs.seed},{DIST_WRAPPER.rank},init_seed)"
            rank_seed = int(hashlib.sha256(hash_string.encode("utf8")).hexdigest(), 16)
            rank_seed = rank_seed % (2**32)
        else:
            rank_seed = self.configs.seed
        seed_everything(
            seed=rank_seed,
            deterministic=self.configs.deterministic,
        )  # diff ddp process got diff seeds

        if self.configs.use_deepspeed_evo_attention:
            env = os.getenv("CUTLASS_PATH", None)
            print(f"env: {env}")
            assert (
                env is not None
            ), "if use ds4sci, set env as https://www.deepspeed.ai/tutorials/ds4sci_evoformerattention/"
        logging.info("Finished init ENV.")

    def init_loss(self):
        self.loss = ProtenixLoss(self.configs)
        self.symmetric_permutation = SymmetricPermutation(
            self.configs, error_dir=self.error_dir
        )
        self.lddt_metrics = LDDTMetrics(self.configs)

    def init_model(self):
        self.raw_model = Protenix(self.configs).to(self.device)
        self.use_ddp = False
        if DIST_WRAPPER.world_size > 1:
            self.print(f"Using DDP")
            self.use_ddp = True
            # Fix DDP/checkpoint https://discuss.pytorch.org/t/ddp-and-gradient-checkpointing/132244
            self.model = DDP(
                self.raw_model,
                find_unused_parameters=self.configs.find_unused_parameters,
                device_ids=[DIST_WRAPPER.local_rank],
                output_device=DIST_WRAPPER.local_rank,
                static_graph=True,
            )
        else:
            self.model = self.raw_model

        def count_parameters(model):
            total_params = sum(p.numel() for p in model.parameters())
            return total_params / 1000.0 / 1000.0

        self.print(f"Model Parameters: {count_parameters(self.model)}")
        if self.configs.get("ema_decay", -1) > 0:
            assert self.configs.ema_decay < 1
            self.ema_wrapper = EMAWrapper(
                self.model,
                self.configs.ema_decay,
                self.configs.ema_mutable_param_keywords,
            )
            self.ema_wrapper.register()

        torch.cuda.empty_cache()
        self.optimizer = get_optimizer(
            self.configs,
            self.model,
            param_names=self.configs.get("finetune_params_with_substring", [""]),
        )
        self.init_scheduler()

    def init_scheduler(self, **kwargs):
        # init finetune lr scheduler if available
        finetune_params = self.configs.get("finetune_params_with_substring", [""])
        is_finetune = len(finetune_params[0]) > 0

        if is_finetune:
            self.lr_scheduler = FinetuneLRScheduler(
                self.optimizer,
                self.configs,
                self.configs.finetune,
                **kwargs,
            )
        else:
            self.lr_scheduler = get_lr_scheduler(self.configs, self.optimizer, **kwargs)

    def init_data(self):
        self.train_dl, self.test_dls = get_dataloaders(
            self.configs,
            DIST_WRAPPER.world_size,
            seed=self.configs.seed,
            error_dir=self.error_dir,
        )

    def save_checkpoint(self, ema_suffix=""):
        if DIST_WRAPPER.rank == 0:
            path = f"{self.checkpoint_dir}/{self.step}{ema_suffix}.pt"
            checkpoint = {
                "model": self.model.state_dict(),
                "optimizer": self.optimizer.state_dict(),
                "scheduler": (
                    self.lr_scheduler.state_dict()
                    if self.lr_scheduler is not None
                    else None
                ),
                "step": self.step,
            }
            torch.save(checkpoint, path)
            self.print(f"Saved checkpoint to {path}")

    def try_load_checkpoint(self):

        def _load_checkpoint(
            checkpoint_path: str,
            load_params_only: bool,
            skip_load_optimizer: bool = False,
            skip_load_step: bool = False,
            skip_load_scheduler: bool = False,
            load_step_for_scheduler: bool = True,
        ):
            if not os.path.exists(checkpoint_path):
                raise Exception(f"Given checkpoint path not exist [{checkpoint_path}]")
            self.print(
                f"Loading from {checkpoint_path}, strict: {self.configs.load_strict}"
            )
            checkpoint = torch.load(checkpoint_path, self.device)
            sample_key = [k for k in checkpoint["model"].keys()][0]
            self.print(f"Sampled key: {sample_key}")
            if sample_key.startswith("module.") and not self.use_ddp:
                # DDP checkpoint has module. prefix
                checkpoint["model"] = {
                    k[len("module.") :]: v for k, v in checkpoint["model"].items()
                }

            self.model.load_state_dict(
                state_dict=checkpoint["model"],
                strict=self.configs.load_strict,
            )
            if not load_params_only:
                if not skip_load_optimizer:
                    self.print(f"Loading optimizer state")
                    self.optimizer.load_state_dict(checkpoint["optimizer"])
                if not skip_load_step:
                    self.print(f"Loading checkpoint step")
                    self.step = checkpoint["step"] + 1
                    self.start_step = self.step
                    self.global_step = self.step * self.iters_to_accumulate
                if not skip_load_scheduler:
                    self.print(f"Loading scheduler state")
                    self.lr_scheduler.load_state_dict(checkpoint["scheduler"])
                elif load_step_for_scheduler:
                    assert (
                        not skip_load_step
                    ), "if load_step_for_scheduler is True, you must load step first"
                    # reinitialize LR scheduler using the updated optimizer and step
                    self.init_scheduler(last_epoch=self.step - 1)

            self.print(f"Finish loading checkpoint, current step: {self.step}")

        # Load EMA model parameters
        if self.configs.load_ema_checkpoint_path:
            _load_checkpoint(
                self.configs.load_ema_checkpoint_path,
                load_params_only=True,
            )
            self.ema_wrapper.register()

        # Load model
        if self.configs.load_checkpoint_path:
            _load_checkpoint(
                self.configs.load_checkpoint_path,
                self.configs.load_params_only,
                skip_load_optimizer=self.configs.skip_load_optimizer,
                skip_load_scheduler=self.configs.skip_load_scheduler,
                skip_load_step=self.configs.skip_load_step,
                load_step_for_scheduler=self.configs.load_step_for_scheduler,
            )

    def print(self, msg: str):
        if DIST_WRAPPER.rank == 0:
            logging.info(msg)

    def model_forward(self, batch: dict, mode: str = "train") -> tuple[dict, dict]:
        assert mode in ["train", "eval"]
        batch["pred_dict"], batch["label_dict"], log_dict = self.model(
            input_feature_dict=batch["input_feature_dict"],
            label_dict=batch["label_dict"],
            label_full_dict=batch["label_full_dict"],
            mode=mode,
            current_step=self.step if mode == "train" else None,
            symmetric_permutation=self.symmetric_permutation,
        )
        return batch, log_dict

    def get_loss(
        self, batch: dict, mode: str = "train"
    ) -> tuple[torch.Tensor, dict, dict]:
        assert mode in ["train", "eval"]

        loss, loss_dict = autocasting_disable_decorator(self.configs.skip_amp.loss)(
            self.loss
        )(
            feat_dict=batch["input_feature_dict"],
            pred_dict=batch["pred_dict"],
            label_dict=batch["label_dict"],
            mode=mode,
        )
        return loss, loss_dict, batch

    @torch.no_grad()
    def get_metrics(self, batch: dict) -> dict:

        lddt_dict = self.lddt_metrics.compute_lddt(
            batch["pred_dict"], batch["label_dict"]
        )

        return lddt_dict

    @torch.no_grad()
    def aggregate_metrics(self, lddt_dict: dict, batch: dict) -> dict:

        simple_metrics, _ = self.lddt_metrics.aggregate_lddt(
            lddt_dict, batch["pred_dict"]["summary_confidence"]
        )

        return simple_metrics

    @torch.no_grad()
    def evaluate(self, mode: str = "eval"):
        if not self.configs.eval_ema_only:
            self._evaluate()
        if hasattr(self, "ema_wrapper"):
            self.ema_wrapper.apply_shadow()
            self._evaluate(ema_suffix=f"ema{self.ema_wrapper.decay}_", mode=mode)
            self.ema_wrapper.restore()

    @torch.no_grad()
    def _evaluate(self, ema_suffix: str = "", mode: str = "eval"):
        # Init Metric Aggregator
        simple_metric_wrapper = SimpleMetricAggregator(["avg"])
        eval_precision = {
            "fp32": torch.float32,
            "bf16": torch.bfloat16,
            "fp16": torch.float16,
        }[self.configs.dtype]
        enable_amp = (
            torch.autocast(device_type="cuda", dtype=eval_precision)
            if torch.cuda.is_available()
            else nullcontext()
        )
        self.model.eval()

        for test_name, test_dl in self.test_dls.items():
            self.print(f"Testing on {test_name}")
            evaluated_pids = []
            total_batch_num = len(test_dl)
            for index, batch in enumerate(tqdm(test_dl)):
                batch = to_device(batch, self.device)
                pid = batch["basic"]["pdb_id"]

                if index + 1 == total_batch_num and DIST_WRAPPER.world_size > 1:
                    # Gather all pids across ranks for avoiding duplicated evaluations when drop_last = False
                    all_data_ids = DIST_WRAPPER.all_gather_object(evaluated_pids)
                    dedup_ids = set(sum(all_data_ids, []))
                    if pid in dedup_ids:
                        print(
                            f"Rank {DIST_WRAPPER.rank}: Drop data_id {pid} as it is already evaluated."
                        )
                        break
                evaluated_pids.append(pid)

                simple_metrics = {}
                with enable_amp:
                    # Model forward
                    batch, _ = self.model_forward(batch, mode=mode)
                    # Loss forward
                    loss, loss_dict, batch = self.get_loss(batch, mode="eval")
                    # lDDT metrics
                    lddt_dict = self.get_metrics(batch)
                    lddt_metrics = self.aggregate_metrics(lddt_dict, batch)
                    simple_metrics.update(
                        {k: v for k, v in lddt_metrics.items() if "diff" not in k}
                    )
                    simple_metrics.update(loss_dict)

                # Metrics
                for key, value in simple_metrics.items():
                    simple_metric_wrapper.add(
                        f"{ema_suffix}{key}", value, namespace=test_name
                    )

                del batch, simple_metrics
                if index % 5 == 0:
                    # Release some memory periodically
                    torch.cuda.empty_cache()

            metrics = simple_metric_wrapper.calc()
            self.print(f"Step {self.step}, eval {test_name}: {metrics}")
            if self.configs.use_wandb and DIST_WRAPPER.rank == 0:
                wandb.log(metrics, step=self.step)

    def update(self):
        # Clip the gradient
        if self.configs.grad_clip_norm != 0.0:
            torch.nn.utils.clip_grad_norm_(
                self.model.parameters(), self.configs.grad_clip_norm
            )

    def train_step(self, batch: dict):
        self.model.train()
        # FP16 training has not been verified yet
        train_precision = {
            "fp32": torch.float32,
            "bf16": torch.bfloat16,
            "fp16": torch.float16,
        }[self.configs.dtype]
        enable_amp = (
            torch.autocast(
                device_type="cuda", dtype=train_precision, cache_enabled=False
            )
            if torch.cuda.is_available()
            else nullcontext()
        )

        scaler = torch.GradScaler(
            device="cuda" if torch.cuda.is_available() else "cpu",
            enabled=(self.configs.dtype == "float16"),
        )

        with enable_amp:
            batch, _ = self.model_forward(batch, mode="train")
            loss, loss_dict, _ = self.get_loss(batch, mode="train")

        if self.configs.dtype in ["bf16", "fp32"]:
            if is_loss_nan_check(loss):
                self.print(f"Skip iteration with NaN loss: {self.step} steps")
                loss = torch.tensor(0.0, device=loss.device, requires_grad=True)
        scaler.scale(loss / self.iters_to_accumulate).backward()

        # For simplicity, the global training step is used
        if (self.global_step + 1) % self.iters_to_accumulate == 0:
            self.print(
                f"self.step {self.step}, self.iters_to_accumulate: {self.iters_to_accumulate}"
            )
            # Unscales the gradients of optimizer's assigned parameters in-place
            scaler.unscale_(self.optimizer)
            # Do grad clip only
            self.update()
            scaler.step(self.optimizer)
            scaler.update()
            self.optimizer.zero_grad(set_to_none=True)
            self.lr_scheduler.step()
        for key, value in loss_dict.items():
            if "loss" not in key:
                continue
            self.train_metric_wrapper.add(key, value, namespace="train")
        torch.cuda.empty_cache()

    def progress_bar(self, desc: str = ""):
        if DIST_WRAPPER.rank != 0:
            return
        if self.global_step % (
            self.configs.eval_interval * self.iters_to_accumulate
        ) == 0 or (not hasattr(self, "_ipbar")):
            # Start a new progress bar
            self._pbar = tqdm(
                range(
                    self.global_step
                    % (self.iters_to_accumulate * self.configs.eval_interval),
                    self.iters_to_accumulate * self.configs.eval_interval,
                )
            )
            self._ipbar = iter(self._pbar)

        step = next(self._ipbar)
        self._pbar.set_description(
            f"[step {self.step}: {step}/{self.iters_to_accumulate * self.configs.eval_interval}] {desc}"
        )
        return

    def run(self):
        """
        Main entry for the AF3Trainer.

        This function handles the training process, evaluation, logging, and checkpoint saving.
        """
        if self.configs.eval_only or self.configs.eval_first:
            self.evaluate()
            if self.configs.eval_only:
                return
        use_ema = hasattr(self, "ema_wrapper")
        self.print(f"Using ema: {use_ema}")

        while True:
            for batch in self.train_dl:
                is_update_step = (self.global_step + 1) % self.iters_to_accumulate == 0
                is_last_step = (self.step + 1) == self.configs.max_steps
                step_need_log = (self.step + 1) % self.configs.log_interval == 0

                step_need_eval = (
                    self.configs.eval_interval > 0
                    and (self.step + 1) % self.configs.eval_interval == 0
                )
                step_need_save = (
                    self.configs.checkpoint_interval > 0
                    and (self.step + 1) % self.configs.checkpoint_interval == 0
                )

                is_last_step &= is_update_step
                step_need_log &= is_update_step
                step_need_eval &= is_update_step
                step_need_save &= is_update_step

                batch = to_device(batch, self.device)
                self.progress_bar()
                self.train_step(batch)
                if use_ema and is_update_step:
                    self.ema_wrapper.update()
                if step_need_log or is_last_step:
                    metrics = self.train_metric_wrapper.calc()
                    self.print(f"Step {self.step} train: {metrics}")
                    last_lr = self.lr_scheduler.get_last_lr()
                    if DIST_WRAPPER.rank == 0:
                        if self.configs.use_wandb:
                            lr_dict = {"train/lr": last_lr[0]}
                            for group_i, group_lr in enumerate(last_lr):
                                lr_dict[f"train/group{group_i}_lr"] = group_lr
                            wandb.log(lr_dict, step=self.step)
                        self.print(f"Step {self.step}, lr: {last_lr}")
                    if self.configs.use_wandb and DIST_WRAPPER.rank == 0:
                        wandb.log(metrics, step=self.step)

                if step_need_save or is_last_step:
                    self.save_checkpoint()
                    if use_ema:
                        self.ema_wrapper.apply_shadow()
                        self.save_checkpoint(
                            ema_suffix=f"_ema_{self.ema_wrapper.decay}"
                        )
                        self.ema_wrapper.restore()

                if step_need_eval or is_last_step:
                    self.evaluate()
                self.global_step += 1
                if self.global_step % self.iters_to_accumulate == 0:
                    self.step += 1
                if self.step >= self.configs.max_steps:
                    self.print(f"Finish training after {self.step} steps")
                    break
            if self.step >= self.configs.max_steps:
                break


def main():
    LOG_FORMAT = "%(asctime)s,%(msecs)-3d %(levelname)-8s [%(filename)s:%(lineno)s %(funcName)s] %(message)s"
    logging.basicConfig(
        format=LOG_FORMAT,
        level=logging.INFO,
        datefmt="%Y-%m-%d %H:%M:%S",
        filemode="w",
    )
    configs_base["use_deepspeed_evo_attention"] = (
        os.environ.get("USE_DEEPSPEED_EVO_ATTENTION", False) == "true"
    )
    configs = {**configs_base, **{"data": data_configs}}
    configs = parse_configs(
        configs,
        parse_sys_args(),
    )
    model_name = configs.model_name
    model_specfics_configs = ConfigDict(model_configs[model_name])
    # update model specific configs
    configs.update(model_specfics_configs)

    print(configs.run_name)
    print(configs)
    trainer = AF3Trainer(configs)
    trainer.run()


if __name__ == "__main__":
    main()