lightning/pytorch_lightning/accelerators/base_accelerator.py

import os
import math
from enum import Enum
from typing import Any

import torch

from pytorch_lightning.utilities import AMPType, rank_zero_warn
from pytorch_lightning.utilities.apply_func import move_data_to_device
from pytorch_lightning.utilities.exceptions import MisconfigurationException
from pytorch_lightning.utilities.parsing import AttributeDict
import torch.distributed as torch_distrib
from pytorch_lightning import _logger as log

try:
    from apex import amp
except ImportError:
    amp = None

EPSILON = 1e-6
EPSILON_FP16 = 1e-5


class Accelerator(object):

    def __init__(self, trainer, cluster_environment=None):
        self.trainer = trainer
        self.cluster_environment = cluster_environment
        self.dist = AttributeDict(rank=0, device=None)
        self.train_loop = self.trainer.train
        self.validation_loop = self.trainer.run_evaluation
        self.test_loop = self.trainer.run_evaluation

    def setup(self, model):
        pass

    def teardown(self):
        pass

    def barrier(self, name: str = None):
        pass

    def broadcast(self, obj, src=0):
        return obj

    def train_or_test(self):
        if self.trainer.testing:
            results = self.trainer.run_test()
        else:
            results = self.trainer.train()
        return results

    def batch_to_device(self, batch: Any, device: torch.device):
        model = self.trainer.get_model()
        if model is not None:
            return model.transfer_batch_to_device(batch, device)
        return move_data_to_device(batch, device)

    def training_step_end(self, output):
        return output

    def test_step_end(self, output):
        return output

    def validation_step_end(self, output):
        return output

    def process_dataloader(self, dataloader):
        return dataloader

    def backward(self, closure_loss, optimizer, opt_idx):
        model_ref = self.trainer.get_model()

        # scale loss for 16 bit
        if self.trainer.precision == 16:
            closure_loss = model_ref.amp_scale_loss(
                closure_loss,
                optimizer,
                opt_idx,
                amp_backend=self.trainer.amp_backend
            )

            # enter amp context
            if self.trainer.amp_backend == AMPType.APEX:
                self.trainer.dev_debugger.track_event('AMP', str(AMPType.APEX))
                context = closure_loss
                closure_loss = closure_loss.__enter__()

        # do backward pass
        model_ref.backward(self, closure_loss, optimizer, opt_idx)

        # exit amp context
        if self.trainer.precision == 16 and self.trainer.amp_backend == AMPType.APEX:
            a, b, c = None, None, None
            error = context.__exit__(a, b, c)
            if error:
                rank_zero_warn(a, b, c)
                raise Exception('apex unscale error')

        # once backward has been applied, release graph
        closure_loss = closure_loss.detach()
        return closure_loss

    def optimizer_step(self, optimizer, batch_idx, opt_idx, lambda_closure):
        model_ref = self.trainer.get_model()
        is_lbfgs = isinstance(optimizer, torch.optim.LBFGS)
        native_amp = self.trainer.amp_backend == AMPType.NATIVE

        # native amp + lbfgs is a no go right now
        if native_amp and is_lbfgs:
            raise MisconfigurationException(
                'native PyTorch amp and lbfgs are not compatible.'
                ' To request, please file a Github issue in PyTorch and tag @mcarilli')

        # model hook
        model_ref.optimizer_step(
            self.trainer.current_epoch,
            batch_idx,
            optimizer,
            opt_idx,
            lambda_closure,
            using_native_amp=native_amp,
            using_lbfgs=is_lbfgs
        )

        # scale when native amp
        if native_amp:
            self.trainer.scaler.update()

    def optimizer_zero_grad(self, batch_idx, optimizer, opt_idx):
        model_ref = self.trainer.get_model()
        model_ref.optimizer_zero_grad(self.trainer.current_epoch, batch_idx, optimizer, opt_idx)

    def clip_gradients(self, optimizer):

        if self.trainer.amp_backend == AMPType.NATIVE:
            self.trainer.scaler.unscale_(optimizer)

        # apply clip gradients
        # TODO: separate TPU case from here
        self._clip_gradients(optimizer)

    def _clip_gradients(self, optimizer):
        # this code is a modification of torch.nn.utils.clip_grad_norm_
        # with TPU support based on https://github.com/pytorch/xla/blob/master/TROUBLESHOOTING.md
        if self.trainer.gradient_clip_val <= 0:
            return

        model = self.trainer.get_model()
        if self.trainer.amp_backend == AMPType.APEX:
            parameters = amp.master_params(optimizer)
        else:
            parameters = model.parameters()

        max_norm = float(self.trainer.gradient_clip_val)
        norm_type = float(2.0)

        if isinstance(parameters, torch.Tensor):
            parameters = [parameters]
        parameters = list(filter(lambda p: p.grad is not None, parameters))

        if norm_type == math.inf:
            total_norm = max(p.grad.data.abs().max() for p in parameters)
        else:
            device = parameters[0].device
            out = torch.empty(len(parameters), device=device)
            for i, p in enumerate(parameters):
                torch.norm(p.grad.data.to(device), norm_type, out=out[i])
            total_norm = torch.norm(out, norm_type)

        eps = EPSILON_FP16 if self.trainer.precision == 16 else EPSILON
        clip_coef = torch.tensor(max_norm, device=device) / (total_norm + eps)
        clip_coef = torch.min(clip_coef, torch.ones_like(clip_coef))
        for p in parameters:
            p.grad.data.mul_(clip_coef.to(p.grad.data.device))

    def on_train_epoch_end(self, outputs):
        pass

    def on_train_end(self):
        pass

    def early_stopping_should_stop(self, pl_module):
        return self.trainer.should_stop

    def setup_optimizers(self, model):
        if self.trainer.testing is True:
            return

        optimizers, lr_schedulers, optimizer_frequencies = self.trainer.init_optimizers(model)
        self.trainer.optimizers = optimizers
        self.trainer.lr_schedulers = lr_schedulers
        self.trainer.optimizer_frequencies = optimizer_frequencies

    def init_ddp_connection(
        self, global_rank: int, world_size: int, is_slurm_managing_tasks: bool = True
    ) -> None:
        if is_slurm_managing_tasks:
            self.trainer.slurm_connector.connect_ddp(global_rank, world_size)
        else:
            self.connect_torchelastic(global_rank, world_size)

    def connect_torchelastic(
        self, global_rank: int, world_size: int
    ) -> None:
        """
        Override to define your custom way of setting up a distributed environment.

        Lightning's implementation uses env:// init by default and sets the first node as root
        for SLURM managed cluster.

        Args:
            global_rank: The global process idx.
            world_size: Number of GPUs being use across all nodes. (num_nodes * num_gpus).
        """

        if "MASTER_ADDR" not in os.environ:
            rank_zero_warn(
                "MASTER_ADDR environment variable is not defined. Set as localhost"
            )
            os.environ["MASTER_ADDR"] = "127.0.0.1"
        log.debug(f"MASTER_ADDR: {os.environ['MASTER_ADDR']}")

        if "MASTER_PORT" not in os.environ:
            rank_zero_warn(
                "MASTER_PORT environment variable is not defined. Set as 12910"
            )
            os.environ["MASTER_PORT"] = "12910"
        log.debug(f"MASTER_PORT: {os.environ['MASTER_PORT']}")

        if "WORLD_SIZE" in os.environ and int(os.environ["WORLD_SIZE"]) != world_size:
            rank_zero_warn(
                f"WORLD_SIZE environment variable ({os.environ['WORLD_SIZE']}) "
                f"is not equal to the computed world size ({world_size}). Ignored."
            )

        torch_backend = "nccl" if self.trainer.on_gpu else "gloo"

        if not torch.distributed.is_initialized():
            log.info(
                f"initializing ddp: GLOBAL_RANK: {global_rank}, MEMBER: {global_rank + 1}/{world_size}"
            )
            torch_distrib.init_process_group(
                torch_backend, rank=global_rank, world_size=world_size
            )


# TODO: allow user to compare with string even internaly we shall use these Enum to prevent typos...
class BackendType(Enum):
    DP = 'dp'
    DDP = 'ddp'
    DDP2 = 'ddp2'
    DDP_SPAWN = 'ddp_spawn'
    # decuple distrib and device
    DDP_CPU = 'ddp_cpu'
    HOROVOD = 'horovod'
    # this is rather device
    TPU = 'tpu'