FFCV

Source code for ffcv.loader.loader

"""
FFCV loader
"""
import enum
from os import environ
import ast
from multiprocessing import cpu_count
from re import sub
from typing import Any, Callable, Mapping, Sequence, Type, Union, Literal
from collections import defaultdict
from enum import Enum, unique, auto
from ffcv.fields.base import Field

import torch as ch
import numpy as np

from .epoch_iterator import EpochIterator
from ..reader import Reader
from ..traversal_order.base import TraversalOrder
from ..traversal_order import Random, Sequential, QuasiRandom
from ..pipeline import Pipeline
from ..pipeline.compiler import Compiler
from ..pipeline.operation import Operation
from ..transforms.ops import ToTensor
from ..transforms.module import ModuleWrapper
from ..memory_managers import (
    ProcessCacheManager, OSCacheManager, MemoryManager
)

[docs]@unique class OrderOption(Enum): SEQUENTIAL = auto() RANDOM = auto() QUASI_RANDOM = auto()
ORDER_TYPE = Union[ TraversalOrder, Literal[OrderOption.SEQUENTIAL, OrderOption.RANDOM] ] ORDER_MAP: Mapping[ORDER_TYPE, TraversalOrder] = { OrderOption.RANDOM: Random, OrderOption.SEQUENTIAL: Sequential, OrderOption.QUASI_RANDOM: QuasiRandom } DEFAULT_PROCESS_CACHE = int(environ.get('FFCV_DEFAULT_CACHE_PROCESS', "0")) DEFAULT_OS_CACHE = not DEFAULT_PROCESS_CACHE
[docs]class Loader: """FFCV loader class that can be used as a drop-in replacement for standard (e.g. PyTorch) data loaders. Parameters ---------- fname: str Full path to the location of the dataset (.beton file format). batch_size : int Batch size. num_workers : int Number of workers used for data loading. Consider using the actual number of cores instead of the number of threads if you only use JITed augmentations as they usually don't benefit from hyper-threading. os_cache : bool Leverages the operating for caching purposes. This is beneficial when there is enough memory to cache the dataset and/or when multiple processes on the same machine training using the same dataset. See https://docs.ffcv.io/performance_guide.html for more information. order : OrderOption Traversal order, one of: SEQEUNTIAL, RANDOM, QUASI_RANDOM QUASI_RANDOM is a random order that tries to be as uniform as possible while minimizing the amount of data read from the disk. Note that it is mostly useful when `os_cache=False`. Currently unavailable in distributed mode. distributed : bool For distributed training (multiple GPUs). Emulates the behavior of DistributedSampler from PyTorch. seed : int Random seed for batch ordering. indices : Sequence[int] Subset of dataset by filtering only some indices. pipelines : Mapping[str, Sequence[Union[Operation, torch.nn.Module]] Dictionary defining for each field the sequence of Decoders and transforms to apply. Fileds with missing entries will use the default pipeline, which consists of the default decoder and `ToTensor()`, but a field can also be disabled by explicitly by passing `None` as its pipeline. custom_fields : Mapping[str, Field] Dictonary informing the loader of the types associated to fields that are using a custom type. drop_last : bool Drop non-full batch in each iteration. batches_ahead : int Number of batches prepared in advance; balances latency and memory. recompile : bool Recompile every iteration. This is necessary if the implementation of some augmentations are expected to change during training. """ def __init__(self, fname: str, batch_size: int, num_workers: int = -1, os_cache: bool = DEFAULT_OS_CACHE, order: ORDER_TYPE = OrderOption.SEQUENTIAL, distributed: bool = False, seed: int = None, # For ordering of samples indices: Sequence[int] = None, # For subset selection pipelines: Mapping[str, Sequence[Union[Operation, ch.nn.Module]]] = {}, custom_fields: Mapping[str, Type[Field]] = {}, drop_last: bool = True, batches_ahead: int = 3, recompile: bool = False, # Recompile at every epoch ): if distributed and order == OrderOption.RANDOM and (seed is None): print('Warning: no ordering seed was specified with distributed=True. ' 'Setting seed to 0 to match PyTorch distributed sampler.') seed = 0 elif seed is None: tinfo = np.iinfo('int32') seed = np.random.randint(0, tinfo.max) # We store the original user arguments to be able to pass it to the # filtered version of the datasets self._args = { 'fname': fname, 'batch_size': batch_size, 'num_workers': num_workers, 'os_cache': os_cache, 'order': order, 'distributed': distributed, 'seed': seed, 'indices': indices, 'pipelines': pipelines, 'drop_last': drop_last, 'batches_ahead': batches_ahead, 'recompile': recompile } self.fname: str = fname self.batch_size: int = batch_size self.batches_ahead = batches_ahead self.seed: int = seed self.reader: Reader = Reader(self.fname, custom_fields) self.num_workers: int = num_workers self.drop_last: bool = drop_last self.distributed: bool = distributed self.code_per_stage = None self.recompile = recompile if self.num_workers < 1: self.num_workers = cpu_count() Compiler.set_num_threads(self.num_workers) if indices is None: self.indices = np.arange(self.reader.num_samples, dtype='uint64') else: self.indices = np.array(indices) if os_cache: self.memory_manager: MemoryManager = OSCacheManager(self.reader) else: self.memory_manager: MemoryManager = ProcessCacheManager( self.reader) self.traversal_order: TraversalOrder = ORDER_MAP[order](self) memory_read = self.memory_manager.compile_reader() self.next_epoch: int = 0 self.pipelines = {} self.field_name_to_f_ix = {} for f_ix, (field_name, field) in enumerate(self.reader.handlers.items()): self.field_name_to_f_ix[field_name] = f_ix DecoderClass = field.get_decoder_class() try: operations = pipelines[field_name] # We check if the user disabled this field if operations is None: continue if not isinstance(operations[0], DecoderClass): msg = "The first operation of the pipeline for " msg += f"'{field_name}' has to be a subclass of " msg += f"{DecoderClass}" raise ValueError(msg) except KeyError: try: operations = [ DecoderClass(), ToTensor() ] except Exception: msg = f"Impossible to create a default pipeline" msg += f"{field_name}, please define one manually" raise ValueError(msg) for i, op in enumerate(operations): assert isinstance(op, (ch.nn.Module, Operation)), op if isinstance(op, ch.nn.Module): operations[i] = ModuleWrapper(op) for op in operations: op.accept_field(field) op.accept_globals(self.reader.metadata[f'f{f_ix}'], memory_read) self.pipelines[field_name] = Pipeline(operations)
[docs] def next_traversal_order(self): return self.traversal_order.sample_order(self.next_epoch)
def __iter__(self): Compiler.set_num_threads(self.num_workers) order = self.next_traversal_order() selected_order = order[:len(self) * self.batch_size] self.next_epoch += 1 # Compile at the first epoch if self.code_per_stage is None or self.recompile: self.generate_code() return EpochIterator(self, selected_order)
[docs] def filter(self, field_name:str, condition: Callable[[Any], bool]) -> 'Loader': new_args = {**self._args} pipelines = {} # Disabling all the other fields for other_field_name in self.reader.handlers.keys(): pipelines[other_field_name] = None # We reuse the original pipeline for the field we care about try: pipelines[field_name] = new_args['pipelines'][field_name] except KeyError: # We keep the default one if the user didn't setup a custom one del pipelines[field_name] pass new_args['pipelines'] = pipelines # We use sequential order for speed and to know which index we are # filtering new_args['order'] = OrderOption.SEQUENTIAL new_args['drop_last'] = False sub_loader = Loader(**new_args) selected_indices = [] # Iterate through the loader and test the user defined condition for i, (batch,) in enumerate(sub_loader): for j, sample in enumerate(batch): sample_id = i * self.batch_size + j if condition(sample): selected_indices.append(sample_id) final_args = {**self._args} final_args['indices'] = np.array(selected_indices) return Loader(**final_args)
def __len__(self): next_order = self.next_traversal_order() if self.drop_last: return len(next_order) // self.batch_size else: return int(np.ceil(len(next_order) / self.batch_size))
[docs] def generate_function_call(self, pipeline_name, op_id, needs_indices): p_ix = self.field_name_to_f_ix[pipeline_name] pipeline_identifier = f'code_{pipeline_name}_{op_id}' memory_identifier = f'memory_{pipeline_name}_{op_id}' result_identifier = f'result_{pipeline_name}' arg_id = result_identifier # This is the decoder so we pass the indices instead of the previous # result if op_id == 0: arg_id = 'batch_indices' tree = ast.parse(f""" {result_identifier} = {pipeline_identifier}({arg_id}, {memory_identifier}) """).body[0] # This is the first call of the pipeline, we pass the metadata and # storage state if op_id == 0: tree.value.args.extend([ ast.Subscript(value=ast.Name(id='metadata', ctx=ast.Load()), slice=ast.Index(value=ast.Constant(value=f'f{p_ix}', kind=None)), ctx=ast.Load()), ast.Name(id='storage_state', ctx=ast.Load()), ]) if needs_indices: tree.value.args.extend([ ast.Name(id='batch_indices', ctx=ast.Load()), ]) return tree
[docs] def generate_stage_code(self, stage, stage_ix, functions): fun_name = f'stage_{stage_ix}' base_code = ast.parse(f""" def {fun_name}(): pass """).body[0] function_calls = [] memory_banks = [] memory_banks_id = [] for p_ix, pipeline_name, op_id, needs_indices in stage: function_calls.append(self.generate_function_call(pipeline_name, op_id, needs_indices)) arg = ast.arg(arg=f'memory_{pipeline_name}_{op_id}') memory_banks.append(arg) memory_banks_id.append((pipeline_name, op_id)) base_code.body.pop() base_code.body.extend(function_calls) return_tuple = ast.Return(value=ast.Tuple(elts=[], ctx=ast.Load())) base_code.args.args.append(ast.arg(arg='batch_indices')) for p_id in self.pipelines.keys(): r = f'result_{p_id}' if stage_ix != 0: base_code.args.args.append(ast.arg(arg=r)) return_tuple.value.elts.append(ast.Name(id=r, ctx=ast.Load())) base_code.body.append(return_tuple) base_code.args.args.extend(memory_banks) base_code.args.args.append(ast.arg(arg='metadata')) base_code.args.args.append(ast.arg(arg='storage_state')) module = ast.fix_missing_locations( ast.Module(body=[base_code], type_ignores=[]) ) namespace = { **functions, } exec(compile(module, '', 'exec'), namespace) final_code = namespace[fun_name] if stage_ix % 2 == 0: final_code = Compiler.compile(final_code) return final_code, memory_banks_id
[docs] def generate_code(self): schedule = defaultdict(lambda: []) compiled_functions = {} for p_ix, (p_id, p) in enumerate(self.pipelines.items()): stage = 0 for jitted_block, block_content in p.operation_blocks: # Even stages are jitted Odds are not # If this doesn't match for this pipeline we # shift the operations if 1 - jitted_block % 2 != stage % 2: stage += 1 for op in block_content: ops_code = p.compiled_ops[op] needs_indices = False if hasattr(ops_code, 'with_indices'): needs_indices = ops_code.with_indices if stage % 2 == 0: ops_code = Compiler.compile(ops_code) compiled_functions[f'code_{p_id}_{op}'] = ops_code schedule[stage].append((p_ix, p_id, op, needs_indices)) stage += 1 memory_bank_keys_per_stage = {} self.code_per_stage = {} for stage_ix, stage in schedule.items(): code_for_stage, mem_banks_ids = self.generate_stage_code(stage, stage_ix, compiled_functions) self.code_per_stage[stage_ix] = code_for_stage memory_bank_keys_per_stage[stage_ix] = mem_banks_ids self.memory_bank_keys_per_stage = memory_bank_keys_per_stage