def _get_process_split(split, process_index, process_count, drop_remainder):
"""Returns the split for the given process given a multi-process setup."""
splits = tfds.even_splits(split,
n=process_count,
drop_remainder=drop_remainder)
process_split = splits[process_index]
return process_split
def make_source_dataset(self, current_host_index, num_hosts):
"""Makes a dataset of dictionaries of images and labels.
Args:
current_host_index: current host index.
num_hosts: total number of hosts.
Returns:
A `tf.data.Dataset` object where each dataset element is a dictionary. For
image classification datasets, the dictionary will contain an 'image' key
with a decoded uint8 image (of shape [height, width, channels]) and a
'label' key with an int64 label. For video classification datasets, the
dictionary will contain a 'video' key with a decoded uint8 video (of shape
[frames, height, width, channels]) and a 'label' key with an int64 label.
"""
split = self.split
if self.mode == enums.ModelMode.TRAIN and self.shard_per_host:
split = tfds.even_splits(split, n=num_hosts)[current_host_index]
# Don't shuffle until after sharding, since otherwise you risk dropping
# samples because the sharding is performed on different shufflings of the
# data for each core.
return tfds.load(name=self.dataset_name,
split=split,
data_dir=self.data_dir,
shuffle_files=False)
def main(_):
log('Loading "coco_captions" from tfds. This will build a copy of '
'the dataset under TFDS_DATA_DIR if one doesn\'t already exist.')
tfds.load('coco_captions')
# Milan split name => constituent tfds subsplits. (Subsplitting enables the
# tfds data to be processed in parallel.)
split_to_coco_subsplits = dict(train=tfds.even_splits('train', 100) +
tfds.even_splits('restval', 100),
dev=tfds.even_splits('val', 50),
test=tfds.even_splits('test', 50))
encoder_params = TfHubBertEncoder.Params().Set(
preprocessor=FLAGS.bert_tfhub_preprocessor,
model=FLAGS.bert_tfhub_model,
max_seq_len=FLAGS.bert_max_length)
def pipeline(root):
for split in FLAGS.splits:
subsplits = split_to_coco_subsplits[split]
_ = root | split.title() >> (
'Create' >> beam.Create(subsplits)
| 'Read' >> beam.ParDo(ReadCocoFromTfdsFn())
| 'Encode' >> beam.ParDo(EncodeCaptionsFn(encoder_params))
| 'ToProto' >> beam.Map(dict_to_example_proto)
| 'Shuffle' >> beam.Reshuffle()
| 'Write' >> beam.io.tfrecordio.WriteToTFRecord(
file_path_prefix=os.path.join(FLAGS.output_dir, split),
coder=beam.coders.ProtoCoder(tf.train.Example),
num_shards=FLAGS.num_output_shards))
pipeline_options = beam.options.pipeline_options.PipelineOptions()
pipeline_options.view_as(beam.options.pipeline_options.DirectOptions
).direct_num_workers = FLAGS.num_workers
with beam.Pipeline(options=pipeline_options) as root:
pipeline(root)
"""### MAIN CELL: Experimenting a range of hyperparameters"""
# at = [0.2, 0.9]
# ll1 = [1e-3]
# ll2 = [1e-3, 1e-4]
# all = []
at = [0.0, 0.2, 0.4, 0.5, 0.7, 0.9]
ll1 = [1e-1, 1e-2, 1e-3]
ll2 = [1e-3, 1e-4, 1e-5]
all = []
n_envs = 2
ds_train_envs = []
batch_size = 10000
splits = tfds.even_splits('train', n=n_envs)
for m in range(n_envs):
ds = tfds.load("mnist:3.*.*", split=splits[m]).cache().repeat()
ds = ds.shuffle(10 * batch_size, seed=0)
ds = ds.batch(batch_size)
ds_train_envs.append(iter(tfds.as_numpy(ds)))
test_ds = tfds.load("mnist:3.*.*", split='test').cache().repeat()
test_ds = test_ds.shuffle(10 * 10000, seed=0)
test_ds = test_ds.batch(10000)
test_ds= iter(tfds.as_numpy(test_ds))
round = 0
for idx, thresh in enumerate(at):
for l1 in ll1:
for l2 in ll2:
def _lazy_init(self):
""" Lazily initialize the dataset.
This is necessary to init the Tensorflow dataset pipeline in the (dataloader) process that
will be using the dataset instance. The __init__ method is called on the main process,
this will be called in a dataloader worker process.
NOTE: There will be problems if you try to re-use this dataset across different loader/worker
instances once it has been initialized. Do not call any dataset methods that can call _lazy_init
before it is passed to dataloader.
"""
worker_info = torch.utils.data.get_worker_info()
# setup input context to split dataset across distributed processes
num_workers = 1
global_worker_id = 0
if worker_info is not None:
self.worker_info = worker_info
self.worker_seed = worker_info.seed
num_workers = worker_info.num_workers
self.global_num_workers = self.dist_num_replicas * num_workers
global_worker_id = self.dist_rank * num_workers + worker_info.id
""" Data sharding
InputContext will assign subset of underlying TFRecord files to each 'pipeline' if used.
My understanding is that using split, the underling TFRecord files will shuffle (shuffle_files=True)
between the splits each iteration, but that understanding could be wrong.
I am currently using a mix of InputContext shard assignment and fine-grained sub-splits for distributing
the data across workers. For training InputContext is used to assign shards to nodes unless num_shards
in dataset < total number of workers. Otherwise sub-split API is used for datasets without enough shards or
for validation where we can't drop examples and need to avoid minimize uneven splits to avoid padding.
"""
should_subsplit = self.global_num_workers > 1 and (
self.split_info.num_shards < self.global_num_workers
or not self.is_training)
if should_subsplit:
# split the dataset w/o using sharding for more even examples / worker, can result in less optimal
# read patterns for distributed training (overlap across shards) so better to use InputContext there
if has_buggy_even_splits:
# my even_split workaround doesn't work on subsplits, upgrade tfds!
if not isinstance(self.split_info,
tfds.core.splits.SubSplitInfo):
subsplits = even_split_indices(self.split,
self.global_num_workers,
self.num_examples)
self.subsplit = subsplits[global_worker_id]
else:
subsplits = tfds.even_splits(self.split,
self.global_num_workers)
self.subsplit = subsplits[global_worker_id]
input_context = None
if self.global_num_workers > 1 and self.subsplit is None:
# set input context to divide shards among distributed replicas
input_context = tf.distribute.InputContext(
num_input_pipelines=self.global_num_workers,
input_pipeline_id=global_worker_id,
num_replicas_in_sync=self.
dist_num_replicas # FIXME does this arg have any impact?
)
read_config = tfds.ReadConfig(shuffle_seed=self.common_seed,
shuffle_reshuffle_each_iteration=True,
input_context=input_context)
ds = self.builder.as_dataset(split=self.subsplit or self.split,
shuffle_files=self.is_training,
read_config=read_config)
# avoid overloading threading w/ combo of TF ds threads + PyTorch workers
options = tf.data.Options()
thread_member = 'threading' if hasattr(
options, 'threading') else 'experimental_threading'
getattr(options, thread_member).private_threadpool_size = max(
1, self.max_threadpool_size // num_workers)
getattr(options, thread_member).max_intra_op_parallelism = 1
ds = ds.with_options(options)
if self.is_training or self.repeats > 1:
# to prevent excessive drop_last batch behaviour w/ IterableDatasets
# see warnings at https://pytorch.org/docs/stable/data.html#multi-process-data-loading
ds = ds.repeat() # allow wrap around and break iteration manually
if self.is_training:
ds = ds.shuffle(min(self.num_examples, self.shuffle_size) //
self.global_num_workers,
seed=self.worker_seed)
ds = ds.prefetch(
min(self.num_examples // self.global_num_workers,
self.prefetch_size))
self.ds = tfds.as_numpy(ds)
https://www.tensorflow.org/datasets/catalog/overview#image_classification
Hacked together by / Copyright 2020 Ross Wightman
"""
import math
import torch
import torch.distributed as dist
from PIL import Image
try:
import tensorflow as tf
tf.config.set_visible_devices(
[], 'GPU') # Hands off my GPU! (or pip install tensorflow-cpu)
import tensorflow_datasets as tfds
try:
tfds.even_splits('', 1, drop_remainder=False
) # non-buggy even_splits has drop_remainder arg
has_buggy_even_splits = False
except TypeError:
print(
"Warning: This version of tfds doesn't have the latest even_splits impl. "
"Please update or use tfds-nightly for better fine-grained split behaviour."
)
has_buggy_even_splits = True
# NOTE uncomment below if having file limit issues on dataset build (or alter your OS defaults)
# import resource
# low, high = resource.getrlimit(resource.RLIMIT_NOFILE)
# resource.setrlimit(resource.RLIMIT_NOFILE, (high, high))
except ImportError as e:
print(e)
print(
"Please install tensorflow_datasets package `pip install tensorflow-datasets`."
make_plot = True
vote_threshold = 0.5
width = 32
height = 32
n_classes = 10
optimizer = tf.keras.optimizers.Adam
l_rate = 0.001
batch_size = 64
loss = "sparse_categorical_crossentropy"
vote_batches = 2
train_datasets, info = tfds.load('cifar10',
split=tfds.even_splits('train', n=n_learners),
as_supervised=True, with_info=True)
n_datapoints = info.splits['train'].num_examples
test_datasets = tfds.load('cifar10',
split=tfds.even_splits('test', n=n_learners),
as_supervised=True)
for i in range(n_learners):
ds_train = train_datasets[i].map(
normalize_img, num_parallel_calls=tf.data.experimental.AUTOTUNE)
ds_train = ds_train.cache()
ds_train = ds_train.shuffle(n_datapoints // n_learners)
ds_train = ds_train.batch(batch_size)
train_datasets[i] = ds_train.prefetch(tf.data.experimental.AUTOTUNE)
