def get_local_path(uri: str,
download_dir: str,
fs: Optional[FileSystem] = None) -> str:
"""Return the path where a local copy of URI should be stored.
If URI is local, return it. If it's remote, we generate a path for it
within download_dir.
Args:
uri: the URI of the file to be copied
download_dir: path of the local directory in which files should
be copied
fs: if supplied, use fs instead of automatically chosen FileSystem for
URI
Returns:
a local path
"""
if uri is None:
return None
if not fs:
fs = FileSystem.get_file_system(uri, 'r')
path = fs.local_path(uri, download_dir)
return path
def upload_or_copy(src_path: str,
dst_uri: str,
fs: Optional[FileSystem] = None) -> List[str]:
"""Upload or copy a file.
If dst_uri is local, the file is copied. Otherwise, it is uploaded.
Args:
src_path: path to source file
dst_uri: URI of destination for file
fs: if supplied, use fs instead of automatically chosen FileSystem for
dst_uri
Raises:
NotWritableError if dst_uri cannot be written to
"""
if dst_uri is None:
return
if not (os.path.isfile(src_path) or os.path.isdir(src_path)):
raise Exception('{} does not exist.'.format(src_path))
if not src_path == dst_uri:
log.info('Uploading {} to {}'.format(src_path, dst_uri))
if not fs:
fs = FileSystem.get_file_system(dst_uri, 'w')
fs.copy_to(src_path, dst_uri)
def download_if_needed(uri: str,
download_dir: str,
fs: Optional[FileSystem] = None) -> str:
"""Download a file into a directory if it's remote.
If uri is local, there is no need to download the file.
Args:
uri: URI of file
download_dir: local directory to download file into
fs: if supplied, use fs instead of automatically chosen FileSystem for
uri
Returns:
path to local file
Raises:
NotReadableError if URI cannot be read from
"""
if uri is None:
return None
if not fs:
fs = FileSystem.get_file_system(uri, 'r')
path = get_local_path(uri, download_dir, fs=fs)
make_dir(path, use_dirname=True)
if path != uri:
log.debug('Downloading {} to {}'.format(uri, path))
fs.copy_from(uri, path)
return path
def test_sync_from_dir_noop_local(self):
path = os.path.join(self.tmp_dir.name, 'lorem', 'ipsum.txt')
src = os.path.join(self.tmp_dir.name, 'lorem')
make_dir(src, check_empty=False)
fs = FileSystem.get_file_system(src, 'r')
fs.write_bytes(path, bytes([0x00, 0x01]))
sync_from_dir(src, src, delete=True)
self.assertEqual(len(list_paths(src)), 1)
def test_last_modified(self):
path = os.path.join(self.tmp_dir.name, 'lorem', 'ipsum1.txt')
directory = os.path.dirname(path)
make_dir(directory, check_empty=False)
fs = FileSystem.get_file_system(path, 'r')
str_to_file(self.lorem, path)
stamp = fs.last_modified(path)
self.assertTrue(isinstance(stamp, datetime.datetime))
def file_exists(uri, fs=None, include_dir=True) -> bool:
"""Check if file exists.
Args:
uri: URI of file
fs: if supplied, use fs instead of automatically chosen FileSystem for
uri
"""
if not fs:
fs = FileSystem.get_file_system(uri, 'r')
return fs.file_exists(uri, include_dir)
def test_bytes_local(self):
path = os.path.join(self.tmp_dir.name, 'lorem', 'ipsum.txt')
directory = os.path.dirname(path)
make_dir(directory, check_empty=False)
expected = bytes([0x00, 0x01, 0x02])
fs = FileSystem.get_file_system(path, 'r')
fs.write_bytes(path, expected)
actual = fs.read_bytes(path)
self.assertEqual(actual, expected)
def test_sync_to_dir_local(self):
path = os.path.join(self.tmp_dir.name, 'lorem', 'ipsum.txt')
src = os.path.dirname(path)
dst = os.path.join(self.tmp_dir.name, 'xxx')
make_dir(src, check_empty=False)
make_dir(dst, check_empty=False)
fs = FileSystem.get_file_system(path, 'r')
fs.write_bytes(path, bytes([0x00, 0x01]))
sync_to_dir(src, dst, delete=True)
self.assertEqual(len(list_paths(dst)), 1)
def test_last_modified_s3(self):
path = os.path.join(self.tmp_dir.name, 'lorem', 'ipsum1.txt')
s3_path = 's3://{}/lorem1.txt'.format(self.bucket_name)
directory = os.path.dirname(path)
make_dir(directory, check_empty=False)
fs = FileSystem.get_file_system(s3_path, 'r')
str_to_file(self.lorem, path)
upload_or_copy(path, s3_path)
stamp = fs.last_modified(s3_path)
self.assertTrue(isinstance(stamp, datetime.datetime))
def str_to_file(content_str: str, uri: str, fs: Optional[FileSystem] = None):
"""Writes string to text file.
Args:
content_str: string to write
uri: URI of file to write
fs: if supplied, use fs instead of automatically chosen FileSystem
Raise:
NotWritableError if uri cannot be written
"""
if not fs:
fs = FileSystem.get_file_system(uri, 'r')
return fs.write_str(uri, content_str)
def test_file_exists(self):
fs = FileSystem.get_file_system(self.tmp_dir.name, 'r')
path1 = os.path.join(self.tmp_dir.name, 'lorem', 'ipsum.txt')
dir1 = os.path.dirname(path1)
make_dir(dir1, check_empty=False)
str_to_file(self.lorem, path1)
self.assertTrue(fs.file_exists(dir1, include_dir=True))
self.assertTrue(fs.file_exists(path1, include_dir=False))
self.assertFalse(fs.file_exists(dir1, include_dir=False))
self.assertFalse(
fs.file_exists(dir1 + 'NOTPOSSIBLE', include_dir=False))
def file_to_str(uri: str, fs: Optional[FileSystem] = None) -> str:
"""Load contents of text file into a string.
Args:
uri: URI of file
fs: if supplied, use fs instead of automatically chosen FileSystem
Returns:
contents of text file
Raises:
NotReadableError if URI cannot be read
"""
if not fs:
fs = FileSystem.get_file_system(uri, 'r')
return fs.read_str(uri)
def list_paths(uri: str, ext: str = '',
fs: Optional[FileSystem] = None) -> List[str]:
"""List paths rooted at URI.
Optionally only includes paths with a certain file extension.
Args:
uri: the URI of a directory
ext: the optional file extension to filter by
fs: if supplied, use fs instead of automatically chosen FileSystem for
uri
"""
if uri is None:
return None
if not fs:
fs = FileSystem.get_file_system(uri, 'r')
return fs.list_paths(uri, ext=ext)
def sync_from_dir(src_dir_uri: str,
dst_dir: str,
delete: bool = False,
fs: Optional[FileSystem] = None):
"""Synchronize a source directory to local destination directory.
Transfers files from source to destination directories so that the
destination has all the source files. If FileSystem is remote, this involves
downloading.
Args:
src_dir_uri: URI of source directory
dst_dir: path of local destination directory
delete: if True, delete files in the destination to match those in the
source directory
fs: if supplied, use fs instead of automatically chosen FileSystem for
dst_dir_uri
"""
if not fs:
fs = FileSystem.get_file_system(src_dir_uri, 'r')
fs.sync_from_dir(src_dir_uri, dst_dir, delete=delete)
def __init__(self,
cfg: LearnerConfig,
tmp_dir: str,
model_path: Optional[str] = None,
model_def_path: Optional[str] = None,
loss_def_path: Optional[str] = None,
training: bool = True):
"""Constructor.
Args:
cfg (LearnerConfig): Configuration.
tmp_dir (str): Root of temp dirs.
model_path (str, optional): A local path to model weights.
Defaults to None.
model_def_path (str, optional): A local path to a directory with a
hubconf.py. If provided, the model definition is imported from
here. Defaults to None.
loss_def_path (str, optional): A local path to a directory with a
hubconf.py. If provided, the loss function definition is
imported from here. Defaults to None.
training (bool, optional): Whether the model is to be used for
training or prediction. If False, the model is put in eval mode
and the loss function, optimizer, etc. are not initialized.
Defaults to True.
"""
log_system_details()
self.cfg = cfg
self.tmp_dir = tmp_dir
self.preview_batch_limit = self.cfg.data.preview_batch_limit
# TODO make cache dirs configurable
torch_cache_dir = '/opt/data/torch-cache'
os.environ['TORCH_HOME'] = torch_cache_dir
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.data_cache_dir = '/opt/data/data-cache'
make_dir(self.data_cache_dir)
if FileSystem.get_file_system(cfg.output_uri) == LocalFileSystem:
self.output_dir = cfg.output_uri
make_dir(self.output_dir)
else:
self.output_dir = get_local_path(cfg.output_uri, tmp_dir)
make_dir(self.output_dir, force_empty=True)
if training and not cfg.overfit_mode:
self.sync_from_cloud()
self.modules_dir = join(self.output_dir, MODULES_DIRNAME)
self.setup_model(model_def_path=model_def_path)
if model_path is not None:
if isfile(model_path):
log.info(f'Loading model weights from: {model_path}')
self.model.load_state_dict(
torch.load(model_path, map_location=self.device))
else:
raise Exception(
'Model could not be found at {}'.format(model_path))
if training:
self.setup_training(loss_def_path=loss_def_path)
else:
self.model.eval()
def test_bytes_local_false(self):
path = os.path.join(self.tmp_dir.name, 'xxx')
fs = FileSystem.get_file_system(path, 'r')
self.assertRaises(NotReadableError, lambda: fs.read_bytes(path))
def test_last_modified_http(self):
uri = 'http://localhost/'
fs = FileSystem.get_file_system(uri, 'r')
self.assertEqual(fs.last_modified(uri), None)
def test_write_bytes_http(self):
uri = 'http://localhost/'
fs = FileSystem.get_file_system(uri, 'r')
self.assertRaises(NotWritableError,
lambda: fs.write_bytes(uri, bytes([0x00, 0x01])))
def __init__(self,
cfg: LearnerConfig,
tmp_dir: str,
model_path: Optional[str] = None,
model_def_path: Optional[str] = None,
loss_def_path: Optional[str] = None):
"""Constructor.
Args:
cfg: configuration
tmp_dir: root of temp dirs
model_path: a local path to model weights. If provided, the model is loaded
and it is assumed that this Learner will be used for prediction only.
model_def_path: a local path to a directory with a hubconf.py. If
provided, the model definition is imported from here.
loss_def_path: a local path to a directory with a hubconf.py. If
provided, the loss function definition is imported from here.
"""
self.cfg = cfg
self.tmp_dir = tmp_dir
# TODO make cache dirs configurable
torch_cache_dir = '/opt/data/torch-cache'
os.environ['TORCH_HOME'] = torch_cache_dir
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.data_cache_dir = '/opt/data/data-cache'
make_dir(self.data_cache_dir)
if FileSystem.get_file_system(cfg.output_uri) == LocalFileSystem:
self.output_dir = cfg.output_uri
make_dir(self.output_dir)
else:
self.output_dir = get_local_path(cfg.output_uri, tmp_dir)
make_dir(self.output_dir, force_empty=True)
if not cfg.overfit_mode:
self.sync_from_cloud()
self.modules_dir = join(self.output_dir, MODULES_DIRNAME)
self.setup_model(model_def_path=model_def_path)
if model_path is not None:
if isfile(model_path):
self.model.load_state_dict(
torch.load(model_path, map_location=self.device))
else:
raise Exception(
'Model could not be found at {}'.format(model_path))
self.model.eval()
else:
log.info(self.cfg)
# ds = dataset, dl = dataloader
self.train_ds = None
self.train_dl = None
self.valid_ds = None
self.valid_dl = None
self.test_ds = None
self.test_dl = None
self.config_path = join(self.output_dir, 'learner-config.json')
str_to_file(self.cfg.json(), self.config_path)
self.log_path = join(self.output_dir, 'log.csv')
self.train_state_path = join(self.output_dir, 'train-state.json')
model_bundle_fname = basename(cfg.get_model_bundle_uri())
self.model_bundle_path = join(self.output_dir, model_bundle_fname)
self.metric_names = self.build_metric_names()
self.last_model_path = join(self.output_dir, 'last-model.pth')
self.load_checkpoint()
self.setup_loss(loss_def_path=loss_def_path)
self.opt = self.build_optimizer()
self.setup_data()
self.start_epoch = self.get_start_epoch()
self.steps_per_epoch = len(
self.train_ds) // self.cfg.solver.batch_sz
self.step_scheduler = self.build_step_scheduler()
self.epoch_scheduler = self.build_epoch_scheduler()
self.setup_tensorboard()
