Sunday01's picture
up
9dce458
raw
history blame
13.3 kB
import os
import stat
import sys
import tempfile
import re
import torch
import shutil
import filecmp
from abc import ABC, abstractmethod
from functools import cached_property
from .generic import (
BASE_PATH,
Context,
download_url_with_progressbar,
prompt_yes_no,
replace_prefix,
get_digest,
get_filename_from_url,
)
from .log import get_logger
class InfererModule(ABC):
def __init__(self):
self.logger = get_logger(self.__class__.__name__)
super().__init__()
def parse_args(self, args: Context):
"""May be overwritten by super classes to parse commandline arguments"""
pass
# class InfererModuleManager(ABC):
# _KEY = ''
# _VARIANTS = []
# def __init__(self):
# self.onstart: Callable = None
# self.onfinish: Callable = None
# def validate(self):
# """
# Throws exception if a
# """
# ...
# async def prepare(self):
# ...
# async def dispatch(self):
# ...
class ModelVerificationException(Exception):
pass
class InvalidModelMappingException(ValueError):
def __init__(self, cls: str, map_key: str, error_msg: str):
error = f'[{cls}->{map_key}] Invalid _MODEL_MAPPING - {error_msg}'
super().__init__(error)
class ModelWrapper(ABC):
r"""
A class that provides a unified interface for downloading models and making forward passes.
All model inferer classes should extend it.
Download specifications can be made through overwriting the `_MODEL_MAPPING` property.
```python
_MODEL_MAPPTING = {
'model_id': {
**PARAMETERS
},
...
}
```
Parameters:
model_id - Used for temporary caches and debug messages
url - A direct download url
hash - Hash of downloaded file, Can be obtained upon ModelVerificationException
file - File download destination, If set to '.' the filename will be inferred
from the url (fallback is `model_id` value)
archive - Dict that contains all files/folders that are to be extracted from
the downloaded archive and their destinations, Mutually exclusive with `file`
executables - List of files that need to have the executable flag set
"""
_MODEL_DIR = os.path.join(BASE_PATH, 'models')
_MODEL_SUB_DIR = ''
_MODEL_MAPPING = {}
_KEY = ''
def __init__(self):
os.makedirs(self.model_dir, exist_ok=True)
self._key = self._KEY or self.__class__.__name__
self._loaded = False
self._check_for_malformed_model_mapping()
self._downloaded = self._check_downloaded()
def is_loaded(self) -> bool:
return self._loaded
def is_downloaded(self) -> bool:
return self._downloaded
@property
def model_dir(self):
return os.path.join(self._MODEL_DIR, self._MODEL_SUB_DIR)
def _get_file_path(self, *args) -> str:
return os.path.join(self.model_dir, *args)
def _get_used_gpu_memory(self) -> bool:
'''
Gets the total amount of GPU memory used by model (Can be used in the future
to determine whether a model should be loaded into vram or ram or automatically choose a model size).
TODO: Use together with `--use-cuda-limited` flag to enforce stricter memory checks
'''
return torch.cuda.mem_get_info()
def _check_for_malformed_model_mapping(self):
for map_key, mapping in self._MODEL_MAPPING.items():
if 'url' not in mapping:
raise InvalidModelMappingException(self._key, map_key, 'Missing url property')
elif not re.search(r'^https?://', mapping['url']):
raise InvalidModelMappingException(self._key, map_key, 'Malformed url property: "%s"' % mapping['url'])
if 'file' not in mapping and 'archive' not in mapping:
mapping['file'] = '.'
elif 'file' in mapping and 'archive' in mapping:
raise InvalidModelMappingException(self._key, map_key, 'Properties file and archive are mutually exclusive')
async def _download_file(self, url: str, path: str):
print(f' -- Downloading: "{url}"')
download_url_with_progressbar(url, path)
async def _verify_file(self, sha256_pre_calculated: str, path: str):
print(f' -- Verifying: "{path}"')
sha256_calculated = get_digest(path).lower()
sha256_pre_calculated = sha256_pre_calculated.lower()
if sha256_calculated != sha256_pre_calculated:
self._on_verify_failure(sha256_calculated, sha256_pre_calculated)
else:
print(' -- Verifying: OK!')
def _on_verify_failure(self, sha256_calculated: str, sha256_pre_calculated: str):
print(f' -- Mismatch between downloaded and created hash: "{sha256_calculated}" <-> "{sha256_pre_calculated}"')
raise ModelVerificationException()
@cached_property
def _temp_working_directory(self):
p = os.path.join(tempfile.gettempdir(), 'manga-image-translator', self._key.lower())
os.makedirs(p, exist_ok=True)
return p
async def download(self, force=False):
'''
Downloads required models.
'''
if force or not self.is_downloaded():
while True:
try:
await self._download()
self._downloaded = True
break
except ModelVerificationException:
if not prompt_yes_no('Failed to verify signature. Do you want to restart the download?', default=True):
print('Aborting.', end='')
raise KeyboardInterrupt()
async def _download(self):
'''
Downloads models as defined in `_MODEL_MAPPING`. Can be overwritten (together
with `_check_downloaded`) to implement unconventional download logic.
'''
print(f'\nDownloading models into {self.model_dir}\n')
for map_key, mapping in self._MODEL_MAPPING.items():
if self._check_downloaded_map(map_key):
print(f' -- Skipping {map_key} as it\'s already downloaded')
continue
is_archive = 'archive' in mapping
if is_archive:
download_path = os.path.join(self._temp_working_directory, map_key, '')
else:
download_path = self._get_file_path(mapping['file'])
if not os.path.basename(download_path):
os.makedirs(download_path, exist_ok=True)
if os.path.basename(download_path) in ('', '.'):
download_path = os.path.join(download_path, get_filename_from_url(mapping['url'], map_key))
if not is_archive:
download_path += '.part'
if 'hash' in mapping:
downloaded = False
if os.path.isfile(download_path):
try:
print(' -- Found existing file')
await self._verify_file(mapping['hash'], download_path)
downloaded = True
except ModelVerificationException:
print(' -- Resuming interrupted download')
if not downloaded:
await self._download_file(mapping['url'], download_path)
await self._verify_file(mapping['hash'], download_path)
else:
await self._download_file(mapping['url'], download_path)
if download_path.endswith('.part'):
p = download_path[:len(download_path)-5]
shutil.move(download_path, p)
download_path = p
if is_archive:
extracted_path = os.path.join(os.path.dirname(download_path), 'extracted')
print(f' -- Extracting files')
shutil.unpack_archive(download_path, extracted_path)
def get_real_archive_files():
archive_files = []
for root, dirs, files in os.walk(extracted_path):
for name in files:
file_path = replace_prefix(os.path.join(root, name), extracted_path, '')
archive_files.append(file_path)
return archive_files
# Move every specified file from archive to destination
for orig, dest in mapping['archive'].items():
p1 = os.path.join(extracted_path, orig)
if os.path.exists(p1):
p2 = self._get_file_path(dest)
if os.path.basename(p2) in ('', '.'):
p2 = os.path.join(p2, os.path.basename(p1))
if os.path.isfile(p2):
if filecmp.cmp(p1, p2):
continue
raise InvalidModelMappingException(self._key, map_key, 'File "{orig}" already exists at "{dest}"')
os.makedirs(os.path.dirname(p2), exist_ok=True)
shutil.move(p1, p2)
else:
raise InvalidModelMappingException(self._key, map_key, f'File "{orig}" does not exist within archive' +
'\nAvailable files:\n%s' % '\n'.join(get_real_archive_files()))
if len(mapping['archive']) == 0:
raise InvalidModelMappingException(self._key, map_key, 'No archive files specified' +
'\nAvailable files:\n%s' % '\n'.join(get_real_archive_files()))
self._grant_execute_permissions(map_key)
# Remove temporary files
try:
os.remove(download_path)
shutil.rmtree(extracted_path)
except Exception:
pass
print()
self._on_download_finished(map_key)
def _on_download_finished(self, map_key):
'''
Can be overwritten to further process the downloaded files
'''
pass
def _check_downloaded(self) -> bool:
'''
Scans filesystem for required files as defined in `_MODEL_MAPPING`.
Returns `False` if files should be redownloaded.
'''
for map_key in self._MODEL_MAPPING:
if not self._check_downloaded_map(map_key):
return False
return True
def _check_downloaded_map(self, map_key: str) -> str:
mapping = self._MODEL_MAPPING[map_key]
if 'file' in mapping:
path = mapping['file']
if os.path.basename(path) in ('.', ''):
path = os.path.join(path, get_filename_from_url(mapping['url'], map_key))
if not os.path.exists(self._get_file_path(path)):
return False
elif 'archive' in mapping:
for orig, dest in mapping['archive'].items():
if os.path.basename(dest) in ('', '.'):
dest = os.path.join(dest, os.path.basename(orig[:-1] if orig.endswith('/') else orig))
if not os.path.exists(self._get_file_path(dest)):
return False
self._grant_execute_permissions(map_key)
return True
def _grant_execute_permissions(self, map_key: str):
mapping = self._MODEL_MAPPING[map_key]
if sys.platform == 'linux':
# Grant permission to executables
for file in mapping.get('executables', []):
p = self._get_file_path(file)
if os.path.basename(p) in ('', '.'):
p = os.path.join(p, file)
if not os.path.isfile(p):
raise InvalidModelMappingException(self._key, map_key, f'File "{file}" does not exist')
if not os.access(p, os.X_OK):
os.chmod(p, os.stat(p).st_mode | stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH)
async def reload(self, device: str, *args, **kwargs):
await self.unload()
await self.load(*args, **kwargs, device=device)
async def load(self, device: str, *args, **kwargs):
'''
Loads models into memory. Has to be called before `forward`.
'''
if not self.is_downloaded():
await self.download()
if not self.is_loaded():
await self._load(*args, **kwargs, device=device)
self._loaded = True
async def unload(self):
if self.is_loaded():
await self._unload()
self._loaded = False
async def infer(self, *args, **kwargs):
'''
Makes a forward pass through the network.
'''
if not self.is_loaded():
raise Exception(f'{self._key}: Tried to forward pass without having loaded the model.')
return await self._infer(*args, **kwargs)
@abstractmethod
async def _load(self, device: str, *args, **kwargs):
pass
@abstractmethod
async def _unload(self):
pass
@abstractmethod
async def _infer(self, *args, **kwargs):
pass