def meme_add_name_idx(name: str):
for idx, meme in enumerate(
training_db.query(MemeCorrectTrain).filter(
cast(ClauseElement, MemeCorrectTrain.name == name))):
meme.name_idx = idx # type:ignore
for idx, meme in enumerate(
training_db.query(MemeCorrectTest).filter(
cast(ClauseElement, MemeCorrectTest.name == name))):
meme.name_idx = idx # type:ignore
for idx, meme in enumerate(
training_db.query(MemeIncorrectTrain).filter(
cast(ClauseElement, MemeIncorrectTrain.name == name))):
meme.name_idx = idx # type:ignore
for idx, meme in enumerate(
training_db.query(MemeIncorrectTest).filter(
cast(ClauseElement, MemeIncorrectTest.name == name))):
meme.name_idx = idx # type:ignore
training_db.commit()
def download_imgflip_memes(fresh: bool = False):
if not training_db.query(Template).count():
raise Exception("No Templates")
for (name, ) in training_db.query(Template.name):
Path(MEMES_REPO + name).mkdir(parents=True, exist_ok=True)
if fresh:
name_page = training_db.query(Template.name, Template.page).all()
else:
names_filled = list(
name for name in os.listdir(MEMES_REPO)
if len(list(os.listdir(MEMES_REPO + name))) >= 1000)
name_page = (training_db.query(Template.name, Template.page).filter(
cast(ClauseElement, ~Template.name.in_(names_filled))).all())
with Pool(cpu_count()) as workers:
_: List[None] = list(
tqdm(
workers.imap_unordered(download_imgs_from_page, name_page),
total=len(name_page),
))
def get_assortment(self):
image_name = self.names.pop()
entity = self.correct_entity
rand = random.randint(0, self.max_name_idx["correct"][image_name])
clause = and_(
cast(ClauseElement, entity.name == image_name),
cast(ClauseElement, entity.name_idx == rand),
)
p = cast(Tuple[str],
training_db.query(entity.path).filter(clause).first())[0]
transforms = toTensorOnly if random.random() < 0.5 else self.transforms
return (transforms(Img.open(p)), 1 if image_name == self.name else 0)
def download_blanks() -> None:
q = training_db.query(Template.name, Template.blank_url).filter(
cast(ClauseElement, Template.name.in_(MEMES_TO_USE)))
path = BLANKS_REPO
shutil.rmtree(path)
Path(path).mkdir(parents=True, exist_ok=True) # type: ignore
with Pool(cpu_count()) as workers:
results: List[bool] = list(
tqdm(
workers.imap_unordered(download_img_from_url_blank,
zip(q, repeat(path))),
total=q.count(),
))
print(f"{sum(results)}/{len(results)}")
def miss_match() -> None:
downloaded_blanks = [
os.path.splitext(filename)[0] for filename in os.listdir(BLANKS_REPO)
]
display_df(
pd.read_sql(
cast(
str,
training_db.query(Template).filter(
cast(ClauseElement,
~Template.name.in_(downloaded_blanks))).statement,
),
training_db.bind,
))
def download_names_not_used():
names = [
name for name in MEMES_TO_USE if not Path(MEMES_REPO + name).is_dir()
or len(list(os.listdir(MEMES_REPO + name))) == 0
]
for name in names:
Path(MEMES_REPO + name).mkdir(parents=True, exist_ok=True)
name_page = (training_db.query(Template.name, Template.page).filter(
cast(ClauseElement, Template.name.in_(names))).all())
with Pool(cpu_count()) as workers:
_: List[None] = list(
tqdm(
workers.imap_unordered(download_imgs_from_page, name_page),
total=len(name_page),
))
def get_not_meme(self):
return (
toTensorOnly(
Img.open(
cast(
Tuple[str],
training_db.query(self.not_a_meme_entity.path).filter(
cast(
ClauseElement,
self.not_a_meme_entity.name_idx
== random.randint(
0, self.max_name_idx["not_a_meme"]),
), ).first(),
)[0])),
0,
)
def get_assortment(self):
name = self.names.pop()
rand = random.randint(0, self.max_name_idx["correct"][name])
clause = and_(
cast(ClauseElement, self.meme_entity.name == name),
cast(ClauseElement, self.meme_entity.name_idx == rand),
)
q = training_db.query(self.meme_entity.path).filter(clause)
path = cast(Tuple[str], q.first())[0]
if self.is_training and self.use_transforms == None:
transforms = toTensorOnly if random.random(
) < 0.05 else trainingTransforms
else:
if self.use_transforms:
transforms = trainingTransforms
else:
transforms = toTensorOnly
return (transforms(Img.open(path)), self.name_num[name])
def get_incorrect(self):
return (
toTensorOnly(
Img.open(
cast(
Tuple[str],
training_db.query(self.incorrect_entity.path).filter(
and_(
cast(
ClauseElement,
self.incorrect_entity.name == self.name,
),
cast(
ClauseElement,
self.incorrect_entity.name_idx ==
random.randint(0, self.my_max_incorrect),
),
)).first(),
)[0])),
0,
)
def build_template_db() -> None:
print(f"num memes to use - {len(MEMES_TO_USE)}")
_: Any = training_db.query(Template).delete()
training_db.commit()
step_size = 50
pages = range(0, step_size)
with Pool(cpu_count() // 2) as workers:
df = pd.DataFrame.from_records(
list(
chain.from_iterable(
cast(
Iterable[List[Dict[str, str]]],
tqdm(
workers.imap_unordered(get_template_data, pages),
total=step_size,
),
)))).drop_duplicates(ignore_index=True)
df.to_sql("templates",
training_db.bind,
if_exists="replace",
index_label="id")
print(f"Total Templates - {training_db.query(Template).count()}")
def build_db_from_imgdir():
_: Any = training_db.query(MemeCorrectTrain).delete()
_: Any = training_db.query(MemeCorrectTest).delete()
_: Any = training_db.query(NotAMemeTrain).delete()
_: Any = training_db.query(NotAMemeTest).delete()
_: Any = training_db.query(NotATemplateTrain).delete()
_: Any = training_db.query(NotATemplateTest).delete()
training_db.commit()
template_names = list(
cast(Set[str],
set.intersection(set(os.listdir(MEMES_REPO)), MEMES_TO_USE)))
files = list(os.listdir(NOT_MEME_REPO))
with Pool(cpu_count()) as workers:
_ = list(
tqdm(workers.imap_unordered(add_not_meme, enumerate(files)),
total=len(files)))
files = list(os.listdir(NOT_TEMPLATE_REPO))
with Pool(cpu_count()) as workers:
_ = list(
tqdm(
workers.imap_unordered(add_not_template, enumerate(files)),
total=len(files),
))
with Pool(cpu_count()) as workers:
_ = list(
tqdm(
workers.imap_unordered(name_imgs_to_db, template_names),
total=len(template_names),
))
q_list: List[Any] = [
NotAMemeTrain,
NotAMemeTest,
NotATemplateTrain,
NotATemplateTest,
]
with Pool(cpu_count()) as workers:
_ = list(
tqdm(workers.imap_unordered(add_name_idx, q_list),
total=len(q_list)))
with Pool(cpu_count()) as workers:
_ = list(
tqdm(
workers.imap_unordered(meme_add_name_idx, template_names),
total=len(template_names),
))
def get_correct(self) -> Tuple[Tensor, int]:
entity = self.correct_entity
rand = random.randint(0, self.my_max_correct)
clause = and_(
cast(ClauseElement, entity.name == self.name),
cast(ClauseElement, entity.name_idx == rand),
)
if self.is_training and self.use_transforms == None:
transforms = toTensorOnly if random.random(
) < 0.15 else trainingTransforms
else:
if self.use_transforms:
transforms = trainingTransforms
else:
transforms = toTensorOnly
return (
transforms(
Img.open(
cast(
Tuple[str],
training_db.query(entity.path).filter(clause).first(),
)[0])),
1,
)
def init_static() -> Static:
names = [
name for name, in cast(
Iterator[str],
training_db.query(MemeCorrectTrain.name).distinct(
MemeCorrectTrain.name),
)
]
names_to_shuffle = deepcopy(names)
name_num = {name: idx for idx, name in enumerate(names)}
num_name = {str(v): k for k, v in name_num.items()}
max_name_idx: TestTrainToMax = {
"train": {
"not_a_meme":
cast(
int,
training_db.query(func.max(NotAMemeTrain.name_idx)).scalar(),
),
"not_a_template":
cast(
int,
training_db.query(func.max(
NotATemplateTrain.name_idx)).scalar(),
),
"correct": {
name: cast(
int,
training_db.query(func.max(
MemeCorrectTrain.name_idx)).filter(
cast(ClauseElement,
MemeCorrectTrain.name == name)).scalar(),
)
for name in names
if name not in ["not_a_meme", "not_a_template"]
},
"incorrect": {
name: cast(
int,
training_db.query(func.max(
MemeIncorrectTrain.name_idx)).filter(
cast(ClauseElement,
MemeIncorrectTrain.name == name)).scalar(),
)
for name in names
if name not in ["not_a_meme", "not_a_template"]
},
},
"test": {
"not_a_meme":
cast(
int,
training_db.query(func.max(NotAMemeTest.name_idx)).scalar(),
),
"not_a_template":
cast(
int,
training_db.query(func.max(
NotATemplateTest.name_idx)).scalar(),
),
"correct": {
name: cast(
int,
training_db.query(func.max(
MemeCorrectTest.name_idx)).filter(
cast(ClauseElement,
MemeCorrectTest.name == name)).scalar(),
)
for name in names
if name not in ["not_a_meme", "not_a_template"]
},
"incorrect": {
name: cast(
int,
training_db.query(func.max(
MemeIncorrectTest.name_idx)).filter(
cast(ClauseElement,
MemeIncorrectTest.name == name)).scalar(),
)
for name in names
if name not in ["not_a_meme", "not_a_template"]
},
},
}
static: Static = {
"names": names,
"names_to_shuffle": names_to_shuffle,
"name_num": name_num,
"num_name": num_name,
"folder_count": {
"not_a_meme": len(os.listdir(NOT_MEME_REPO)),
"not_a_template": len(os.listdir(NOT_TEMPLATE_REPO)),
**{
name: len(os.listdir(MEMES_REPO + name))
for name in os.listdir(MEMES_REPO)
},
},
"max_name_idx": max_name_idx,
}
return static
def add_name_idx(entity: Any):
for idx, meme in enumerate(training_db.query(entity)):
meme.name_idx = idx
training_db.commit()
def get_not_template(self):
rand = random.randint(0, self.max_name_idx["not_a_template"])
clause = cast(ClauseElement, self.not_a_template.name_idx == rand)
q = training_db.query(self.not_a_template.path).filter(clause)
path = cast(Tuple[str], q.first())[0]
return (toTensorOnly(Img.open(path)), self.name_size)
