def normalize_triplet_funcs(mean: FloatTensor,
std: FloatTensor,
do_x: bool = True,
do_y: bool = False) -> Tuple[Callable, Callable]:
"Create normalize/denormalize func using `mean` and `std`, can specify `do_y` and `device`."
mean, std = tensor(mean), tensor(std)
return (
partial(_normalize_triplet_batch,
mean=mean,
std=std,
do_x=do_x,
do_y=do_y),
partial(denormalize_triplet, mean=mean, std=std, do_x=do_x),
)
def create_dataset(path_fullRes: Path, path_list, downsize=True):
il = ImageList.from_folder(path_fullRes)
for p, size, qf in path_list:
if not p.exists():
print(f"Creating {p}")
print(f"Size: {size} with {qf} quality factor")
parallel(partial(create_training_images,
p_hr=path_fullRes,
p_lr=p,
size=size,
qualityFactor=qf,
downsize=downsize), il.items)
def get_all_issue_text(owner, repo, inf_wrapper, workers=64):
"""
Prepare embedding features of all issues in a given repository.
Returns
------
dict
{'features':list, 'labels':list, 'nums':list}
"""
# prepare list of issue nums
max_num = find_max_issue_num(owner, repo)
get = partial(get_issue_text, owner=owner, repo=repo, skip_issue=True)
issues = parallel(get, list(range(1, max_num + 1)), max_workers=workers)
# filter out issues with problems
filtered_issues = []
for issue in issues:
if issue:
filtered_issues.append(issue)
logging.info(f'Retrieved {len(filtered_issues)} issues.')
features = []
labels = []
nums = []
for issue in tqdm(filtered_issues):
labels.append(issue['labels'])
nums.append(issue['num'])
# calculate embedding
text = inf_wrapper.process_dict(issue)['text']
feature = inf_wrapper.get_pooled_features(text).detach().cpu()
# only need the first 1600 dimensions
features.append(feature[:, :1600])
assert len(features) == len(
labels
), 'Error you have mismatch b/w number of observations and labels.'
return {
'features': torch.cat(features).numpy(),
'labels': labels,
'nums': nums
}
def get_all_issue_text(owner, repo, inf_wrapper, workers=64):
"""
Prepare embedding features of all issues in a given repository.
Returns
------
dict
{'features':list, 'labels':list, 'nums':list}
"""
# prepare list of issue nums
max_num = find_max_issue_num(owner, repo)
get = partial(get_issue_text, owner=owner, repo=repo, skip_issue=True)
issues = parallel(get, list(range(1, max_num + 1)), max_workers=workers)
# filter out issues with problems
filtered_issues = []
if not issues:
raise ValueError(f"No issues retrieved for {owner}/{repo}")
for issue in issues:
if issue:
filtered_issues.append(issue)
logging.info(
f'Repo {owner}/{repo} Retrieved {len(filtered_issues)} issues.')
features = []
labels = []
nums = []
issues_dict = {'title': [], 'body': []}
for issue in tqdm(filtered_issues):
labels.append(issue['labels'])
nums.append(issue['num'])
issues_dict['title'].append(issue['title'])
issues_dict['body'].append(issue['body'])
features = inf_wrapper.df_to_embedding(pd.DataFrame.from_dict(issues_dict))
assert len(features) == len(
labels
), 'Error you have mismatch b/w number of observations and labels.'
return {'features': features[:, :1600], 'labels': labels, 'nums': nums}
from fastai.vision import get_transforms, ImageList, FloatList
from fastai.vision import cnn_learner, Learner
from optim.ranger import Ranger
from fastai.core import partial
from torch import Tensor
def accuracy_thresh(y_pred, y_true, thresh:float=0.5, sigmoid:bool=False):
"Computes accuracy when `y_pred` and `y_true` are the same size."
if sigmoid: y_pred = y_pred.sigmoid()
y_pred = np.array(y_pred).flatten()
y_true = np.array(y_true).flatten()
return Tensor([((y_pred>thresh)==y_true).mean()])
optar = partial(Ranger)
# In[25]:
CV=1
seed = CV
bs = 20
tfms = get_transforms(flip_vert=True, do_flip=True, max_zoom=1.05, max_lighting=0.2,
max_warp=0.05, max_rotate=5.)
data = (ImageList.from_df(df=image_df,path=DATA_BASE_PATH / 'train_images',cols='ImageId')
.split_from_df()
.label_from_df(cols='Detected',label_cls=FloatList)
.transform(tfms)
.databunch(bs=bs,num_workers=4)
