This notebook describes how to perform Scalar regression on images
from fastai.vision.all import *
Below you will find the exact imports for everything we use today
from fastcore.basics import Int
from fastcore.transform import DisplayedTransform
from fastcore.xtras import Path
from fastai.callback.progress import ProgressCallback
from fastai.callback.schedule import fine_tune
from fastai.data.block import DataBlock, RegressionBlock
from fastai.data.transforms import RandomSplitter, get_image_files, ColReader
from fastai.losses import MSELossFlat
from fastai.torch_core import ShowTitle, show_title, TitledFloat
from fastai.vision.augment import Resize, aug_transforms, Flip, Rotate, Zoom, Warp
from fastai.vision.data import ImageBlock
from fastai.vision.learner import cnn_learner
from torchvision.models.resnet import resnet50
from datetime import datetime, timedelta
import pandas as pd
from scipy.io import loadmat
import numpy as np
Dataset
The dataset is the IMDB-WIKI dataset in which we try to guess the expected age from an indivudal image. We'll focus on the WIKI dataset.
url = 'https://data.vision.ee.ethz.ch/cvl/rrothe/imdb-wiki/static/wiki_crop.tar'
!wget {url}
!tar -xvf 'wiki_crop.tar'
Now the dataset itself is pretty easy, there is an included .mat file we can convert to Pandas. Let's go through that process:
We are following the same preperation code as seen here
- Note: the video and this notebook differ here. This is because I was labelling the data wrong intially, and the floats are now inheritly supported
from scipy.io import loadmat
from datetime import datetime, timedelta
def prepare_dataframe(path:Path):
"Builds wiki_crop DataFrame based on https://sefiks.com/2019/02/13/apparent-age-and-gender-prediction-in-keras/"
mat = loadmat(path)
instances = mat['wiki'][0][0][0].shape[1]
columns = ['dob', 'photo_taken', 'full_path', 'gender', 'name', 'face_location', 'face_score', 'second_face_score']
df = pd.DataFrame(index=range(0, instances), columns=columns)
for i in mat:
if i == 'wiki':
curr_arr = mat[i][0][0]
for j, val in enumerate(curr_arr):
df[columns[j]] = pd.DataFrame(val[0])
### Python datatime conversion
df['date_of_birth'] = df['dob'].apply(datenum_to_datetime)
df['age'] = df['photo_taken'] - df['date_of_birth']
## Data cleaning
df = df[df['face_score'] != -np.inf] # picture doesn't include faces
df = df[df['second_face_score'].isna()] # more than one face
df = df[df['face_score'] > 3] # has a score
df = df[df['age'] < 100] # older than 100 (mostly paintings)
df = df[df['age'] > 0] # unborn people
df = df[['full_path', 'age']]
df['full_path'] = df['full_path'].apply(lambda x: x[0])
return df
def datenum_to_datetime(datenum):
days = datenum % 1
hours = days % 1 * 24
minutes = hours % 1 * 60
seconds = minutes % 1 * 60
exact_date = datetime.fromordinal(int(datenum)) \
+ timedelta(days=int(days)) + timedelta(hours=int(hours)) \
+ timedelta(minutes=int(minutes)) + timedelta(seconds=round(seconds)) \
- timedelta(days=366)
return exact_date.year
path = Path('wiki_crop/wiki.mat')
df = prepare_dataframe(path)
df.head()
Let's make a get_x and a get_y since we are using a DataFrame
get_x = ColReader('full_path', pref=Path('wiki_crop'))
get_y = ColReader('age')
block = DataBlock(blocks=(ImageBlock, RegressionBlock()),
get_x = get_x,
get_y = get_y,
splitter=RandomSplitter(),
item_tfms=Resize(460, method='squish'),
batch_tfms=[*aug_transforms(size=224, max_warp=0)])
Let's look at a batch of data;
dls = block.dataloaders(df)
dls.show_batch()
dls.c
And now we can run this through cnn_learner
from fastai.metrics import mae
learn = cnn_learner(dls, resnet50, loss_func=MSELossFlat(), y_range=(0,100), metrics=mae)
learn.lr_find()
learn.fine_tune(5, 1e-2)