Skip to content

Commit f25a6ee

Browse files
aaminiaamini
committed
lab 2 funcs
1 parent 822748d commit f25a6ee

File tree

4 files changed

+179
-178
lines changed

4 files changed

+179
-178
lines changed

‎lab2/util.py

Lines changed: 0 additions & 174 deletions
Original file line numberDiff line numberDiff line change
@@ -1,174 +0,0 @@
1-
import cv2
2-
import os
3-
import matplotlib.pyplot as plt
4-
import numpy as np
5-
import tensorflow as tf
6-
import time
7-
import h5py
8-
import sys
9-
10-
IM_SHAPE = (64, 64, 3)
11-
12-
13-
class TrainingDatasetLoader(object):
14-
def __init__(self, data_path):
15-
16-
print ("Opening {}".format(data_path))
17-
sys.stdout.flush()
18-
19-
self.cache = h5py.File(data_path, 'r')
20-
21-
print ("Loading data into memory...")
22-
sys.stdout.flush()
23-
self.images = self.cache['images'][:]
24-
self.labels = self.cache['labels'][:]
25-
self.image_dims = self.images.shape
26-
n_train_samples = self.image_dims[0]
27-
28-
self.train_inds = np.random.permutation(np.arange(n_train_samples))
29-
30-
self.pos_train_inds = self.train_inds[ self.labels[self.train_inds, 0] == 1.0 ]
31-
self.neg_train_inds = self.train_inds[ self.labels[self.train_inds, 0] != 1.0 ]
32-
33-
def get_train_size(self):
34-
return self.train_inds.shape[0]
35-
36-
def get_train_steps_per_epoch(self, batch_size, factor=10):
37-
return self.get_train_size()//factor//batch_size
38-
39-
def get_batch(self, n, only_faces=False, p_pos=None, p_neg=None, return_inds=False):
40-
if only_faces:
41-
selected_inds = np.random.choice(self.pos_train_inds, size=n, replace=False, p=p_pos)
42-
else:
43-
selected_pos_inds = np.random.choice(self.pos_train_inds, size=n//2, replace=False, p=p_pos)
44-
selected_neg_inds = np.random.choice(self.neg_train_inds, size=n//2, replace=False, p=p_neg)
45-
selected_inds = np.concatenate((selected_pos_inds, selected_neg_inds))
46-
47-
sorted_inds = np.sort(selected_inds)
48-
train_img = self.images[sorted_inds,:,:,::-1]/255.
49-
train_label = self.labels[sorted_inds,...]
50-
return (train_img, train_label, sorted_inds) if return_inds else (train_img, train_label)
51-
52-
def get_n_most_prob_faces(self, prob, n):
53-
idx = np.argsort(prob)[::-1]
54-
most_prob_inds = self.pos_train_inds[idx[:10*n:10]]
55-
return self.images[most_prob_inds,...]/255.
56-
57-
def get_all_train_faces(self):
58-
return self.images[ self.pos_train_inds ]
59-
60-
61-
62-
class PPBFaceEvaluator:
63-
''' Evaluate on the PPB dataset'''
64-
def __init__(self, skip=4):
65-
66-
path_to_faces = tf.keras.utils.get_file('ppb', 'https://www.dropbox.com/s/l0lp6qxeplumouf/PPB.tar?dl=1', extract=True)
67-
self.ppb_root = os.path.join(os.path.split(path_to_faces)[0], 'PPB-2017')
68-
69-
ppb_anno = os.path.join(self.ppb_root,'PPB-2017-metadata.csv')
70-
71-
self.anno_dict = {}
72-
with open(ppb_anno) as f:
73-
for line in f.read().split('\r'):
74-
ind, name, gender, numeric, skin, country = line.split(',')
75-
self.anno_dict[name] = (gender.lower(),skin.lower())
76-
77-
image_dir = os.path.join(self.ppb_root, "imgs")
78-
image_files = sorted(os.listdir(image_dir))[::skip] #sample every 4 images for computation time in the lab
79-
80-
self.raw_images = {
81-
'male_darker':[],
82-
'male_lighter':[],
83-
'female_darker':[],
84-
'female_lighter':[],
85-
}
86-
87-
for filename in image_files:
88-
if not filename.endswith(".jpg"):
89-
continue
90-
image = cv2.imread(os.path.join(image_dir,filename))[:,:,::-1]
91-
gender, skin = self.anno_dict[filename]
92-
self.raw_images[gender+'_'+skin].append(image)
93-
94-
95-
def get_sample_faces_from_demographic(self, gender, skin_color):
96-
key = self.__get_key(gender, skin_color)
97-
data = self.raw_images[key][50]/255.
98-
return data
99-
100-
101-
def evaluate(self, models_to_test, gender, skin_color, output_idx=None, from_logit=False, patch_stride=0.2, patch_depth=5):
102-
correct_predictions = [0.0]*len(models_to_test)
103-
104-
key = self.__get_key(gender, skin_color)
105-
num_faces = len(self.raw_images[key])
106-
107-
import progressbar
108-
bar = progressbar.ProgressBar()
109-
for face_idx in bar(range(num_faces)):
110-
111-
image = self.raw_images[key][face_idx]
112-
height, width, _ = image.shape
113-
114-
patches, bboxes = slide_square(image, patch_stride, width/2, width, patch_depth)
115-
patches = tf.cast(tf.constant(patches, dtype=tf.uint8), tf.float32)/255.
116-
117-
for model_idx, model in enumerate(models_to_test):
118-
out = model(patches)
119-
y = out if output_idx is None else out[output_idx]
120-
y = y.numpy()
121-
y_inds = np.argsort(y.flatten())
122-
most_likely_prob = y[y_inds[-1]]
123-
if (from_logit and most_likely_prob >= 0.0) or \
124-
(not from_logit and most_likely_prob >= 0.5):
125-
correct_predictions[model_idx] += 1
126-
127-
accuracy = [correct_predictions[i]/num_faces for i,_ in enumerate(models_to_test)]
128-
return accuracy
129-
130-
131-
def __get_key(self, gender, skin_color):
132-
gender = gender.lower()
133-
skin_color = skin_color.lower()
134-
assert gender in ['male', 'female']
135-
assert skin_color in ['lighter', 'darker']
136-
return '{}_{}'.format(gender, skin_color)
137-
138-
139-
140-
''' function to slide a square across image and extract square regions
141-
img = the image
142-
stride = (0,1], provides the fraction of the dimension for which to slide to generate a crop
143-
max_size = maximum square size
144-
min_size = minimum square size
145-
n = number of different sizes including min_size, max_size '''
146-
def slide_square(img, stride, min_size, max_size, n):
147-
img_h, img_w = img.shape[:2]
148-
149-
square_sizes = np.linspace(min_size, max_size, n, dtype=np.int32)
150-
square_images = []
151-
square_bbox = [] # list of list of tuples: [(i1,j1), (i2,j2)] where i1,j1 is the top left corner; i2,j2 is bottom right corner
152-
# for each of the square_sizes
153-
for level, sq_dim in enumerate(square_sizes):
154-
155-
stride_length = int(stride*sq_dim)
156-
stride_start_i = xrange(0,int(img_h-sq_dim+1),stride_length)
157-
stride_start_j = xrange(0,int(img_w-sq_dim+1),stride_length)
158-
for i in stride_start_i:
159-
for j in stride_start_j:
160-
square_top_left = (i,j)
161-
162-
square_bottom_right = (i+sq_dim,j+sq_dim)
163-
164-
square_corners = (square_top_left,square_bottom_right)
165-
square_corners_show = ((j,i),(j+sq_dim,i+sq_dim))
166-
square_image = img[i:i+sq_dim,j:j+sq_dim]
167-
168-
square_resize = cv2.resize(square_image, IM_SHAPE[:2], interpolation=cv2.INTER_NEAREST)
169-
# append to list of images and bounding boxes
170-
171-
square_images.append(square_resize)
172-
square_bbox.append(square_corners)
173-
174-
return square_images, square_bbox

‎mitdeeplearning/__init__.py

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1,2 +1,4 @@
1-
import mitdeeplearning.lab1
21
import mitdeeplearning.util
2+
3+
import mitdeeplearning.lab1
4+
import mitdeeplearning.lab2

‎mitdeeplearning/lab2.py

Lines changed: 174 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,174 @@
1+
import cv2
2+
import os
3+
import matplotlib.pyplot as plt
4+
import numpy as np
5+
import tensorflow as tf
6+
import time
7+
import h5py
8+
import sys
9+
10+
IM_SHAPE = (64, 64, 3)
11+
12+
13+
class TrainingDatasetLoader(object):
14+
def __init__(self, data_path):
15+
16+
print ("Opening {}".format(data_path))
17+
sys.stdout.flush()
18+
19+
self.cache = h5py.File(data_path, 'r')
20+
21+
print ("Loading data into memory...")
22+
sys.stdout.flush()
23+
self.images = self.cache['images'][:]
24+
self.labels = self.cache['labels'][:]
25+
self.image_dims = self.images.shape
26+
n_train_samples = self.image_dims[0]
27+
28+
self.train_inds = np.random.permutation(np.arange(n_train_samples))
29+
30+
self.pos_train_inds = self.train_inds[ self.labels[self.train_inds, 0] == 1.0 ]
31+
self.neg_train_inds = self.train_inds[ self.labels[self.train_inds, 0] != 1.0 ]
32+
33+
def get_train_size(self):
34+
return self.train_inds.shape[0]
35+
36+
def get_train_steps_per_epoch(self, batch_size, factor=10):
37+
return self.get_train_size()//factor//batch_size
38+
39+
def get_batch(self, n, only_faces=False, p_pos=None, p_neg=None, return_inds=False):
40+
if only_faces:
41+
selected_inds = np.random.choice(self.pos_train_inds, size=n, replace=False, p=p_pos)
42+
else:
43+
selected_pos_inds = np.random.choice(self.pos_train_inds, size=n//2, replace=False, p=p_pos)
44+
selected_neg_inds = np.random.choice(self.neg_train_inds, size=n//2, replace=False, p=p_neg)
45+
selected_inds = np.concatenate((selected_pos_inds, selected_neg_inds))
46+
47+
sorted_inds = np.sort(selected_inds)
48+
train_img = self.images[sorted_inds,:,:,::-1]/255.
49+
train_label = self.labels[sorted_inds,...]
50+
return (train_img, train_label, sorted_inds) if return_inds else (train_img, train_label)
51+
52+
def get_n_most_prob_faces(self, prob, n):
53+
idx = np.argsort(prob)[::-1]
54+
most_prob_inds = self.pos_train_inds[idx[:10*n:10]]
55+
return self.images[most_prob_inds,...]/255.
56+
57+
def get_all_train_faces(self):
58+
return self.images[ self.pos_train_inds ]
59+
60+
61+
62+
class PPBFaceEvaluator:
63+
''' Evaluate on the PPB dataset'''
64+
def __init__(self, skip=4):
65+
66+
path_to_faces = tf.keras.utils.get_file('ppb', 'https://www.dropbox.com/s/l0lp6qxeplumouf/PPB.tar?dl=1', extract=True)
67+
self.ppb_root = os.path.join(os.path.split(path_to_faces)[0], 'PPB-2017')
68+
69+
ppb_anno = os.path.join(self.ppb_root,'PPB-2017-metadata.csv')
70+
71+
self.anno_dict = {}
72+
with open(ppb_anno) as f:
73+
for line in f.read().split('\r'):
74+
ind, name, gender, numeric, skin, country = line.split(',')
75+
self.anno_dict[name] = (gender.lower(),skin.lower())
76+
77+
image_dir = os.path.join(self.ppb_root, "imgs")
78+
image_files = sorted(os.listdir(image_dir))[::skip] #sample every 4 images for computation time in the lab
79+
80+
self.raw_images = {
81+
'male_darker':[],
82+
'male_lighter':[],
83+
'female_darker':[],
84+
'female_lighter':[],
85+
}
86+
87+
for filename in image_files:
88+
if not filename.endswith(".jpg"):
89+
continue
90+
image = cv2.imread(os.path.join(image_dir,filename))[:,:,::-1]
91+
gender, skin = self.anno_dict[filename]
92+
self.raw_images[gender+'_'+skin].append(image)
93+
94+
95+
def get_sample_faces_from_demographic(self, gender, skin_color):
96+
key = self.__get_key(gender, skin_color)
97+
data = self.raw_images[key][50]/255.
98+
return data
99+
100+
101+
def evaluate(self, models_to_test, gender, skin_color, output_idx=None, from_logit=False, patch_stride=0.2, patch_depth=5):
102+
correct_predictions = [0.0]*len(models_to_test)
103+
104+
key = self.__get_key(gender, skin_color)
105+
num_faces = len(self.raw_images[key])
106+
107+
import progressbar
108+
bar = progressbar.ProgressBar()
109+
for face_idx in bar(range(num_faces)):
110+
111+
image = self.raw_images[key][face_idx]
112+
height, width, _ = image.shape
113+
114+
patches, bboxes = slide_square(image, patch_stride, width/2, width, patch_depth)
115+
patches = tf.cast(tf.constant(patches, dtype=tf.uint8), tf.float32)/255.
116+
117+
for model_idx, model in enumerate(models_to_test):
118+
out = model(patches)
119+
y = out if output_idx is None else out[output_idx]
120+
y = y.numpy()
121+
y_inds = np.argsort(y.flatten())
122+
most_likely_prob = y[y_inds[-1]]
123+
if (from_logit and most_likely_prob >= 0.0) or \
124+
(not from_logit and most_likely_prob >= 0.5):
125+
correct_predictions[model_idx] += 1
126+
127+
accuracy = [correct_predictions[i]/num_faces for i,_ in enumerate(models_to_test)]
128+
return accuracy
129+
130+
131+
def __get_key(self, gender, skin_color):
132+
gender = gender.lower()
133+
skin_color = skin_color.lower()
134+
assert gender in ['male', 'female']
135+
assert skin_color in ['lighter', 'darker']
136+
return '{}_{}'.format(gender, skin_color)
137+
138+
139+
140+
''' function to slide a square across image and extract square regions
141+
img = the image
142+
stride = (0,1], provides the fraction of the dimension for which to slide to generate a crop
143+
max_size = maximum square size
144+
min_size = minimum square size
145+
n = number of different sizes including min_size, max_size '''
146+
def slide_square(img, stride, min_size, max_size, n):
147+
img_h, img_w = img.shape[:2]
148+
149+
square_sizes = np.linspace(min_size, max_size, n, dtype=np.int32)
150+
square_images = []
151+
square_bbox = [] # list of list of tuples: [(i1,j1), (i2,j2)] where i1,j1 is the top left corner; i2,j2 is bottom right corner
152+
# for each of the square_sizes
153+
for level, sq_dim in enumerate(square_sizes):
154+
155+
stride_length = int(stride*sq_dim)
156+
stride_start_i = xrange(0,int(img_h-sq_dim+1),stride_length)
157+
stride_start_j = xrange(0,int(img_w-sq_dim+1),stride_length)
158+
for i in stride_start_i:
159+
for j in stride_start_j:
160+
square_top_left = (i,j)
161+
162+
square_bottom_right = (i+sq_dim,j+sq_dim)
163+
164+
square_corners = (square_top_left,square_bottom_right)
165+
square_corners_show = ((j,i),(j+sq_dim,i+sq_dim))
166+
square_image = img[i:i+sq_dim,j:j+sq_dim]
167+
168+
square_resize = cv2.resize(square_image, IM_SHAPE[:2], interpolation=cv2.INTER_NEAREST)
169+
# append to list of images and bounding boxes
170+
171+
square_images.append(square_resize)
172+
square_bbox.append(square_corners)
173+
174+
return square_images, square_bbox

‎setup.py

Lines changed: 2 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -11,7 +11,6 @@ def get_dist(pkgname):
1111
install_deps = [
1212
'numpy',
1313
'regex',
14-
'progressbar',
1514
'tqdm',
1615

1716
]
@@ -22,13 +21,13 @@ def get_dist(pkgname):
2221
setup(
2322
name = 'mitdeeplearning', # How you named your package folder (MyLib)
2423
packages = ['mitdeeplearning'], # Chose the same as "name"
25-
version = '0.3.7', # Start with a small number and increase it with every change you make
24+
version = '0.3.8', # Start with a small number and increase it with every change you make
2625
license='MIT', # Chose a license from here: https://help.github.com/articles/licensing-a-repository
2726
description = 'Official software labs for MIT Introduction to Deep Learning (http://introtodeeplearning.com)', # Give a short description about your library
2827
author = 'Alexander Amini', # Type in your name
2928
author_email = '[email protected]', # Type in your E-Mail
3029
url = 'http://introtodeeplearning.com', # Provide either the link to your github or to your website
31-
download_url = 'https://github.com/aamini/introtodeeplearning_labs/archive/v0.3.7.tar.gz', # I explain this later on
30+
download_url = 'https://github.com/aamini/introtodeeplearning_labs/archive/v0.3.8.tar.gz', # I explain this later on
3231
keywords = ['deep learning', 'neural networks', 'tensorflow', 'introduction'], # Keywords that define your package best
3332
install_requires=install_deps,
3433
classifiers=[

0 commit comments

Comments
 (0)