def testRotation(self):
from generator import ImageGenerator
batch1 = ImageGenerator(self.file_path, self.label_path, 100, [32, 32, 3], rotation=False, mirroring=False, shuffle=False).next()[0]
batch2 = ImageGenerator(self.file_path, self.label_path, 100, [32, 32, 3], rotation=True, mirroring=False, shuffle=False).next()[0]
# determine the images which were augmented
augmented_images_indices = np.sum(np.abs(batch1 - batch2), axis=(1, 2, 3)).astype(np.bool_)
# extract corner points for each sample and reduce augmented to one characateristic row
# this row is also inverted for simpler computations
augmented_corners = self._get_corner_points(batch2[augmented_images_indices])
characteristic_corners = augmented_corners[:, 0, ::-1]
original_corners = self._get_corner_points(batch1[augmented_images_indices])
# subtract characteristic corners to original corners. after summing spacial and channel dimensions, 0 are expected
# e.g. if sample 2 is rotated by 90, we have a 0 in rot1 at the 2nd posistion see similiar to testMirroring
rot1 = np.sum(original_corners[:, :, 0] - characteristic_corners, axis=(1, 2))
rot2 = np.sum(original_corners[:, 1, :] - characteristic_corners, axis=(1, 2))
rot3 = np.sum(original_corners[:, ::-1, 1] - characteristic_corners, axis=(1, 2))
# assumption is that augmented images are either rotated by 90 (rot1), 180 (ro2) or 270 (rot3) degrees, thus
# their elementwise product must be zero
np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, batch1, batch2)
np.testing.assert_almost_equal(np.sum(rot1 * rot2 * rot3), 0)
def testShuffle(self):
# Creates two image generator objects.
# Since shuffle is enabled for one image generator the output should be different.
from generator import ImageGenerator
gen = ImageGenerator(self.file_path, self.label_path, 10, [32, 32, 3], rotation=False, mirroring=False, shuffle=True)
gen2 = ImageGenerator(self.file_path, self.label_path, 10, [32, 32, 3], rotation=False, mirroring=False, shuffle=False)
np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, gen.next()[0], gen2.next()[0])
np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, gen.next()[1], gen2.next()[1])
def testResetIndex(self):
# Data contains 100 image samples, for a batchsize of 60 an
# overlap of 20 occurs, therefore the first 20 elements
# of the first batch should be equal to the last 20 of the second batch
from generator import ImageGenerator
gen = ImageGenerator(self.file_path, self.label_path, 60, [32, 32, 3], rotation=False, mirroring=False, shuffle=False)
b1 = gen.next()[0]
b2 = gen.next()[0]
np.testing.assert_almost_equal(b1[:20], b2[40:])
def testResetIndex(self):
from generator import ImageGenerator
# Data contains 100 image samples, create two overlapping batches
gen = ImageGenerator(self.file_path,
self.json_path,
60, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
b1 = gen.next()[0]
b2 = gen.next()[0]
np.testing.assert_almost_equal(b1[:20], b2[40:])
def testLabels(self):
# this test makes sure your generator returns integers as labels and not any strings or floating point values
# if this one fails make sure you cast your array values to integers
from generator import ImageGenerator
label = ImageGenerator(self.file_path, self.label_path, 12, [50, 50, 3], rotation=False, mirroring=False,
shuffle=False).next()[1]
self.assertFalse(isinstance(label[0], str))
self.assertTrue(np.issubdtype(np.array(label).dtype, np.integer))
def testInit(self):
# Creates two image generator objects without shuffling.
# Calling next on either one should result in the same output
from generator import ImageGenerator
gen = ImageGenerator(self.file_path,
self.label_path,
12, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
gen2 = ImageGenerator(self.file_path,
self.label_path,
12, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
np.testing.assert_almost_equal(gen.next[0], gen2.next[0])
np.testing.assert_almost_equal(gen.next[1], gen2.next[1])
def testResize(self):
from generator import ImageGenerator
batch = ImageGenerator(self.file_path,
self.json_path,
12, [50, 50, 3],
rotation=False,
mirroring=False,
shuffle=False).next()[0]
self.assertEqual(batch.shape, (12, 50, 50, 3))
def testMirroring(self):
from generator import ImageGenerator
batch1 = ImageGenerator(self.file_path, self.label_path, 12, [32, 32, 3], rotation=False, mirroring=False, shuffle=False).next()[0]
batch2 = ImageGenerator(self.file_path, self.label_path, 12, [32, 32, 3], rotation=False, mirroring=True, shuffle=False).next()[0]
# determine the images which were augmented
augmented_images_indices = np.sum(np.abs(batch1 - batch2), axis=(1, 2, 3)).astype(np.bool_)
# extract corner points for each sample
augmented_corners = self._get_corner_points(batch2[augmented_images_indices])
original_corners = self._get_corner_points(batch1[augmented_images_indices])
# pick vertical ref points and compare them with opposites in original corners
# compute then the diff and sum over vertical axis and channels. those images which were flipped vertically
# contain now zeros
vertical_augmented_corners = augmented_corners[:, :, 0]
vertical = np.sum(original_corners[:, :, 1, :] - vertical_augmented_corners, axis=(1, 2))
# pick horizontal ref points and compare them with opposites in original corners
# compute then the diff and sum over horizontal axis and channels. those images which were flipped horizontally
# contain now zeros
horizontal_augmented_corners = augmented_corners[:, 0, :]
horizontal = np.sum(original_corners[:, 1, :, :] - horizontal_augmented_corners, axis=(1, 2))
# pick top left corner and bottom right corner (diagonals are flipped if double mirror)
vertical_horizontal_augmented_corners = np.stack(
list(zip(augmented_corners[:, 0, 0, :], augmented_corners[:, 1, 1, :])))
original_corner_diagonals = np.stack(
list(zip(original_corners[:, 1, 1, :], original_corners[:, 0, 0, :])))
horizontal_vertical = np.sum(original_corner_diagonals - vertical_horizontal_augmented_corners, axis=(1, 2))
# the elementwise product of horizontal and vertical must be zero
# since the images can only be augmented with vertical or horizontal mirroring
np.testing.assert_raises(AssertionError, np.testing.assert_array_equal, batch1, batch2)
np.testing.assert_almost_equal(np.sum(vertical * horizontal * horizontal_vertical), 0)
def fit(self):
self.model = Sequential()
self.model.add(
Conv2D(32, (2, 2),
padding='valid',
strides=1,
input_shape=(1024, 1024, 1)))
self.model.add(Activation('relu'))
self.model.add(Conv2D(32, (2, 2)))
self.model.add(Activation('relu'))
self.model.add(Conv2D(32, (2, 2)))
self.model.add(Activation('relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
self.model.add(Conv2D(64, (4, 4)))
self.model.add(Activation('relu'))
self.model.add(Conv2D(64, (4, 4)))
self.model.add(Activation('relu'))
self.model.add(MaxPooling2D(pool_size=(2, 2)))
self.model.add(Flatten())
self.model.add(Dense(4096))
self.model.add(Activation('relu'))
self.model.add(Dropout(0.2))
self.model.add(Dense(4096))
self.model.add(Activation('relu'))
self.model.add(Dropout(0.2))
self.model.add(Dense(self.classes))
self.model.add(Activation('softmax'))
self.model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
self.model.fit_generator(generator=ImageGenerator(
X_train, y_train, self.batch_size),
batch_size=self.batch_size,
epochs=self.epoch,
verbose=1,
validation_split=0.2,
class_weight='auto',
callbacks=[
EarlyStopping(monitor='acc',
min_delta=0.001,
patience=2,
verbose=0,
mode='auto')
])
def testShuffle(self):
from generator import ImageGenerator
gen = ImageGenerator(self.file_path,
self.json_path,
10, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=True)
gen2 = ImageGenerator(self.file_path,
self.json_path,
10, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
np.testing.assert_raises(AssertionError, np.testing.assert_array_equal,
gen.next()[0],
gen2.next()[0])
np.testing.assert_raises(AssertionError, np.testing.assert_array_equal,
gen.next()[1],
gen2.next()[1])
def testInit(self):
from generator import ImageGenerator
gen = ImageGenerator(self.file_path,
self.json_path,
12, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
gen2 = ImageGenerator(self.file_path,
self.json_path,
12, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
np.testing.assert_almost_equal(gen.next()[0], gen2.next()[0])
np.testing.assert_almost_equal(gen.next()[1], gen2.next()[1])
#C = pat.Checkers(100, 25)
#C.draw()
#C.show()
#RGB = pat.Spectrum(255)
#RGB.draw()
#RGB.show()
#Circ = pat.Circle(1024, 200, (512, 256))
#Circ.draw()
#Circ.show()
from generator import ImageGenerator
label_path = '/home/shrihari/Downloads/Deep Learning/Exercise1/src_to_implement/data/Labels.json'
file_path = '/home/shrihari/Downloads/Deep Learning/Exercise1/src_to_implement/data/exercise_data/'
gen = ImageGenerator(file_path,
label_path,
12, [32, 32, 3],
rotation=True,
mirroring=False,
shuffle=True)
#b,l = gen.next()
#print(b)
#print(l)
gen.show()
print('Test Image Count: {:d}'.format(len(test_images)))
model_option = 'fcn'
#Resize images to be fed into model
if model_option=='random':
resize_images = (350,525) #Random baseline resize
elif model_option=='fcn':
resize_images = (448,672) #FCN resize
elif model_option=='mask-rcnn':
resize_images = (1400,2100) #Resize is handled by config
else:
print('Please choose a valid model option')
batch_size = math.ceil(len(test_images) / 100)
#Use for random baseline and FCN
test_generator = ImageGenerator(test_images, batch_size=batch_size, resize=resize_images)
#Open file for writing
f = open('submission.csv', 'w+')
f.write('Image_Label,EncodedPixels\n')
#Write pixel encoded image masks out to file in batches
stop_flag = False
iteration = 1
for b, batch in enumerate(test_generator):
while not stop_flag:
if iteration*batch_size > len(test_images):
batch = batch[0:(len(test_images)-iteration*batch_size)]
stop_flag = True
#Write predictions to submission file
print('Running Batch {:d}'.format(iteration))
from pattern import Checker
from pattern import Circle
from generator import ImageGenerator
# Checkerboard
a = Checker(250, 25)
a.draw()
a.show()
# Circle
b = Circle(1024, 200, (512, 256))
b.draw()
b.show()
# Image generator
ima_gen = ImageGenerator('./exercise_data/',
'./Labels.json',
12, [32, 32, 3],
rotation=False,
mirroring=False,
shuffle=False)
ima_gen.show(resize=True)
from flask import Flask, escape, request, jsonify, abort, request, send_from_directory
from flask_restplus import Api, Resource, reqparse, fields
from generator import ImageGenerator
import json
import uuid
import os
from cleanup import clean_images
IMAGE_DIR = "{}/api/static/cleaned-images".format(os.getcwd())
# flask app configuration
app = Flask(__name__, static_folder='static')
api = Api(app, doc="/swagger/")
# image gnerator
generator = ImageGenerator()
UPLOAD_DIRECTORY = "./api/static/images"
fields = api.model(
'NewPokemonImageRequestBody', {
'id1': fields.Integer(description='pokedex id', required=True, min=1),
'id2': fields.Integer(description='pokedex id', required=True, min=1),
})
@api.route("/home")
class Home(Resource):
@api.doc("home")
def get(self):
return jsonify("Welcome to PokeMate image generator!")
