Skip to content
Switch branches/tags
Go to file
Cannot retrieve contributors at this time
import argparse, glob, os, sys
import numpy as np
from random import shuffle
from PIL import Image
from sklearn.linear_model import LogisticRegressionCV
from sklearn import preprocessing
from enum import Enum
from timeit import default_timer as timer
class Solver(Enum):
lbfgs = "lbfgs"
sag = "sag"
saga = "saga"
# magic methods for argparse compatibility
def __str__(self):
def __repr__(self):
return str(self)
def argparse(s):
return Solver[s]
except KeyError:
return s
DEFAULT_SOLVER = Solver.lbfgs
def main():
parser = argparse.ArgumentParser(description="Train a image classifier using Logistic Regression.")
parser.add_argument("--imageSize", type=int , help="The image size to use" , default=DEFAULT_IMG_SIZE)
parser.add_argument("--solver" , type=Solver.argparse, help="This to use for optimization" , default=DEFAULT_SOLVER, choices=list(Solver))
parser.add_argument("--rescale" , dest='rescale', action='store_true' , help="Rescale images between 0 and 1.")
parser.add_argument("--no-rescale", dest='rescale', action='store_false', help="Don't rescale images.")
parser.add_argument("--debug" , dest='debug', action='store_true' , help="Enable debugging (uses only 100 training examples).")
parser.add_argument("--no-debug", dest='debug', action='store_false', help="Disable debugging.")
args = parser.parse_args()
start1 = timer()
X_train, y_train, X_test, y_test = loadDataSet(args.imageSize, args.rescale)
if args.debug:
X_train = X_train[0:100, :]
y_train = y_train[0:100]
X_test = X_test [0:100, :]
y_test = y_test [0:100]
print("Size and type of training data : {} and {}".format(X_train.shape, X_train.dtype))
print("Size and type of training labels: {} and {}".format(y_train.shape, y_train.dtype))
print("Size and type of testing data : {} and {}".format(X_test.shape, X_test.dtype))
print("Size and type of testing labels: {} and {}".format(y_test.shape, y_test.dtype))
end1 = timer()
timeToLoad = end1 - start1
print("Time taken to load model = {:.1f} seconds".format(timeToLoad))
start2 = timer()
print("Fitting logistic classifier on the data using {} ...".format(args.solver))
classifier = LogisticRegressionCV(cv=3, n_jobs=8, max_iter=100, verbose=0, solver=args.solver.value), y_train)"LRModel_{}_{}_{}".format(args.imageSize, "Rescaled" if args.rescale else "", args.solver), classifier)
end2 = timer()
timeToFit = end2 - start2
print("Time taken to fit model = {:.1f} seconds".format(timeToFit))
print("Calculating training accuracy...")
train_acc = classifier.score(X_train, y_train) * 100
print("Training accuracy: {:.2f}%".format(train_acc))
print("Calculating testing accuracy...")
test_acc = classifier.score(X_test, y_test) * 100
print("Testing accuracy: {:.2f}%".format(test_acc))
end3 = timer()
totalTime = end3 - start1
print("Total time taken for logistic regression = {:.1f} seconds".format(totalTime))
if not os.path.isfile("LogisticRegression.csv"):
with open("LogisticRegression.csv", "w") as _file:
_file.write("Solver, ImageSize, Rescaled, TrainAcc, ValidateAcc, TimeToFit, TotalTime\n")
with open("LogisticRegression.csv", "a") as _file:
_file.write("{}, {}, {}, {:0.3f}, {:0.3f}, {:0.2f}, {:0.2f}\n".format(args.solver.value, args.imageSize, args.rescale, train_acc, test_acc, timeToFit, totalTime))
def loadDataSet(imgSize, rescale):
dataSetFileName = "DogsCats_{}{}.npz".format(imgSize, "_Rescaled" if rescale else "")
if os.path.isfile(dataSetFileName):
print("Loading existing dataset")
Data = np.load(dataSetFileName)
X_train = Data["X_train"]
y_train = Data["y_train"].flatten()
X_test = Data["X_test"]
y_test = Data["y_test"].flatten()
print("Creating a new dataset")
X_train, y_train = loadImages(["data/train/**/*"] , imgSize, rescale, "training")
X_test , y_test = loadImages("data/validate/**/*", imgSize, rescale, "testing")
y_train = y_train.flatten()
y_test = y_test.flatten()
if rescale:
print("Rescaling images...")
print("Sum of feature-wise mean before rescaling: {}".format(np.mean(X_train, axis=0).sum()))
print("Sum of feature-wise std before rescaling: {}".format(np.std(X_train, axis=0).sum()))
scaler = preprocessing.StandardScaler().fit(X_train)
X_train = scaler.transform(X_train)
X_test = scaler.transform(X_test)
# Make sure processing is correct.
mean = abs(np.mean(X_train, axis=0).sum()/X_train.shape[1])
std = round(np.std(X_train, axis=0).sum()/X_train.shape[1])
print("Fature-wise mean after rescaling: {}".format(mean))
print("Feature-wise std after rescaling: {}".format(std))
if mean>1e-6 or std!=1.0:
print("Error normalizing images")
np.savez(dataSetFileName, X_train=X_train, y_train=y_train, X_test=X_test, y_test=y_test)
return X_train, y_train, X_test, y_test
def loadImages(filePath, imgSize, rescale, label):
fileNames = []
if isinstance(filePath, list):
for fp in filePath:
fileNames = glob.glob(filePath)
numImages = len(fileNames)
print("Number of {} images to load: {}".format(numImages, label))
X = np.zeros((numImages, imgSize*imgSize*3), dtype = np.float32 if rescale else np.int32)
Y = np.zeros((numImages, 1), dtype = np.float32 if rescale else np.int32)
for i in range(0, numImages):
image =[i])
image = image.resize((imgSize, imgSize), resample=Image.LANCZOS)
image = image.convert("RGB")
if rescale:
image = np.asarray(image, dtype=np.float32) / 255.0
image = np.asarray(image, dtype=np.int32)
red = image[:, :, 0].reshape(-1)
green = image[:, :, 1].reshape(-1)
blue = image[:, :, 2].reshape(-1)
image = np.vstack((red, green, blue)).reshape((-1,), order="F") # Interleave three channels.
X[i, :] = image
if "dog" in fileNames[i].lower():
Y[i, 0] = 1
elif "cat" in fileNames[i].lower() :
Y[i, 0] = 0
print("Unknown label")
if i%1000 == 0:
print("Processed {} of {}".format(i, numImages))
return X, Y
if __name__ == "__main__":