def get_feature_map(idx):
model = resnet18Conv5().cuda()
mnist = MNIST(download=True, train=True, root=".").train_data.float()
data_transform = Compose([
Resize((224, 224)),
ToTensor(),
Normalize((mnist.mean() / 255, ), (mnist.std() / 255, ))
])
mnist_dataset = MNIST(download=True,
root=".",
transform=data_transform,
train=True)
print(len(mnist))
data = mnist_dataset[idx]
print("data shape = ", data[0].shape) # data shape 1x224x224\
print("data label = ", data[1]) # label number
X = data[0].cuda()
X = X.unsqueeze(0)
outputs = model(X)
outputs = outputs.detach().cpu().numpy()
print(outputs.shape)
return outputs
def get_data_loaders(train_batch_size=256, val_batch_size=64):
mnist = MNIST(download=True, train=True, root=".").train_data.float()
# Ugly method to hardcode convert 1 channel grayscale to 3 channel RGB-like
# but it is still gray input (due to MNIST from torchvision)
data_transform = Compose([
Resize((224, 224)),
Grayscale(3),
ToTensor(),
Normalize((mnist.mean() / 255, ), (mnist.std() / 255, ))
])
train_loader = DataLoader(MNIST(download=True,
root=".",
transform=data_transform,
train=True),
batch_size=train_batch_size,
shuffle=True)
val_loader = DataLoader(MNIST(download=True,
root=".",
transform=data_transform,
train=False),
batch_size=val_batch_size,
shuffle=False)
return train_loader, val_loader
def get_data_loaders(train_batch_size, val_batch_size):
mnist = MNIST(download=False, train=True, root=".").train_data.float()
data_transform = Compose([ Resize((224, 224)),ToTensor(), Normalize((mnist.mean()/255,), (mnist.std()/255,))])
train_loader = DataLoader(MnistDataset(root_dir=''),
batch_size=train_batch_size, shuffle=True)
val_loader = DataLoader(MnistDataset(root_dir='', training=False),
batch_size=val_batch_size, shuffle=False)
return train_loader, val_loader
def get_feature_map(idx):
model = resnet18fc1000().cuda()
mnist = MNIST(download=True, train=True, root=".").train_data.float()
data_transform = Compose([ Resize((224, 224)),ToTensor(), Normalize((mnist.mean()/255,), (mnist.std()/255,))])
mnist_dataset = MNIST(download=True, root=".", transform=data_transform, train=True)
data = mnist_dataset[idx]
X = data[0].cuda()
X = X.unsqueeze(0)
outputs = model(X)
return outputs
def get_data_loaders(train_batch_size, val_batch_size):
mnist = MNIST(download=False, train=True, root="./raw/").train_data.float()
train_data_transform = Compose([
RandomAffine(degrees=30,
translate=(0.15, 0.15),
scale=(0.9, 1.1),
shear=10),
Resize((48, 48)),
ToTensor(),
Normalize((mnist.mean() / 255, ), (mnist.std() / 255, ))
])
valid_data_transform = Compose([
#RandomAffine(degrees=6, translate=(0.15, 0.15)),
Resize((48, 48)),
ToTensor(),
Normalize((mnist.mean() / 255, ), (mnist.std() / 255, ))
])
train_loader = DataLoader(MNIST(download=False,
root=".",
transform=train_data_transform,
train=True),
batch_size=train_batch_size,
shuffle=True,
num_workers=4)
val_loader = DataLoader(MNIST(download=False,
root=".",
transform=valid_data_transform,
train=False),
batch_size=val_batch_size,
shuffle=False,
num_workers=4)
return train_loader, val_loader
import torch import matplotlib.pyplot as plt from biolayer import BioLinear from visualization import LinearLayerVisualizer from torchvision.datasets import MNIST MNIST_DIR = '~/DataSets/' Nc = 10 # num. of classes N = 784 # Sample size Nep = 300 # Number of epochs Num = 100 # Batch size eps0 = 2e-3 # Learning rate Kx=5 Ky=5 hid=Kx*Ky # number of hidden units that are displayed in Ky by Kx array M = MNIST(MNIST_DIR, download=True).data.detach().view(-1, 28*28).float() M -= M.mean(0) bio_linear = BioLinear(N, hid) vis = LinearLayerVisualizer(bio_linear, as_heatmap=True) try: for weight in bio_linear.train(M, batch_size=Num, epsilon=eps0): vis.update() except KeyboardInterrupt: vis.close()
import torch
### Load dataset
from torchvision.datasets import MNIST
from resnet_vis import *
import os
from sklearn.multiclass import OneVsRestClassifier
from sklearn.model_selection import train_test_split, GridSearchCV
#os.system('nvidia-smi -q -d Memory |grep -A4 GPU|grep Free >tmp')
#os.environ['CUDA_VISIBLE_DEVICES'] = str(np.argmax([int(x.split()[2]) for x in open('tmp', 'r').readlines()]))
#os.system('rm tmp')
mnist_train = MNIST(download=True, train=True, root=".").train_data.float()
data_transform = Compose([
Resize((224, 224)),
ToTensor(),
Normalize((mnist_train.mean() / 255, ), (mnist_train.std() / 255, ))
])
mnist_dataset_train = MNIST(download=True,
root=".",
transform=data_transform,
train=True)
mnist_dataset_test = MNIST(download=True,
train=False,
root=".",
transform=data_transform)
#### Load ResNet 18 model to get features
resnet_model = resnet18Conv5().cuda()
### Generate train features
class MyModel:
def __init__(self):
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
self.model = MnistResNet().to(self.device)
try:
self.model.load_state_dict(
torch.load("./model/mnist_model_46_36.pth",
map_location="cpu"))
except:
self.model.load_state_dict(
torch.load("mnist_model_46_36.pth", map_location="cpu"))
self.model.eval()
self.mnist = MNIST(download=True, train=True,
root=".").train_data.float()
self.test_transforms = Compose([
Resize(((46, 46))),
ToTensor(),
Normalize((self.mnist.mean() / 255, ), (self.mnist.std() / 255, ))
])
self.BINARY_THREHOLD = 180
def finetune(self):
for param in self.model.parameters():
param.require_grad = False
print(self.model)
def predict(self, image):
with torch.no_grad():
self.model.eval()
image_tensor = self.test_transforms(image).float()
image_tensor = image_tensor.unsqueeze_(0)
input_ = Variable(image_tensor)
input_ = input_.to(self.device)
output = self.model(input_)
index = output.data.cpu().numpy().argmax()
return index
def image_smoothening(self, img):
ret1, th1 = cv2.threshold(img, self.BINARY_THREHOLD, 255,
cv2.THRESH_BINARY)
ret2, th2 = cv2.threshold(th1, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)
blur = cv2.GaussianBlur(th2, (1, 1), 0)
ret3, th3 = cv2.threshold(blur, 0, 255,
cv2.THRESH_BINARY + cv2.THRESH_OTSU)
return th3
def remove_noise_and_smooth(self, file_name):
img = cv2.imread(file_name, 0)
h, w = img.shape
#img[img <128] = 0
img = img[int(w * 0.2):int(w * 0.9), int(h * 0.2):int(h * 0.9)]
filtered = cv2.adaptiveThreshold(img.astype(np.uint8), 255,
cv2.ADAPTIVE_THRESH_MEAN_C,
cv2.THRESH_BINARY, 9, 41)
kernel = np.ones((1, 1), np.uint8)
opening = cv2.morphologyEx(filtered, cv2.MORPH_OPEN, kernel)
closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel)
img = self.image_smoothening(img)
or_image = cv2.bitwise_or(img, closing)
return or_image
def remove_noise_and_smooth_numpy(self, image_np):
img = cv2.cvtColor(image_np, cv2.COLOR_BGR2GRAY)
h, w = img.shape
#img[img <128] = 0
img = img[int(w * 0.2):int(w * 0.9), int(h * 0.2):int(h * 0.9)]
filtered = cv2.adaptiveThreshold(img.astype(np.uint8), 255,
cv2.ADAPTIVE_THRESH_MEAN_C,
cv2.THRESH_BINARY, 9, 41)
kernel = np.ones((1, 1), np.uint8)
opening = cv2.morphologyEx(filtered, cv2.MORPH_OPEN, kernel)
closing = cv2.morphologyEx(opening, cv2.MORPH_CLOSE, kernel)
img = self.image_smoothening(img)
or_image = cv2.bitwise_or(img, closing)
return or_image
def preprocess_image(self, image):
image = cv2.GaussianBlur(image, (7, 7), 0)
h, w, c = image.shape
image[image < 128] = 0
# Convert BGR to HSV
hsv = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
lower_val = np.array([110, 25, 50])
upper_val = np.array([255, 255, 255])
# Threshold the HSV image to get only black colors
mask = cv2.inRange(hsv, lower_val, upper_val)
# Bitwise-AND mask and original image
image = cv2.bitwise_and(image, image, mask=mask)
# invert the mask to get black letters on white background
image = cv2.bitwise_not(image)
#image = image[int(w*0.2): int(w*0.9),int(h*0.2): int(h*0.8)]
#rimage = cv2.resize(image,(28,28))
#rimage= np.moveaxis(rimage, 2, 0)
#resized_image = np.reshape(rimage,(28,28,1))
im_gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# rimage = np.expand_dims(im_gray, axis=0)
# rimage= np.rollaxis(rimage, 3, 1)
#
return im_gray
def main(self):
file_in = "211851-1-2 copy.jpg"
#image = cv2.imread(file_in)
image = self.remove_noise_and_smooth(file_in)
#image = preprocess_image(image)
cv2.imwrite("test.jpg", image)
to_pil = transforms.ToPILImage()
pil_image = to_pil(image)
index = predict(pil_image)
print(index)
def main_prediction(self, image_np):
image_np = self.remove_noise_and_smooth_numpy(image_np)
to_pil = transforms.ToPILImage()
pil_image = to_pil(image_np)
index = self.predict(pil_image)
return index