class Inferencer:
def __init__(self, path, small):
self.net = Net()
self.path = path
self.small = small
self.net.load_state_dict(torch.load(path))
def change_format(self, board):
np_onehot = self.to_one_hot(board)
tensor_onthot = torch.from_numpy(np_onehot).type(torch.float32)
return tensor_onthot
def inference(self, board):
with torch.no_grad():
tensor = self.change_format(board)
output = self.net(tensor)
_, predicted = torch.max(output.data, 1)
return self.transform(predicted.item())
def transform(self, value):
if (self.small):
return value * 3
else:
return value * 30
def to_one_hot(self, board):
onehot = np.zeros((1, NUM_OF_COLOR + 1, ROW_DIM, COLUMN_DIM))
for row in range(ROW_DIM):
for col in range(COLUMN_DIM):
color = board[row][col]
onehot[0, color, row, col] = 1
return onehot
def train(train_files,
test_files,
train_batch_size,
eval_batch_size,
model_file,
vocab_size,
num_classes,
n_epoch,
print_every=50,
eval_every=500):
torch.multiprocessing.set_sharing_strategy('file_system')
torch.backends.cudnn.benchmark = True
print "Setting seed..."
seed = 1234
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
# setup CNN model
CONFIG["vocab_size"] = vocab_size
CONFIG["num_classes"] = num_classes
model = Net()
if torch.cuda.is_available():
print "CUDA is available on this machine. Moving model to GPU..."
model.cuda()
print model
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters())
train_set = HDF5Dataset(train_files)
test_set = HDF5Dataset(test_files)
train_loader = DataLoader(dataset=train_set,
batch_size=train_batch_size,
shuffle=True,
num_workers=2)
test_loader = DataLoader(dataset=test_set,
batch_size=eval_batch_size,
num_workers=2)
_train_loop(train_loader=train_loader,
test_loader=test_loader,
model=model,
criterion=criterion,
optimizer=optimizer,
n_epoch=n_epoch,
print_every=print_every,
eval_every=eval_every,
model_file=model_file)
def setUp(self):
dataset, classes = prepare_cifar10()
self.models_path = "./models"
self.trainer = SimpleTrainer(pruning=True,
datasets=dataset,
models_path=self.models_path)
experiment = ax.Experiment(
name="model_test",
search_space=search_space(),
)
sobol = ax.Models.SOBOL(search_space=experiment.search_space)
self.param = sobol.gen(1).arms[0].parameters
self.net = Net(self.param, classes=classes, input_shape=(3, 32, 32))
if not os.path.exists(self.models_path):
os.mkdir(self.models_path)
class TrainingTestCase(unittest.TestCase):
""" Test cases to see if we can build a model,
train it, evaluate, quantize it and prune it"""
def setUp(self):
dataset, classes = prepare_cifar10()
self.models_path = "./models"
self.trainer = SimpleTrainer(pruning=True,
datasets=dataset,
models_path=self.models_path)
experiment = ax.Experiment(
name="model_test",
search_space=search_space(),
)
sobol = ax.Models.SOBOL(search_space=experiment.search_space)
self.param = sobol.gen(1).arms[0].parameters
self.net = Net(self.param, classes=classes, input_shape=(3, 32, 32))
if not os.path.exists(self.models_path):
os.mkdir(self.models_path)
def tearDown(self):
shutil.rmtree(self.models_path)
shutil.rmtree("./data")
def test_training(self):
self.trainer.load_dataloaders(batch_size=4, collate_fn=None)
self.trainer.train(self.net,
self.param,
name="0",
epochs=1,
reload=False,
old_net=None)
self.trainer.evaluate(self.net.to('cpu'), quant_mode=False)
from cnn import Net
import data_processing
from data_processing import process_data
import torch
import torch.nn as nn
import torch.optim as optim
from tqdm import tqdm
train, test, val = process_data()
device = torch.device('cuda:0')
net = Net().to(device)
def train_model(net):
X = torch.Tensor([i[0] for i in train]).view(-1, 98, 98).to(device)
# Normalize data
X = X / 255.0
y = torch.Tensor([i[1] for i in train]).to(device)
optimizer = optim.Adam(net.parameters(), lr=0.001)
loss_function = nn.MSELoss()
epochs = 2
batch_size = 32
for epoch in tqdm(range(epochs)):
running_loss = 0.0
for i in range(0, len(X), batch_size):
batch_X = X[i: i+batch_size].view(-1, 1, 98, 98).to(device)
batch_y = y[i: i+batch_size].to(device)
def main():
# load images as a numpy array
train_dataset = np.array(
np.load('/content/drive/My Drive/McGill/comp551/data/train_max_x',
allow_pickle=True))
train_dataset = train_dataset / 255.0
train_dataset = train_dataset.astype('float32')
targets = pd.read_csv(
'/content/drive/My Drive/McGill/comp551/data/train_max_y.csv',
delimiter=',',
skipinitialspace=True)
targets = targets.to_numpy()
# remove id column
targets = targets[:, 1]
targets = targets.astype(int)
X_train, X_test, y_train, y_test = train_test_split(train_dataset,
targets,
test_size=0.2,
random_state=42)
# Clean memory
train_dataset = None
# converting training images into torch format
dim1, dim2, dim3 = X_train.shape
X_train = X_train.reshape(dim1, 1, dim2, dim3)
X_train = torch.from_numpy(X_train)
y_train = torch.from_numpy(y_train)
# converting validation images into torch format
dim1, dim2, dim3 = X_test.shape
X_test = X_test.reshape(dim1, 1, dim2, dim3)
X_test = torch.from_numpy(X_test)
y_test = torch.from_numpy(y_test)
# defining the model
model = Net()
criterion = nn.NLLLoss()
optimizer = optim.SGD(model.parameters(), lr=0.003, momentum=0.9)
if torch.cuda.is_available():
model = model.cuda()
criterion = criterion.cuda()
print(model)
time0 = time()
epochs = 1
for e in range(epochs):
model.train()
running_loss = 0
x_train, y_train = Variable(X_train).cuda(), Variable(y_train).cuda()
x_val, y_val = Variable(X_test).cuda(), Variable(y_test).cuda()
# converting the data into GPU format
# if torch.cuda.is_available():
# x_train = x_train.cuda()
# y_train = y_train.cuda()
# x_val = x_val.cuda()
# y_val = y_val.cuda()
# clearing the Gradients of the model parameters
optimizer.zero_grad()
# prediction for training and validation set
output_train = model(x_train)
output_val = model(x_val)
# computing the training and validation loss
loss_train = criterion(output_train, y_train)
loss_val = criterion(output_val, y_val)
# computing the updated weights of all the model parameters
loss_train.backward()
# And optimizes its weights here
optimizer.step()
running_loss += loss_train.item()
print("Epoch {} - Training loss: {}".format(
e, running_loss / len(train_dataset)))
print("\nTraining Time (in minutes) =", (time() - time0) / 60)
# prediction for validation set
with torch.no_grad():
output = model(x_val.cuda())
ps = torch.exp(output).cpu()
probab = list(ps.numpy())
predictions = np.argmax(probab, axis=1)
# accuracy on validation set
print("\nModel Accuracy =", (accuracy_score(y_val, predictions)))
np.shape(images))
# hyperparameters (start with r = for reducing layer, start with n = for nonreducing layer)
lr_rate = float(param["learning_rate"])
rKern = int(param["reducing_conv_kernel"])
nKern = int(param["nonreducing_conv_kernel"])
rStride = int(param["reducing_stride"])
nStride = int(param["nonreducing_stride"])
rOut = int(param["reducing_out"])
nOut = int(param["nonreducing_out"])
# constructing a model (converting model to double precision)
model = Net(red_kernel=rKern,
nonred_kernel=nKern,
red_stride=rStride,
nonred_stride=nStride,
red_out=rOut,
nonred_out=nOut,
d="cuda:0" if enableCuda else "cpu")
model.double()
if enableCuda:
model.cuda()
# declare optimizer and gradient and loss function
optimizer = optim.Adadelta(model.parameters(), lr=lr_rate)
loss = torch.nn.MSELoss(reduction='mean')
print("Loading model")
model.load_state_dict(torch.load(modelFile))
elif args.log_level == "critical" or args.log_level == "CRITICAL":
logger.setLevel(logging.CRITICAL)
for arg in vars(args):
logger.info("{}: {}".format(arg, getattr(args, arg)))
##
torch.backends.cudnn.enabled = False
momentum = 0.5
learning_rate = 0.01
n_epochs = 3
training_data, testing_data, example_data, example_targets = getData()
# Save
network = Net()
optimizer = optim.SGD(network.parameters(),
lr=learning_rate,
momentum=momentum)
loss = 0
test(network, testing_data)
for epoch in range(1, n_epochs + 1):
train(epoch, network, optimizer, loss, training_data)
test(network, testing_data)
# Load
continued_network = Net()
continued_optimizer = optim.SGD(network.parameters(),
lr=learning_rate,
momentum=momentum)
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import torch.optim as optim
from cnn import Net
net = Net()
PATH = './cifar_net.pth'
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse',
'ship', 'truck')
testset = torchvision.datasets.CIFAR10(root='./data',
train=False,
download=True,
transform=transform)
testloader = torch.utils.data.DataLoader(testset,
batch_size=4,
shuffle=False,
num_workers=2)
# dataiter = iter(testloader)
# images, labels = dataiter.next()
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import torch.optim as optim
from cnn import Net
bs = 16
model_name = 'cnn_adaptive_7.pt'
#load model
model = Net()
model.load_state_dict(torch.load('models/' + model_name))
#data
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
testset = torchvision.datasets.CIFAR10(root='./data',
train=False,
download=False,
transform=transform)
testloader = torch.utils.data.DataLoader(testset,
batch_size=bs,
shuffle=False,
num_workers=2)
sample = trainDataset[0]
angle = 22.5
sample_x = transform(sample)
display(sample_x['image'], sample_x['keypoints'])
print(sample['keypoints'][0][1])
'''
trainLoader = DataLoader(trainDataset,
batch_size=batch_size,
shuffle=True,
num_workers=0)
testLoader = DataLoader(testDataset, batch_size=1, shuffle=True, num_workers=0)
model = Net()
critirion = torch.nn.SmoothL1Loss()
optimizer = optim.Adam(model.parameters(), lr=0.00001)
print(model)
epoch = 1000
model.cuda()
model, optimizer = loadModel(model, optimizer)
#train(epoch=epoch, train_loader=trainLoader, optimizer= optimizer, critirion=critirion, model=model, testLoader= testLoader, batch_size= batch_size)
torch.save(model.state_dict(), 'SavedModels/pull_model_saved')
#test(model=model, testLoader=testLoader, batch_size=batch_size, im_num=12)
#model = loadModel(model)
#weights = model.conv2.weight.data.numpy()
# function to show an image
def imshow(img):
img = img/2+0.5
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1,2,0)))
plt.show()
if __name__ == '__main__':
# dataiter = iter(trainloader)
# images, labels = dataiter.next()
# imshow(torchvision.utils.make_grid(images, padding=2))
# print(' '.join(classes[labels[j]] for j in range(4)))
bn = False
net = Net(bn)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if torch.cuda.device_count() > 1:
net = nn.DataParallel(net)
net.to(device)
# training the network
for epoch in range(2):
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
inputs, labels = data
inputs, labels = inputs.to(device), labels.to(device)
# hyperparameters (start with r = for reducing layer, start with n = for nonreducing layer)
lr_rate = float(param["learning_rate"])
rKern = int(param["reducing_conv_kernel"])
nKern = int(param["nonreducing_conv_kernel"])
rStride = int(param["reducing_stride"])
nStride = int(param["nonreducing_stride"])
rOut = int(param["reducing_out"])
nOut = int(param["nonreducing_out"])
# epoch hyperparameters
num_epochs = int(param["num_epochs_per_batch"])
disp_epochs = param["display_epochs"]
# constructing a model (converting model to double precision)
model = Net(red_kernel=rKern, nonred_kernel=nKern, red_stride=rStride,
nonred_stride=nStride, red_out=rOut, nonred_out=nOut,
d="cuda:0" if enableCuda else "cpu")
model.double()
if enableCuda:
model.cuda()
# declare optimizer and gradient and loss function
optimizer = optim.Adadelta(model.parameters(), lr=lr_rate)
loss = torch.nn.MSELoss(reduction='mean')
# storing the training and test loss values
obj_vals = []
#load test dataset
t_img, t_nrg = loadimg(imdir + "\\" + files[-1])
def capture_solve(input_image):
# Grayscale and adaptive threshold
img = cv2.GaussianBlur(input_image, (5, 5), 0)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
mask = np.zeros(gray.shape, np.uint8)
kernel1 = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (11, 11))
close = cv2.morphologyEx(gray, cv2.MORPH_CLOSE, kernel1)
div = np.float32(gray) / close
isolate = np.uint8(cv2.normalize(div, div, 0, 255, cv2.NORM_MINMAX))
res2 = cv2.cvtColor(isolate, cv2.COLOR_GRAY2BGR)
thresh = cv2.adaptiveThreshold(isolate, 255, 0, 1, 19, 2)
contour, hier = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
# Grab the biggest contour (the board)
max_area = 0
best_cnt = None
for cnt in contour:
area = cv2.contourArea(cnt)
if area > 50000:
if area > max_area:
max_area = area
best_cnt = cnt
# Isolate board
cv2.drawContours(mask, [best_cnt], 0, 255, -1)
cv2.drawContours(mask, [best_cnt], 0, 0, 2)
isolate = cv2.bitwise_and(isolate, mask)
# === Use second order derivative filter to find the vertical and horizontal lines === #
# Obtain vertical contours
kernel_x = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 10))
dx = cv2.Sobel(isolate, cv2.CV_16S, 1, 0)
dx = cv2.convertScaleAbs(dx)
cv2.normalize(dx, dx, 0, 255, cv2.NORM_MINMAX)
_, close = cv2.threshold(dx, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
close = cv2.morphologyEx(close, cv2.MORPH_DILATE, kernel_x, iterations=1)
# Remove number contours, isolating vertical lines
contour, _ = cv2.findContours(close, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
for cnt in contour:
x, y, w, h = cv2.boundingRect(cnt)
if h / w > 5:
cv2.drawContours(close, [cnt], 0, 255, -1)
else:
cv2.drawContours(close, [cnt], 0, 0, -1)
close_x = cv2.morphologyEx(close, cv2.MORPH_CLOSE, None, iterations=2)
# Obtain horizontal contours
kernel_y = cv2.getStructuringElement(cv2.MORPH_RECT, (10, 2))
dy = cv2.Sobel(isolate, cv2.CV_16S, 0, 2)
dy = cv2.convertScaleAbs(dy)
cv2.normalize(dy, dy, 0, 255, cv2.NORM_MINMAX)
ret, close = cv2.threshold(dy, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
close = cv2.morphologyEx(close, cv2.MORPH_DILATE, kernel_y)
# Remove number contours, isolating vertical lines
contour, _ = cv2.findContours(close, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
for cnt in contour:
x, y, w, h = cv2.boundingRect(cnt)
if w / h > 5:
cv2.drawContours(close, [cnt], 0, 255, -1)
else:
cv2.drawContours(close, [cnt], 0, 0, -1)
close_y = cv2.morphologyEx(close, cv2.MORPH_DILATE, None, iterations=2)
# Get intersections
isolate = cv2.bitwise_and(close_x, close_y)
# Obtain centroids
contour, _ = cv2.findContours(isolate, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
centroids = []
for cnt in contour:
mom = cv2.moments(cnt)
(x, y) = int(mom['m10'] / mom['m00']), int(mom['m01'] / mom['m00'])
cv2.circle(img, (x, y), 4, (0, 255, 0), -1)
centroids.append((x, y))
# Re-order centroids
centroids = np.array(centroids, dtype=np.float32)
c = centroids.reshape((100, 2))
c2 = c[np.argsort(c[:, 1])]
b = np.vstack([c2[i * 10:(i + 1) * 10][np.argsort(c2[i * 10:(i + 1) * 10, 0])] for i in range(10)])
bm = b.reshape((10, 10, 2))
# Apply perspective transform and warp to 450x450
image = np.zeros((450, 450, 3), np.uint8)
for i, j in enumerate(b):
ri = i // 10
ci = i % 10
if ci != 9 and ri != 9:
src = bm[ri:ri + 2, ci:ci + 2, :].reshape((4, 2))
dst = np.array([[ci * 50, ri * 50],
[(ci + 1) * 50 - 1, ri * 50],
[ci * 50, (ri + 1) * 50 - 1],
[(ci + 1) * 50 - 1, (ri + 1) * 50 - 1]], np.float32)
retval = cv2.getPerspectiveTransform(src, dst)
warp = cv2.warpPerspective(res2, retval, (450, 450))
image[
ri * 50:(ri + 1) * 50 - 1,
ci * 50:(ci + 1) * 50 - 1
] = warp[ri * 50:(ri + 1) * 50 - 1, ci * 50:(ci + 1) * 50 - 1].copy()
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
thresh = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 57, 5)
# Filter out all numbers and noise to isolate boxes
cnts = cv2.findContours(thresh, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
for c in cnts:
area = cv2.contourArea(c)
if area < 1000:
cv2.drawContours(thresh, [c], -1, (0, 0, 0), -1)
# Fix horizontal and vertical lines
vertical_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 5))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, vertical_kernel, iterations=9)
horizontal_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 1))
thresh = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, horizontal_kernel, iterations=4)
invert = 255 - thresh
cnts = cv2.findContours(invert, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
cnts, _ = contours.sort_contours(cnts, method="top-to-bottom")
sudoku_rows = []
row = []
for (i, c) in enumerate(cnts, 1):
area = cv2.contourArea(c)
if area < 50000:
row.append(c)
if i % 9 == 0:
(cnts, _) = contours.sort_contours(row, method="left-to-right")
sudoku_rows.append(cnts)
row = []
squares = []
for row in sudoku_rows:
for cc in row:
x, y, w, h = cv2.boundingRect(cc)
squares.append(image[y:y + h, x:x + w])
squares = np.reshape(squares, (-1, 9))
board = np.zeros((9, 9))
mean_list = []
for x in squares:
for box in x:
gray = cv2.cvtColor(box, cv2.COLOR_BGR2GRAY)
img = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 57, 5)
denoise = ndimage.median_filter(img, 5)
white_pix_count = np.count_nonzero(denoise)
mean_list.append(white_pix_count)
else:
mean = np.mean(mean_list)
for col, x in enumerate(squares):
for row, box in enumerate(x):
gray = cv2.cvtColor(box, cv2.COLOR_BGR2GRAY)
img = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 57, 5)
denoise = ndimage.median_filter(img, 5)
white_pix_count = np.count_nonzero(denoise)
if white_pix_count > mean:
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Net()
model.load_state_dict(torch.load('model.pth', map_location=torch.device(device)))
model.eval()
p = transforms.Compose([transforms.Resize((28, 28)),
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))])
img = Image.fromarray(denoise)
img = p(img)
img = img.view(1, 28, 28)
img = img.unsqueeze(0)
output = model(img)
prediction = list(output.cpu()[0])
board[col][row] = prediction.index(max(prediction))
solve(board)
return board
def __init__(self, path, small):
self.net = Net()
self.path = path
self.small = small
self.net.load_state_dict(torch.load(path))
def __execute(self, model: cnn.Net, image_paths: List[str]):
processed_data = {
'vanilla': [],
'deconv': [],
'gbp': [],
'gcam': [],
'ggcam': [],
}
device = next(model.parameters()).device
model.eval()
images, raw_images = load_images(image_paths, self.input_size)
images = torch.stack(images).to(device)
cls_num = len(self.classes)
save_dir = Path(self.save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
# --- Vanilla Backpropagation ---
bp = BackPropagation(model=model)
probs, ids = bp.forward(images) # sorted
# --- Deconvolution ---
deconv = None
if self.is_deconv:
deconv = Deconvnet(model=model)
_ = deconv.forward(images)
# --- Grad-CAM / Guided Backpropagation / Guided Grad-CAM ---
gcam = None
gbp = None
if self.is_gradcam:
gcam = GradCAM(model=model)
_ = gcam.forward(images)
gbp = GuidedBackPropagation(model=model)
_ = gbp.forward(images)
# probs = probs.detach().cpu().numpy() # to numpy
# ids_np = ids.detach().cpu().numpy() # to numpy
pbar = tqdm(range(cls_num),
total=cls_num,
ncols=100,
bar_format='{l_bar}{bar:30}{r_bar}',
leave=False)
pbar.set_description('Grad-CAM')
for i in pbar:
if self.is_vanilla:
bp.backward(ids=ids[:, [i]])
gradients = bp.generate()
# Save results as image files
for j in range(len(images)):
# fmt = '%d-{}-%s.png' % (j, self.classes[ids_np[j, i]])
# print("\t#{}: {} ({:.5f})".format(j, classes[ids[j, i]], probs[j, i]))
# append
_grad = get_gradient_data(gradients[j])
processed_data['vanilla'].append(_grad)
# save as image
# _p = save_dir.joinpath(fmt.format('vanilla'))
# save_gradient(str(_p), gradients[j])
if self.is_deconv:
deconv.backward(ids=ids[:, [i]])
gradients = deconv.generate()
for j in range(len(images)):
# fmt = '%d-{}-%s.png' % (j, self.classes[ids_np[j, i]])
# print("\t#{}: {} ({:.5f})".format(j, classes[ids[j, i]], probs[j, i]))
# append
_grad = get_gradient_data(gradients[j])
processed_data['deconv'].append(_grad)
# save as image
# _p = save_dir.joinpath(fmt.format('deconvnet'))
# save_gradient(str(_p), gradients[j])
# Grad-CAM / Guided Grad-CAM / Guided Backpropagation
if self.is_gradcam:
gbp.backward(ids=ids[:, [i]])
gradients = gbp.generate()
# Grad-CAM
gcam.backward(ids=ids[:, [i]])
regions = gcam.generate(target_layer=self.target_layer)
for j in range(len(images)):
# fmt = '%d-{}-%s.png' % (j, self.classes[ids_np[j, i]])
# print("\t#{}: {} ({:.5f})".format(j, classes[ids[j, i]], probs[j, i]))
# append
_grad = get_gradient_data(gradients[j])
processed_data['gbp'].append(_grad)
_grad = get_gradcam_data(regions[j, 0], raw_images[j])
processed_data['gcam'].append(_grad)
_grad = get_gradient_data(torch.mul(regions, gradients)[j])
processed_data['ggcam'].append(_grad)
# save as image - Guided Backpropagation
# _p = save_dir.joinpath(fmt.format('guided-bp'))
# save_gradient(str(_p), gradients[j])
# save as image - Grad-CAM
# _p = save_dir.joinpath(fmt.format(f'gradcam-{self.target_layer}'))
# save_gradcam(str(_p), regions[j, 0], raw_images[j])
# save as image - Guided Grad-CAM
# _p = save_dir.joinpath(fmt.format(f'guided_gradcam-{self.target_layer}'))
# save_gradient(str(_p), torch.mul(regions, gradients)[j])
# Remove all the hook function in the 'model'
bp.remove_hook()
if self.is_deconv:
deconv.remove_hook()
if self.is_gradcam:
gcam.remove_hook()
gbp.remove_hook()
return processed_data
def model(input):
net = Net()
net = loadModel(net)
return net(input)
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='Visualization tool for CNN Pytorch model')
parser.add_argument('--use_cuda',
action='store_true',
default=False,
help='enables CUDA training')
parser.add_argument('--image', help='path to the image', required=True)
parser.add_argument('--image_cls',
help='cllass of the image',
required=True)
parser.add_argument('--load_weights',
help='path to saved model\'s file',
required=True)
parser.add_argument(
'--dataset',
help='Path to dataset from which to load the images for visualizatiom')
args = parser.parse_args()
use_cuda = args.use_cuda and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
model = Net().to(device)
model.load_state_dict(torch.load(args.load_weights))
model.eval()
image = np.array(Image.open(args.image)).astype(np.float32) / 255
image_tensor = torch.tensor(image).permute(2, 0, 1)
plot_pixelwise_gradients(model, image_tensor, args.image_cls)
plot_occlusion_heatmap(image_tensor, args.image_cls)
trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=4, shuffle=True, num_workers=2)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')
def imshow(img):
img = img / 2 + 0.5
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1,2,0)))
plt.show()
dataiter = iter(trainloader)
images, labels = dataiter.next()
net = Net()
# define loss function and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
for epoch in range(2):
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
inputs, labels = data
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
acc += partial_acc
#accuracy on whole test set
acc /= (nbs * bs)
print('Test accuracy on this epoch: {}'.format(acc))
test_acc_values.append(acc)
#save model
torch.save(net.state_dict(), 'models/' + model_name)
return lr_values, running_loss_values, train_acc_values, test_acc_values
if __name__ == '__main__':
net = Net()
print(net)
lr_values, running_loss_values, train_acc_values, test_acc_values = make_iterations(
net, lr_ini)
plt.plot(running_loss_values)
plt.ylim((0, 5))
plt.title('Loss over iterations')
plt.show()
plt.plot(lr_values)
plt.ylim((0, 20 * lr_ini))
plt.title('Learning rate over iterations')
plt.show()
def get_chars(path_name):
gcs = storage.Client()
bucket = gcs.get_bucket('mail-scanner-bucket')
blob = bucket.blob(path_name)
blob.download_to_filename('image.jpg')
image = cv2.imread('image.jpg')
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
edged = cv2.Canny(gray, 100, 200)
cnts = cv2.findContours(edged.copy(), cv2.RETR_LIST,
cv2.CHAIN_APPROX_SIMPLE)
cnts = imutils.grab_contours(cnts)
cnts = sorted(cnts, key=cv2.contourArea, reverse=True)[:5]
for c in cnts:
peri = cv2.arcLength(c, True)
approx = cv2.approxPolyDP(c, 0.02 * peri, True)
if len(approx) == 4:
screenCnt = approx
break
cv2.drawContours(image, [screenCnt], -1, (0, 255, 0), 2)
warped = four_point_transform(
image,
screenCnt.reshape(4, 2) * (image.shape[0] / 480.0))
warped = cv2.cvtColor(warped, cv2.COLOR_BGR2GRAY)
warped = cv2.adaptiveThreshold(warped, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, 17, 4)
image = imutils.resize(warped, width=500)
# Remove Salt and pepper noise
saltpep = cv2.fastNlMeansDenoising(image, None, 9, 13)
# blur
blured = cv2.blur(saltpep, (3, 3))
# binary
ret, thresh = cv2.threshold(gray, 127, 255, cv2.THRESH_BINARY_INV)
# dilation
kernel = np.ones((5, 100), np.uint8)
img_dilation = cv2.dilate(thresh, kernel, iterations=1)
# find contours
im2, ctrs, hier = cv2.findContours(img_dilation.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
# sort contours
sorted_ctrs = sorted(ctrs, key=lambda ctr: cv2.boundingRect(ctr)[1])
matrix = [[], [], []]
for i, ctr in enumerate(sorted_ctrs):
# Get bounding box
x, y, w, h = cv2.boundingRect(ctr)
# Getting ROI
roi = image[y:y + h, x:x + w]
im = cv2.resize(roi, None, fx=4, fy=4, interpolation=cv2.INTER_CUBIC)
ret_1, thresh_1 = cv2.threshold(im, 127, 255, cv2.THRESH_BINARY_INV)
im, ctrs_1, hier = cv2.findContours(thresh_1, cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
# sort contours
sorted_ctrs_1 = sorted(ctrs_1,
key=lambda ctr: cv2.boundingRect(ctr)[0])
count, p_height = 0, 0
for j, ctr_1 in enumerate(sorted_ctrs_1):
# Get bounding box
x_1, y_1, w_1, h_1 = cv2.boundingRect(ctr_1)
if w_1 > 50 and h_1 > 50:
# Getting ROI
roi_1 = thresh_1[y_1:y_1 + h_1, x_1:x_1 + w_1]
if j != 0:
# print(h_1, p_height)
if p_height < h_1:
count = 0
matrix[count].append(roi_1)
p_height = h_1
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = Net()
model.load_state_dict(
torch.load('letter_model.pt', map_location=torch.device(device)))
p = transforms.Compose([
transforms.Resize((28, 28)),
transforms.ToTensor(),
transforms.Normalize((0.5, ), (0.5, ))
])
real_pred = ''
for img in matrix[0]:
img = Image.fromarray(img)
img = p(img)
img = img.view(1, 28, 28)
img = img.unsqueeze(0)
img = model(img)
prediction = list(img.cpu().detach().numpy()[0])
classes = '0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
max_pred = prediction.index(max(prediction))
real_pred += classes[max_pred]
return real_pred
def main():
parser = argparse.ArgumentParser(description='Prediction of TCR binding to peptide-MHC complexes')
parser.add_argument('--infile', type=str,
help='input file for training')
parser.add_argument('--indepfile', type=str, default=None,
help='independent test file')
parser.add_argument('--blosum', type=str, default='data/BLOSUM50',
help='file with BLOSUM matrix')
parser.add_argument('--batch_size', type=int, default=50, metavar='N',
help='batch size')
parser.add_argument('--model_name', type=str, default='original.ckpt',
help = 'if train is True, model name to be saved, otherwise model name to be loaded')
parser.add_argument('--epoch', type = int, default=200, metavar='N',
help='number of epoch to train')
parser.add_argument('--lr', type=float, default=0.001, metavar='LR',
help='learning rate')
parser.add_argument('--cuda', type = str2bool, default=True,
help = 'enable cuda')
parser.add_argument('--seed', type=int, default=7405,
help='random seed')
parser.add_argument('--mode', default = 'train', type=str,
help = 'train or test')
parser.add_argument('--model', type=str, default='cnn',
help='cnn, resnet')
args = parser.parse_args()
if args.mode is 'test':
assert args.indepfile is not None, '--indepfile is missing!'
## cuda
if torch.cuda.is_available() and not args.cuda:
print("WARNING: You have a CUDA device, so you should probably run with --cuda")
args.cuda = (args.cuda and torch.cuda.is_available())
device = torch.device('cuda' if args.cuda else 'cpu')
## set random seed
seed = args.seed
torch.manual_seed(seed)
if args.cuda:
torch.cuda.manual_seed(seed) if args.cuda else None
# embedding matrix
embedding = load_embedding(args.blosum)
## read data
X_pep, X_tcr, y = data_io_tf.read_pTCR(args.infile)
y = np.array(y)
n_total = len(y)
n_train = int(round(n_total * 0.8))
n_valid = int(round(n_total * 0.1))
n_test = n_total - n_train - n_valid
idx_shuffled = np.arange(n_total); np.random.shuffle(idx_shuffled)
idx_train, idx_valid, idx_test = idx_shuffled[:n_train], \
idx_shuffled[n_train:(n_train+n_valid)], \
idx_shuffled[(n_train+n_valid):]
## define dataloader
train_loader = define_dataloader(X_pep[idx_train], X_tcr[idx_train], y[idx_train], None,
None, None,
batch_size=args.batch_size, device=device)
valid_loader = define_dataloader(X_pep[idx_valid], X_tcr[idx_valid], y[idx_valid], None,
maxlen_pep=train_loader['pep_length'],
maxlen_tcr=train_loader['tcr_length'],
batch_size=args.batch_size, device=device)
test_loader = define_dataloader(X_pep[idx_test], X_tcr[idx_test], y[idx_test], None,
maxlen_pep=train_loader['pep_length'],
maxlen_tcr=train_loader['tcr_length'],
batch_size=args.batch_size, device=device)
## read indep data
if args.indepfile is not None:
X_indep_pep, X_indep_tcr, y_indep = data_io_tf.read_pTCR(args.indepfile)
y_indep = np.array(y_indep)
indep_loader = define_dataloader(X_indep_pep, X_indep_tcr, y_indep, None,
maxlen_pep=train_loader['pep_length'],
maxlen_tcr=train_loader['tcr_length'],
batch_size=args.batch_size, device=device)
if args.model == 'cnn':
from cnn import Net
#if args.model == 'resnet':
#
# from resnet import Net
# Net = models.resnet18
else:
raise ValueError('unknown model name')
## define model
model = Net(embedding, train_loader['pep_length'], train_loader['tcr_length']).to(device)
optimizer = optim.Adam(model.parameters(), lr=args.lr)
if 'models' not in os.listdir('.'):
os.mkdir('models')
if 'result' not in os.listdir('.'):
os.mkdir('result')
## fit model
if args.mode == 'train' :
model_name = check_model_name(args.model_name)
model_name = check_model_name(model_name, './models')
model_name = args.model_name
wf_open = open('result/'+os.path.splitext(os.path.basename(args.infile))[0]+'_'+os.path.splitext(os.path.basename(args.model_name))[0]+'_valid.csv', 'w')
wf_colnames = ['loss', 'accuracy',
'precision1', 'precision0',
'recall1', 'recall0',
'f1macro','f1micro', 'auc']
wf = csv.DictWriter(wf_open, wf_colnames, delimiter='\t')
t0 = time.time()
for epoch in range(1, args.epoch + 1):
train(args, model, device, train_loader['loader'], optimizer, epoch)
## evaluate performance
perf_train = get_performance_batchiter(train_loader['loader'], model, device)
perf_valid = get_performance_batchiter(valid_loader['loader'], model, device)
## print performance
print('Epoch {} TimeSince {}\n'.format(epoch, timeSince(t0)))
print('[TRAIN] {} ----------------'.format(epoch))
print_performance(perf_train)
print('[VALID] {} ----------------'.format(epoch))
print_performance(perf_valid, writeif=True, wf=wf)
## evaluate and print test-set performance
print('[TEST ] {} ----------------'.format(epoch))
perf_test = get_performance_batchiter(test_loader['loader'], model, device)
print_performance(perf_test)
model_name = './models/' + model_name
torch.save(model.state_dict(), model_name)
elif args.mode == 'test' :
model_name = args.model_name
assert model_name in os.listdir('./models')
model_name = './models/' + model_name
model.load_state_dict(torch.load(model_name))
## evaluate and print independent-test-set performance
print('[INDEP] {} ----------------')
perf_indep = get_performance_batchiter(indep_loader['loader'], model, device)
print_performance(perf_indep)
## write blackbox output
wf_bb_open = open('data/testblackboxpred_' + os.path.basename(args.indepfile), 'w')
wf_bb = csv.writer(wf_bb_open, delimiter='\t')
write_blackbox_output_batchiter(indep_loader, model, wf_bb, device)
wf_bb_open1 = open('data/testblackboxpredscore_' + os.path.basename(args.indepfile), 'w')
wf_bb1 = csv.writer(wf_bb_open1, delimiter='\t')
write_blackbox_output_batchiter(indep_loader, model, wf_bb1, device, ifscore=True)
else :
print('\nError: "--mode train" or "--mode test" expected')
def main():
try:
args, check = argvcontrol()
if check:
if platform.system() == "Windows" or platform.system() == "win32":
args.model = os.path.abspath(".") + "\\" + args.model
args.path = os.path.abspath(".") + "\\" + args.path
else:
args.model = os.path.abspath(".") + "/" + args.model
args.path = os.path.abspath(".") + "/" + args.path
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog',
'horse', 'ship', 'truck')
net = Net()
if args.cuda:
net.cuda()
if args.model and args.training:
if not os.path.exists(args.model):
print("Model will be created")
else:
loadModel(net, args.model)
elif args.model and not args.training:
if os.path.exists(args.model):
loadModel(net, args.model)
else:
print(
"Invalid model checkpoint file. Please enter a valid path."
)
test_transform, train_transform = getTransformtions()
if args.training:
trainloader, testloader, criterion = CIFAR10Init(
args.cuda, args.path, int(args.batch_size),
int(args.workers))
if trainloader and testloader:
frequency = int(len(trainloader) / 4)
training(net, args.cuda, int(args.epochs),
trainloader, frequency, criterion,
float(args.learning_rate), int(args.batch_size),
int(args.workers), args.model)
else:
exit()
elif args.image:
_class = recognition(net, args.image, test_transform, classes)
if _class:
print("\nClassification: %s\n" % (_class))
else:
trainloader, testloader, criterion = CIFAR10Init(
args.cuda, args.path, int(args.batch_size),
int(args.workers))
if trainloader and testloader:
validation(net, args.cuda, testloader, classes, args.model,
int(args.batch_size))
else:
exit()
else:
print("\nTraining: python image-recognizer.py --train")
print("Validation: python image-recognizer.py --no-train")
print(
"Recognition: python image-recognizer.py --image PATH_TO_IMAGE"
)
print(
"You can set parameters for this operations. Add --help for more informations.\n"
)
except (KeyboardInterrupt, SystemExit):
pass
import torch
import torch.onnx
from cnn import Net
#TODO Get this working
# https://pytorch.org/tutorials/advanced/super_resolution_with_caffe2.html
# Online web service that enhances photography.
# A model class instance (class not shown)
model = Net()
model.train(False)
# Load the weights from a file (.pth usually)
state_dict = torch.load("model.pth")
# Load the weights now into a model net architecture defined by our class
print("loading model")
model.load_state_dict(state_dict['model_state_dict'])
# Input to the model
x = torch.randn(10, 3, 224, 224, requires_grad=True)
# Export the model
torch_out = torch.onnx._export(
model, # model being run
x, # model input (or a tuple for multiple inputs)
"model.onnx", # where to save the model (can be a file or file-like object)
export_params=True
) # store the trained parameter weights inside the model file
print("exported model")
acc += partial_acc
#accuracy on whole test set
acc /= (nbs * bs)
print('Test accuracy on this epoch: {}'.format(acc))
test_acc_values.append(acc)
#save model
torch.save(net.state_dict(), 'models/' + model_name)
return lr_values, running_loss_values, train_acc_values, test_acc_values
if __name__ == '__main__':
net = Net()
net.cuda()
print(net)
lr_values, running_loss_values, train_acc_values, test_acc_values = make_iterations(
net, lr_ini)
plt.plot(running_loss_values)
plt.ylim((0, 5))
plt.title('Loss over iterations')
plt.show()
plt.plot(lr_values)
plt.ylim((0, 20 * lr_ini))
plt.title('Learning rate over iterations')
def main(model: cnn.Net, classes: List[str], input_size: Tuple[int, int]):
# print("Mode:", ctx.invoked_subcommand)
# classes = ['crossing', 'klaxon', 'noise']
# input_size = (60, 60)
# model = cnn.Net(input_size)
device = next(model.parameters()).device
# model.to(device)
model.eval()
image_paths = [
'./recognition_datasets/Images/crossing/crossing-samp1_3_4.jpg',
'./recognition_datasets/Images/crossing/crossing-samp1_3_3.jpg'
]
images, raw_images = load_images(image_paths, input_size)
images = torch.stack(images).to(device)
bp = BackPropagation(model=model)
probs, ids = bp.forward(images) # sorted
ids = ids.cpu().numpy() # numpy
gcam = GradCAM(model=model)
_ = gcam.forward(images)
gbp = GuidedBackPropagation(model=model)
_ = gbp.forward(images)
topk = 3
target_layer = 'conv5'
output_dir = Path('results')
output_dir.mkdir(parents=True, exist_ok=True)
for i in range(topk):
# Guided Backpropagation
gbp.backward(ids=ids[:, [i]])
gradients = gbp.generate()
# Grad-CAM
gcam.backward(ids=ids[:, [i]])
regions = gcam.generate(target_layer=target_layer)
for j in range(len(images)):
name_fmt = f'{j}-' + '{}' + f'-{classes[ids[j, i]]}.png'
print("\t#{}: {} ({:.5f})".format(j, classes[ids[j, i]], probs[j,
i]))
# Guided Backpropagation
path = output_dir.joinpath(name_fmt.format('guided')).as_posix()
save_gradient(filename=path, gradient=gradients[j])
# Grad-CAM
path = Path(output_dir,
name_fmt.format(f'gradcam-{target_layer}')).as_posix()
save_gradcam(filename=path,
gcam=regions[j, 0],
raw_image=raw_images[j])
# Guided Grad-CAM
path = Path(
output_dir,
name_fmt.format(f'guided_gradcam-{target_layer}')).as_posix()
save_gradient(filename=path,
gradient=torch.mul(regions, gradients)[j])