batch_size=batchSize,
shuffle=False,
num_workers=0)
classes = ('normal', 'TB')
net = Net() #network1 calling, comment this line to test other network.
#net = Net1() #network2 calling, un-comment this line to test network2.
#net = NewNet() #network3 calling, un-comment this line to test network3.
net.to(device)
#loss and optimiser defining
import torch.optim as optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(),
lr=LR) # changes in learing rate can be made here.
#validation function definition
def val(epoch):
net.eval()
val_loss = 0.0
val_total = 0
val_correct = 0
val_acc = 0
for i, (inputs, labels) in enumerate(validloader, 0):
# get the validation inputs
inputs = Variable(inputs)
labels = Variable(labels).long()
inputs, labels = inputs.to(device), labels.to(device)
class object_detector():
def __init__(self, start):
self.cap = cv2.VideoCapture(0)
self.start_time = start
self.stored_flag = False
self.trained_flag = False
self.milstone_flag = False
self.incremental_train_flag = False
self.tracking_flag = False
self.boxls = None
self.count = 1
self.new_count = 1
self.path = "/home/intuitivecompting/Desktop/color/Smart-Projector/script/datasets/"
if MODE == 'all':
self.file = open(self.path + "read.txt", "w")
self.milestone_file = open(self.path + "mileston_read.txt", "w")
self.user_input = 0
self.predict = None
self.memory = cache(10)
self.memory1 = cache(10)
self.hand_memory = cache(10)
self.node_sequence = []
#-----------------------create GUI-----------------------#
self.gui_img = np.zeros((130,640,3), np.uint8)
cv2.circle(self.gui_img,(160,50),30,(255,0,0),-1)
cv2.putText(self.gui_img,"start",(130,110),cv2.FONT_HERSHEY_SIMPLEX, 1.0,(255,0,0))
cv2.circle(self.gui_img,(320,50),30,(0,255,0),-1)
cv2.putText(self.gui_img,"stop",(290,110),cv2.FONT_HERSHEY_SIMPLEX, 1.0,(0,255,0))
cv2.circle(self.gui_img,(480,50),30,(0,0,255),-1)
cv2.putText(self.gui_img,"quit",(450,110),cv2.FONT_HERSHEY_SIMPLEX, 1.0,(0,0,255))
cv2.namedWindow('gui_img')
cv2.namedWindow('gui_img1')
cv2.setMouseCallback('gui_img',self.gui_callback)
cv2.setMouseCallback('gui_img1',self.gui_callback)
#-----------------------Training sign--------------#
self.training_surface = np.ones((610,640,3), np.uint8) * 255
cv2.putText(self.training_surface,'Training...',(120,300),cv2.FONT_HERSHEY_SIMPLEX, 3.0,(255,192,203), 5)
#----------------------new coming item id------------------#
self.new_come_id = None
self.old_come_id = None
self.new_come_side = None
self.old_come_side = None
self.new_coming_lock = True
self.once_lock = True
#---------------------set some flag-------------------#
self.storing = None
self.quit = None
self.once = True
#---------------------set gui image----------------------#
self.temp_surface = None
#----------------------for easlier developing-----------------#
if MODE == 'test':
if not GPU:
self.net = Net()
else:
self.net = Net().cuda()
self.net.load_state_dict(torch.load(f=self.path + 'model'))
self.user_input = 5
self.stored_flag = True
def update(self, save=True, train=False):
self.boxls = []
OK, origin = self.cap.read()
if OK:
rect = self.camrectify(origin)
#-------------warp the image---------------------#
warp = self.warp(rect)
#-------------segment the object----------------#
hsv = cv2.cvtColor(warp,cv2.COLOR_BGR2HSV)
green_mask = cv2.inRange(hsv, Green_low, Green_high)
# green_mask = cv2.inRange(hsv, np.array([45,90,29]), np.array([85,255,255]))
hand_mask = cv2.inRange(hsv, Hand_low, Hand_high)
hand_mask = cv2.dilate(hand_mask, kernel = np.ones((7,7),np.uint8))
skin_mask = cv2.inRange(hsv, Skin_low, Skin_high)
skin_mask = cv2.dilate(skin_mask, kernel = np.ones((7,7),np.uint8))
thresh = 255 - green_mask
thresh = cv2.subtract(thresh, hand_mask)
thresh = cv2.subtract(thresh, skin_mask)
thresh[477:, 50:610] = 0
#thresh = cv2.dilate(thresh, kernel = np.ones((11,11),np.uint8))
cv2.imshow('afg', thresh)
draw_img1 = warp.copy()
draw_img2 = warp.copy()
draw_img3 = warp.copy()
self.train_img = warp.copy()
#-------------get the bounding box--------------
self.get_bound(draw_img1, thresh, hand_mask, only=False, visualization=True)
#--------------get bags of words and training-------#
if MODE == 'all':
#----------------------------storing image for each item---------#
if not self.stored_flag:
self.temp_surface = np.vstack((draw_img1, self.gui_img))
self.stored_flag = self.store()
cv2.imshow('gui_img', self.temp_surface)
#--------------------------training, just once------------------#
if self.stored_flag and not self.trained_flag:
cv2.destroyWindow('gui_img')
#cv2.imshow('training', self.training_surface)
self.trained_flag = self.train()
#------------------------assembling and saving milstone---------#
if self.trained_flag and not self.milstone_flag:
self.test(draw_img2)
self.temp_surface = np.vstack((draw_img2, self.gui_img))
cv2.imshow('gui_img1', self.temp_surface)
#-----------------------training saved milstone image---------#
if self.milstone_flag and not self.incremental_train_flag:
cv2.destroyWindow('gui_img1')
self.incremental_train_flag = self.train(is_incremental=True)
#-----------------------finalized tracking------------------#
if self.incremental_train_flag and not self.tracking_flag:
self.test(draw_img3, is_tracking=True)
cv2.imshow('tracking', draw_img3)
elif MODE == 'test':
self.test(draw_img2)
self.temp_surface = np.vstack((draw_img2, self.gui_img))
cv2.imshow('gui_img', self.temp_surface)
#cv2.imshow('track', draw_img2)
#-----------------------training saved milstone image---------#
if self.milstone_flag and not self.incremental_train_flag:
cv2.destroyWindow('gui_img')
self.incremental_train_flag = self.train(is_incremental=True)
#-----------------------finalized tracking------------------#
if self.incremental_train_flag and not self.tracking_flag:
self.test(draw_img3, is_tracking=True)
cv2.imshow('gui_img1', draw_img3)
elif MODE == 'train':
if not self.trained_flag:
#cv2.destroyWindow('gui_img')
#cv2.imshow('training', self.training_surface)
self.trained_flag = self.train()
#------------------------assembling and saving milstone---------#
if self.trained_flag and not self.milstone_flag:
self.test(draw_img2)
self.temp_surface = np.vstack((draw_img2, self.gui_img))
cv2.imshow('gui_img1', self.temp_surface)
#-----------------------training saved milstone image---------#
if self.milstone_flag and not self.incremental_train_flag:
cv2.destroyWindow('gui_img1')
self.incremental_train_flag = self.train(is_incremental=True)
#-----------------------finalized tracking------------------#
if self.incremental_train_flag and not self.tracking_flag:
self.test(draw_img3, is_tracking=True)
cv2.imshow('tracking', draw_img3)
def get_bound(self, img, object_mask, hand_mask, only=True, visualization=True):
(_,object_contours, object_hierarchy)=cv2.findContours(object_mask,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
(_,hand_contours, hand_hierarchy)=cv2.findContours(hand_mask,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
hand_m_ls = []
object_m_ls = []
if len(hand_contours) > 0:
for i , contour in enumerate(hand_contours):
area = cv2.contourArea(contour)
if area>600 and area < 100000 and hand_hierarchy[0, i, 3] == -1:
M = cv2.moments(contour)
cx = int(M['m10']/M['m00'])
cy = int(M['m01']/M['m00'])
hand_m_ls.append((cx, cy))
if len(object_contours) > 0:
for i , contour in enumerate(object_contours):
area = cv2.contourArea(contour)
if area>100 and area < 100000 and object_hierarchy[0, i, 3] == -1:
M = cv2.moments(contour)
cx = int(M['m10']/M['m00'])
cy = int(M['m01']/M['m00'])
object_m_ls.append((cx, cy))
x,y,w,h = cv2.boundingRect(contour)
self.boxls.append([x, y, w, h])
temp_i = []
temp_j = []
for (x3, y3) in hand_m_ls:
for i in range(len(object_m_ls)):
for j in range(i + 1, len(object_m_ls)):
x1, y1 = object_m_ls[i]
x2, y2 = object_m_ls[j]
d12 = distant((x1, y1), (x2, y2))
d13 = distant((x1, y1), (x3, y3))
d23 = distant((x2, y2), (x3, y3))
# dis = d13 * d23 / d12
# if dis < 60 and d12 < 140 and d13 < 100 and d23 < 100:
# temp_i.append(i)
# temp_j.append(j)
dis = self.get_k_dis((x1, y1), (x2, y2), (x3, y3))
if dis < 60 and d12 < 140 and d13 < 100 and d23 < 100:
temp_i.append(i)
temp_j.append(j)
# print(dis, d12, d13, d23)
if len(temp_i) > 0 and len(temp_j) > 0 and len(self.boxls) >= 1:
for (i, j) in zip(temp_i, temp_j):
if self.boxls[i] != 0 and self.boxls[j] != 0:
x, y = np.min([self.boxls[i][0], self.boxls[j][0]]), np.min([self.boxls[i][1], self.boxls[j][1]])
x_max, y_max = np.max([self.boxls[i][0] + self.boxls[i][2], self.boxls[j][0] + self.boxls[j][2]]), np.max([self.boxls[i][1] + self.boxls[i][3], self.boxls[j][1] + self.boxls[j][3]])
w, h = x_max - x, y_max - y
self.boxls[i] = 0
self.boxls[j] = [x, y, w, h]
self.boxls = filter(lambda a: a != 0, self.boxls)
#---------------sorting the list according to the x coordinate of each item
if len(self.boxls) > 0:
boxls_arr = np.array(self.boxls)
self.boxls = boxls_arr[boxls_arr[:, 0].argsort()].tolist()
for i in range(len(self.boxls)):
if visualization:
ind = max(range(len(self.boxls)), key=lambda i:self.boxls[i][2]*self.boxls[i][3])
x,y,w,h = self.boxls[ind]
cv2.rectangle(img,(x,y),(x+w,y+h),(0,0,255),2)
cv2.putText(img,str(self.user_input),(x,y),cv2.FONT_HERSHEY_SIMPLEX, 1.0,(0,0,255))
def gui_callback(self, event, x, y, flags, param):
if event == cv2.EVENT_LBUTTONDBLCLK and (self.temp_surface[y, x] == np.array([255, 0, 0])).all() and not self.storing:
self.count = 1
self.user_input += 1
self.storing = True
if self.user_input > 5:
if self.once:
temp_node = node((self.new_come_id, self.old_come_id), (self.new_come_side, self.old_come_side),self.user_input)
self.once = False
else:
temp_node = node((self.new_come_id, self.user_input - 1), (self.new_come_side, self.old_come_side), self.user_input)
self.node_sequence.append(temp_node)
print("start")
if event == cv2.EVENT_LBUTTONDBLCLK and (self.temp_surface[y, x] == np.array([0, 255, 0])).all() and self.storing:
self.storing = False
self.new_coming_lock = True
self.new_come_id = None
self.old_come_id = None
self.new_come_side = None
self.old_come_side = None
print("stop")
if event == cv2.EVENT_LBUTTONDBLCLK and (self.temp_surface[y, x] == np.array([0, 0, 255])).all():
self.storing = None
self.quit = True
print("quit")
if self.stored_flag:
x,y,w,h = self.boxls[0]
sub_img = self.train_img[y:y+h, x:x+w, :]
cv2.imwrite('test_imgs/saved' + str(self.user_input) + '.jpg', sub_img)
# if event == cv2.EVENT_LBUTTONDBLCLK and (self.temp_surface[y, x] == np.array([255, 0, 255])).all():
# self.saving_milstone = True
# self.user_input += 1
def store(self, is_milestone=False):
# if is_milestone:
# file = self.milestone_file
# img_dir = os.path.join(self.path + "milestone_image", str(self.count) + ".jpg")
# self.createFolder(self.path + "milestone_image")
# else:
if is_milestone:
self.file = open(self.path + "read.txt", "a")
img_dir = os.path.join(self.path + "image", "milstone" + str(self.new_count) + ".jpg")
else:
img_dir = os.path.join(self.path + "image", str(self.new_count) + ".jpg")
file = self.file
self.createFolder(self.path + "image")
if self.quit:
file.close()
print('finish output')
return True
if len(self.boxls) > 0:
if self.storing:
cv2.putText(self.temp_surface,"recording",(450,50),cv2.FONT_HERSHEY_SIMPLEX, 1.0,(0,0,255), 2)
frame = self.train_img
ind = max(range(len(self.boxls)), key=lambda i:self.boxls[i][2]*self.boxls[i][3])
#-------------capturing img for each of item--------------#
x,y,w,h = self.boxls[ind]
temp = frame[y:y+h, x:x+w, :]
cv2.imwrite(img_dir, temp)
file.write(img_dir + " " + str(self.user_input) + "\n")
if self.count % 100 == 0:
print('output imgs ' + str(self.count) + 'img' )
self.count += 1
self.new_count += 1
return False
#-----------------get to the next item-----------
else:
return False
def train(self, is_incremental=False):
if is_incremental:
pickle.dump(node.pair_list ,open("node.p", "wb"))
start_time = time.time()
if not is_incremental:
reader_train = self.reader(self.path, "read.txt")
if not GPU:
self.net = Net()
else:
self.net = Net().cuda()
else:
if not GPU:
self.net = Net()
else:
self.net = Net().cuda()
reader_train = self.reader(self.path, "read.txt")
#self.net.load_state_dict(torch.load(f=self.path + 'model'))
optimizer = optim.SGD(self.net.parameters(), lr=LR, momentum=MOMENTUM, nesterov=True)
#optimizer = optim.Adam(self.net.parameters(), lr=LR, weight_decay=0.01)
schedule = optim.lr_scheduler.StepLR(optimizer, step_size=STEP, gamma=GAMMA)
trainset = CovnetDataset(reader=reader_train, transforms=transforms.Compose([transforms.Resize((200, 100)),
transforms.ToTensor()
]))
#trainset = CovnetDataset(reader=reader_train, transforms=transforms.Compose([transforms.Pad(30),
# transforms.ToTensor()
# ]))
trainloader = DataLoader(dataset=trainset, batch_size=BATCH_SIZE, shuffle=True, num_workers=2)
#-----------------------------------training----------------------------------------------------------------
if True:
loss_ls = []
count = 0
count_ls = []
t = tqdm.trange(EPOTH, desc='Training')
temp = 0
for _ in t: # loop over the dataset multiple times
schedule.step()
running_loss = 0.0
i = 0
for data in trainloader:
# get the inputs
inputs, labels = data
if GPU:
inputs, labels = inputs.cuda(), labels.cuda()
inputs, labels = Variable(inputs), Variable(labels.long())
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = self.net(inputs)
# print(outputs)
# print(labels.view(1, -1)[0])
loss = F.cross_entropy(outputs, labels.view(1, -1)[0])
loss.backward()
optimizer.step()
t.set_description('loss=%g' %(temp))
loss_ls.append(loss.item())
count += 1
count_ls.append(count)
running_loss += loss.item()
if i % 10 == 9:
temp = running_loss/10
running_loss = 0.0
i += 1
plt.plot(count_ls, loss_ls)
plt.show(block=False)
print('Finished Training, using {} second'.format(int(time.time() - start_time)))
self.quit = None
if not is_incremental:
self.user_input = 5
torch.save(self.net.state_dict(), f=self.path + 'model')
else:
torch.save(self.net.state_dict(), f=self.path + 'milestone_model')
# try:
# node_file = open(self.path + "node.txt", "w")
# for pair in node.pair_list:
# node_file.write(str(pair[0][0]) + "" + str(pair[0][1]) + "" +str(pair[1][0]) + "" + str(pair[1][1]) + "\n")
# except:
# print("fail to save")
return True
#---------------------------------testing-----------------------------------------------
def test(self, draw_img, is_tracking=False):
self.predict = []
net = self.net
num_object = len(self.boxls)
frame = self.train_img
preprocess = transforms.Compose([transforms.Resize((200, 100)),
transforms.ToTensor()])
#preprocess = transforms.Compose([transforms.Pad(30),
# transforms.ToTensor()])
for i in range(num_object):
x,y,w,h = self.boxls[i]
temp = frame[y:y+h, x:x+w, :]
temp = cv2.cvtColor(temp,cv2.COLOR_BGR2RGB)
image = Image.fromarray(temp)
img_tensor = preprocess(image)
img_tensor.unsqueeze_(0)
img_variable = Variable(img_tensor).cuda()
if GPU:
img_variable = Variable(img_tensor).cuda()
out = np.argmax(net(img_variable).cpu().data.numpy()[0])
else:
img_variable = Variable(img_tensor)
out = np.argmax(net(img_variable).data.numpy()[0])
# if np.max(net(img_variable).cpu().data.numpy()[0]) > 0.9:
# out = np.argmax(net(img_variable).cpu().data.numpy()[0])
# else:
# out = -1
cv2.rectangle(draw_img,(x,y),(x+w,y+h),(0,0,255),2)
cv2.putText(draw_img,str(out),(x,y),cv2.FONT_HERSHEY_SIMPLEX, 1.0,(0,0,255))
self.predict.append(((x, y, w, h), out))
if not is_tracking:
if self.old_come_side is not None and self.new_come_side is None:
cv2.putText(draw_img,"Point to next!",(220,50),cv2.FONT_HERSHEY_SIMPLEX, 0.7,(0,0,255), 2)
if self.new_come_side is not None and self.old_come_side is not None:
cv2.putText(draw_img,"Start Assemble! Click Start when finish",(180,50),cv2.FONT_HERSHEY_SIMPLEX, 0.7,(0,0,255), 2)
lab, color, ind, coord = self.store_side(frame)
if lab:
self.get_pair(frame.copy(), num_object, lab, color, ind, coord)
self.milstone_flag = self.store(is_milestone=True)
# self.memory.append(self.predict)
#print(len(self.memory.list))
def store_side(self, frame):
img = frame.copy()
point, center = hand_tracking(img, self.hand_memory).get_result()
if point and len(self.boxls) > 0:
red_center = side_finder(img, color='red')
blue_center = side_finder(img, color='blue')
tape = red_center + blue_center
length_ls = []
for (x, y) in tape:
length_ls.append((self.get_k_dis((point[0], point[1]), (center[0], center[1]), (x, y)), (x, y)))
x,y = min(length_ls, key=lambda x: x[0])[1]
cv2.circle(img, (x,y), 10, [255, 255, 0], -1)
ind = test_insdie((x, y), self.boxls)
# x,y,w,h = self.boxls[ind]
# line_canvas = np.zeros((h, w))
# cx, cy = center
# x1, y1 = point
# k = (y1-cy)/float(x1-cx)
# cv2.line(line_canvas, point, (x1-50, y1-50*k), (255,0,0), 5)
# frame_copy = frame.copy()
# sub_img = frame_copy[y:y+h, x:x+w, :]
# hsv = cv2.cvtColor(img,cv2.COLOR_BGR2HSV)
# object_mask = cv2.subtract(cv2.inRange(hsv, Green_low, Green_high),cv2.inRange(hsv, Hand_low, Hand_high))
# object_mask = 255 - object_mask
# (_,object_contours, object_hierarchy)=cv2.findContours(object_mask,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
# max_area = 0
# cnt = None
# for i , contour in enumerate(object_contours):
# area = cv2.contourArea(contour)
# if object_hierarchy[0, i, 3] == -1 and area > max_area:
# max_area = area
# cnt = contour
# cnt_canvas = np.zeros((h, w))
# cv2.imshow("point", img)
# print(ind, self.predict)
if ind is not None:
color = None
if (x, y) in red_center:
color = 'red'
elif (x, y) in blue_center:
color = 'blue'
return self.predict[ind][1], color, ind, (x, y)
else:
return None, None, None, None
else:
return None, None, None, None
#
def get_pair(self,image, num_object, label, color, index, coord):
'''
pointing from left to right
'''
if self.once and self.once_lock and num_object == 2:
if index == 0:
self.memory.append(self.predict[0][1])
if self.memory.full and self.new_come_id is None:
self.old_come_id = max(set(self.memory.list), key=self.memory.list.count)
# if self.new_come_id == label:
self.old_come_side = self.draw_point(image, coord, index)
# cv2.circle(image, coord, 5, (125, 125), 1)
# x,y,w,h = self.boxls[index]
# sub_img = image[y:y+h, x:x+w, :]
# cv2.imwrite('saved' + str(self.predict[0][1]) + '.jpg', sub_img)
# else:
# self.memory.clear()
if self.memory.full and index == 1:
self.memory1.append(self.predict[-1][1])
if self.memory1.full:
self.new_come_id = max(set(self.memory1.list), key=self.memory1.list.count)
# if self.old_come_id == label:
self.new_come_side = self.draw_point(image, coord, index)
# cv2.circle(image, coord, 5, (125, 125), 1)
# x,y,w,h = self.boxls[index]
# sub_img = image[y:y+h, x:x+w, :]
# cv2.imwrite('saved' + str(self.predict[-1][1]) + '.jpg', sub_img)
# else:
# self.memory1.clear()
if self.memory.full and self.memory1.full:
self.once_lock = False
self.memory.clear()
self.memory1.clear()
print("new_come_id:{}, old_come_id:{}".format(self.new_come_id, self.old_come_id))
print("new_come_side:{}, old_come_side:{}".format(self.new_come_side, self.old_come_side))
print(self.new_come_side, self.old_come_side)
'''
pointing from left to right
'''
if not self.once and num_object == 2 and self.new_coming_lock:
if index == 0:
self.memory.append(0)
if self.memory.full:
self.old_come_side = self.draw_point(image, coord, index, is_milestone=True)
self.memory.clear()
# cv2.circle(image, coord, 5, (125, 125), 1)
# x,y,w,h = self.boxls[index]
# sub_img = image[y:y+h, x:x+w, :]
# cv2.imwrite('saved' + str(self.user_input + 1) + '.jpg', sub_img)
elif index == 1 and self.new_come_id is None:
self.memory1.append(self.predict[-1][1])
if self.memory1.full:
self.new_come_id = max(set(self.memory1.list), key=self.memory1.list.count)
self.new_come_side = self.draw_point(image, coord, index)
self.memory1.clear()
# cv2.circle(image, coord, 5, (125, 125), 1)
# x,y,w,h = self.boxls[index]
# sub_img = image[y:y+h, x:x+w, :]
# cv2.imwrite('saved' + str(self.predict[1][1]) + '.jpg', sub_img)
if self.new_come_side and self.old_come_side:
self.new_coming_lock = False
print("new_come_id:{}".format(self.new_come_id))
print("new_come_side:{}, old_come_side:{}".format(self.new_come_side, self.old_come_side))
def draw_point(self, image, coord, index, is_milestone=False):
#cv2.circle(image, coord, 5, (125, 125), 1)
x,y,w,h = self.boxls[index]
sub_img = image[y:y+h, x:x+w, :]
#cv2.circle(sub_img, (coord[0] - x, coord[1] - y) , 5, (125, 125), -1)
if not is_milestone:
cv2.imwrite('test_imgs/saved' + str(self.predict[index][1]) + '.jpg', sub_img)
return (coord[0] - x, coord[1] - y)
else:
cv2.imwrite('test_imgs/saved' + str(self.user_input) + '.jpg', sub_img)
return (coord[0] - x, coord[1] - y)
def warp(self, img):
#pts1 = np.float32([[115,124],[520,112],[2,476],[640,480]])
pts1 = np.float32([[101,160],[531,133],[0,480],[640,480]])
pts2 = np.float32([[0,0],[640,0],[0,480],[640,480]])
M = cv2.getPerspectiveTransform(pts1,pts2)
dst = cv2.warpPerspective(img,M,(640,480))
return dst
@staticmethod
def get_k_dis((x1, y1), (x2, y2), (x, y)):
coord = ((x, y), (x1, y1), (x2, y2))
return Polygon(coord).area / distant((x1, y1), (x2, y2))
def main():
# Training settings
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument(
'--batch-size',
type=int,
default=64,
metavar='N',
help='input batch size for training (default: %(default)s)')
parser.add_argument(
'--test-batch-size',
type=int,
default=100,
metavar='N',
help='input batch size for testing (default: %(default)s)')
parser.add_argument(
'--epochs',
type=int,
default=30,
metavar='N',
help='number of epochs to train (default: %(default)s)')
parser.add_argument('--lr',
type=float,
default=1e-3,
metavar='LR',
help='learning rate (default: %(default)s)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: %(default)s)')
parser.add_argument(
'--log-interval',
type=int,
default=10,
metavar='N',
help=
'how many batches to wait before logging training status (default: %(default)s)'
)
parser.add_argument('--dataset',
choices=['mnist', 'fashion-mnist'],
default='mnist',
metavar='D',
help='mnist/fashion-mnist (default: %(default)s)')
parser.add_argument('--nonlin',
choices=['softplus', 'sigmoid', 'tanh'],
default='softplus',
metavar='D',
help='softplus/sigmoid/tanh (default: %(default)s)')
parser.add_argument('--num-layers',
choices=['2', '3', '4'],
default=2,
metavar='N',
help='2/3/4 (default: %(default)s)')
parser.add_argument('--epsilon',
type=float,
default=1.58,
metavar='E',
help='ball radius (default: %(default)s)')
parser.add_argument('--test-epsilon',
type=float,
default=1.58,
metavar='E',
help='ball radius (default: %(default)s)')
parser.add_argument(
'--step-size',
type=float,
default=0.005,
metavar='L',
help='step size for finding adversarial example (default: %(default)s)'
)
parser.add_argument(
'--num-steps',
type=int,
default=200,
metavar='L',
help=
'number of steps for finding adversarial example (default: %(default)s)'
)
parser.add_argument(
'--beta',
type=float,
default=0.005,
metavar='L',
help='regularization coefficient for Lipschitz (default: %(default)s)')
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
if args.dataset == 'mnist':
dataset = datasets.MNIST
elif args.dataset == 'fashion-mnist':
dataset = datasets.FashionMNIST
else:
raise ValueError('Unknown dataset %s', args.dataset)
torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(dataset(
'./' + args.dataset,
train=True,
download=True,
transform=transforms.Compose([transforms.ToTensor()])),
batch_size=args.batch_size,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(dataset(
'./' + args.dataset,
train=False,
transform=transforms.Compose([transforms.ToTensor()])),
batch_size=args.test_batch_size,
shuffle=False,
**kwargs)
model = Net(int(args.num_layers), args.nonlin).to(device)
optimizer = optim.Adam(model.parameters(), lr=args.lr)
model_name = 'saved_models/' + args.dataset + '_' + str(
args.num_layers) + '_' + args.nonlin + '_L2_' + str(
args.epsilon) + '_EIGEN_' + str(args.beta)
print(args)
print(model_name)
acc, empirical_acc = test_standard_adv(args, model, device, test_loader)
certified_acc = test_cert(args, model, device, test_loader)
best_acc = 0.
best_empirical_acc = 0.
best_certified_acc = 0.
for epoch in range(1, args.epochs + 1):
train_robust(args, model, device, train_loader, optimizer, epoch)
acc, empirical_acc = test_standard_adv(args, model, device,
test_loader)
certified_acc = test_cert(args, model, device, test_loader)
if acc > best_acc:
best_acc = acc
best_empirical_acc = empirical_acc
best_certified_acc = certified_acc
torch.save(model.state_dict(), model_name)
print('Saved model: Accuracy: {:.4f}, Empirical Robust Accuracy: {:.4f}, Certified Robust Accuracy: {:.4f}\n'.\
format(best_acc, best_empirical_acc, best_certified_acc))
class DQN:
def __init__(self,
memory_size=50000,
batch_size=128,
gamma=0.99,
lr=1e-3,
n_step=500000):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.gamma = gamma
# memory
self.memory_size = memory_size
self.Memory = ReplayMemory(self.memory_size)
self.batch_size = batch_size
# network
self.target_net = Net().to(self.device)
self.eval_net = Net().to(self.device)
self.target_update() # initialize same weight
self.target_net.eval()
# optim
self.optimizer = optim.Adam(self.eval_net.parameters(), lr=lr)
def select_action(self, state, eps):
prob = random.random()
if prob > eps:
return self.eval_net.act(state), False
else:
return (torch.tensor(
[[random.randrange(0, 9)]],
device=self.device,
dtype=torch.long,
), True)
def select_dummy_action(self, state):
state = state.reshape(3, 3, 3)
open_spots = state[:, :, 0].reshape(-1)
p = open_spots / open_spots.sum()
return np.random.choice(np.arange(9), p=p)
def target_update(self):
self.target_net.load_state_dict(self.eval_net.state_dict())
def learn(self):
if self.Memory.__len__() < self.batch_size:
return
# random batch sampling
transitions = self.Memory.sampling(self.batch_size)
batch = Transition(*zip(*transitions))
non_final_mask = torch.tensor(
tuple(map(lambda s: s is not None, batch.next_state)),
device=self.device,
dtype=torch.bool,
)
non_final_next_states = torch.cat(
[s for s in batch.next_state if s is not None]).to(self.device)
state_batch = torch.cat(batch.state).to(self.device)
action_batch = torch.cat(batch.action).to(self.device)
reward_batch = torch.cat(batch.reward).to(self.device)
# Q(s)
Q_s = self.eval_net(state_batch).gather(1, action_batch)
# maxQ(s') no grad for target_net
Q_s_ = torch.zeros(self.batch_size, device=self.device)
Q_s_[non_final_mask] = self.target_net(non_final_next_states).max(
1)[0].detach()
# Q_target=R+γ*maxQ(s')
Q_target = reward_batch + (Q_s_ * self.gamma)
# loss_fnc---(R+γ*maxQ(s'))-Q(s)
# huber loss with delta=1
loss = F.smooth_l1_loss(Q_s, Q_target.unsqueeze(1))
# Optimize the model
self.optimizer.zero_grad()
loss.backward()
for param in self.eval_net.parameters():
param.grad.data.clamp_(-1, 1)
self.optimizer.step()
def load_net(self, name):
self.action_net = torch.load(name).cpu()
def load_weight(self, name):
self.eval_net.load_state_dict(torch.load(name))
self.eval_net = self.eval_net.cpu()
def act(self, state):
with torch.no_grad():
p = F.softmax(self.action_net.forward(state)).cpu().numpy()
valid_moves = (state.cpu().numpy().reshape(
3, 3, 3).argmax(axis=2).reshape(-1) == 0)
p = valid_moves * p
return p.argmax()