Related
I am trying to train an adversarial patch located at the bottom left corner of the image to cause a misclassification. Currently, I am using these parameters to normalize the CIFAR10 dataset.
transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914,0.4822,0.4465),(0.2023,0.1994,0.201))]
This would result in the images having a maximum and minimum value of around 2.55 and -2.55 respectively. However, I'm not sure how to work with this range when training my patch. I struggle between converting the patch from a range of (0,1) to (-2.55,2.55). Any help is appreciated!
My code for training is below: (I don't think its training properly for now)
import torch
import torchvision
import torchvision.transforms as transforms
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models
from torch.utils.data import DataLoader
from torch.utils.data.sampler import SubsetRandomSampler
import matplotlib.pyplot as plt
import numpy as np
from torch.autograd import Variable
import torchattacks
import random
import torch.nn.functional as F
dictionary ={
'0':'airplane',
'1':'automobile',
'2':'bird',
'3':'cat',
'4':'deer',
'5':'dog',
'6':'frog',
'7':'horse',
'8':'ship',
'9':'truck',
}
transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914,0.4822,0.4465),(0.2023,0.1994,0.201))])
#transform1 = transforms.Compose([transforms.ToTensor()])
normalize = transforms.Normalize((0.4914,0.4822,0.4465),(0.2023,0.1994,0.201))
mean =(0.4914,0.4822,0.4465)
std =(0.2023,0.1994,0.201)
inv_normalize = transforms.Normalize(
mean=[-0.4914/0.2023, -0.4822/0.1994, -0.4465/0.201],
std=[1/0.2023, 1/0.1994, 1/0.201])
batch_size = 1
trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
shuffle=False, num_workers=2)
testset = torchvision.datasets.CIFAR10(root='./data', train=False,
download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size,
shuffle=True, num_workers=2)
model = torch.hub.load("chenyaofo/pytorch-cifar-models", "cifar10_resnet20", pretrained=True)
model = model.cuda()
import os
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
patch = np.random.rand(3,32,32)
model.eval()
def mask_generation(mask_type='rectangle', patch = patch, image_size=(3, 7, 7)):
applied_patch = np.zeros(image_size) #0,1
#patch = torch.tensor(patch)
#padding = (3,3,3,3)
#patch = F.pad(patch, padding)
if mask_type == 'rectangle':
rotation_angle = 0
for i in range(patch.shape[0]):
patch[i] = np.rot90(patch[i], rotation_angle)
x_location , y_location = 25,0
for i in range(patch.shape[0]):
applied_patch[:, x_location:x_location + patch.shape[1], y_location:y_location + patch.shape[2]] = patch
mask = applied_patch.copy()
mask[mask != 0] = 1.0
return patch , applied_patch, mask, x_location, y_location , rotation_angle
def patch_attack(image, applied_patch, mask, target, probability_threshold, model, lr, max_iteration):
applied_patch = torch.from_numpy(applied_patch)
mask = torch.from_numpy(mask)
image = inv_normalize(image)
target_probability, count = 0,0
perturbated_image = torch.mul(mask.type(torch.FloatTensor), applied_patch.type(torch.FloatTensor)) + torch.mul((1 - mask.type(torch.FloatTensor)), image.type(torch.FloatTensor))
perturbated_image = normalize(perturbated_image)
while target_probability < probability_threshold and count < max_iteration:
count += 1
# Optimize the patch
perturbated_image = Variable(perturbated_image.data, requires_grad=True)
per_image = perturbated_image.cuda()
output = model(per_image)
target_log_softmax = torch.nn.functional.log_softmax(output, dim=1)[0][target]
target_log_softmax.backward()
patch_grad = perturbated_image.grad.clone().cpu()
applied_patch = (lr * patch_grad) + applied_patch.type(torch.FloatTensor)
applied_patch = torch.clamp(applied_patch,0,1)
perturbated_image.grad.data.zero_()
# Test the patch
perturbated_image = torch.mul(mask.type(torch.FloatTensor), applied_patch.type(torch.FloatTensor)) + torch.mul((1-mask.type(torch.FloatTensor)), image.type(torch.FloatTensor))
perturbated_image = normalize(perturbated_image)
perturbated_image = perturbated_image.cuda()
output = model(perturbated_image)
target_probability = torch.nn.functional.softmax(output, dim=1).data[0][target]
perturbated_image = perturbated_image.detach().cpu().numpy()
applied_patch = applied_patch.cpu().numpy()
return perturbated_image, applied_patch
def test_patch(patch_type, target, patch, test_loader, model):
test_total, test_actual_total, test_success = 0, 0, 0
for (image, label) in test_loader:
test_total += label.shape[0]
assert image.shape[0] == 1, 'Only one picture should be loaded each time.'
image = image.cuda() #-3,3
label = label.cuda()
output = model(image)
_, predicted = torch.max(output.data, 1)
if predicted[0] != label and predicted[0].data.cpu().numpy() != target:
test_actual_total += 1
patch ,applied_patch, mask, x_location, y_location = mask_generation('rectangle', patch, (3, 32, 32))
applied_patch = torch.from_numpy(applied_patch)
mask = torch.from_numpy(mask)
mask = normalize(mask)
applied_patch = normalize(applied_patch)
perturbated_image = torch.mul(mask.type(torch.FloatTensor), applied_patch.type(torch.FloatTensor)) + torch.mul((1 - mask.type(torch.FloatTensor)), image.type(torch.FloatTensor))
perturbated_image = perturbated_image.cuda() #-3,3
output = model(perturbated_image)
_, predicted = torch.max(output.data, 1)
if predicted[0].data.cpu().numpy() == target:
test_success += 1
return test_success / test_actual_total
#training parameters
epochs = 1
target = 0
probability_threshold = 0.99
lr = 1/255
max_iteration = 1
runs = 0
for epoch in range(epochs):
train_total, train_actual_total, train_success = 0, 0, 0
for (image, label) in trainloader:
runs+=1
assert image.shape[0] == 1
image = image.cuda()
label = label.cuda()
train_total += label.shape[0]
output = model(image)
_, predicted = torch.max(output.data, 1)
if predicted[0] != label or predicted[0].data.cpu().numpy() != target:
train_actual_total += 1
patch , applied_patch, mask, x_location, y_location ,rotation_angle = mask_generation('rectangle', patch, (3, 32, 32))
perturbated_image, applied_patch = patch_attack(image, applied_patch, mask, target, probability_threshold, model, lr,max_iteration)
perturbated_image = torch.from_numpy(perturbated_image).cuda()
output = model(perturbated_image)
_, predicted = torch.max(output.data, 1)
if predicted[0].data.cpu().numpy() == target:
train_success += 1
patch = applied_patch[0][:, x_location:x_location + patch.shape[1], y_location:y_location + patch.shape[2]]
patch = np.array(patch)
To convert a number x in the range [0,1] to the range [-2.55,2.55]:
Multiply by size of final range / size of original range or in this case 5.1/1.0.
Add min of final range - min of starting range to the result, so in this case -2.55+0 = 0.
I recently tried implementing perceptron algorithm but I was not getting the desired output.
Here is the code:
import numpy as np
import pandas as pd
with open("D:/data.txt",'r') as data: #importing the data
column = data.read()
split = np.array(column.split('\n'))
final =[]
for string in split:
final.append(string.split(','))
df = pd.DataFrame(final,columns=['x','y','response'])
df['x'] = df['x'].astype(float)
df['y'] = df['y'].astype(float)
df['response'] = df['response'].astype(int)
X = np.array(df[['x','y']])
y = np.array(df['response'])
def perceptron_algorithm(x,y,learning_rate=0.01,num_epoch=25):
np.random.seed(2)
x_min, x_max = min(x.T[0]), max(x.T[0])
y_min, y_max = min(x.T[1]), max(x.T[0])
w = np.array(np.random.rand(2,1))
b = np.random.rand(1)[0] + x_max
print(w,b)
for i in range(num_epoch):
w,b = perceptronstep(x,y,w,b,learning_rate)
print(w,b)
return w,b
def perceptronstep(x,y,w,b,learning_rate):
for i in range(len(x)):
y_hat = prediction(x[i],w,b)
if y_hat-y[i] == 1:
for j in range(len(w)):
w[j] += x[i][j]*learning_rate
b += learning_rate
elif y_hat-y[i] == -1:
for j in range(len(w)):
w[j] -= x[i][j]*learning_rate
b -= learning_rate
return w,b
def prediction(x,w,b):
return step(np.matmul(x,w)+b)
def step(t):
if t >=0:
return 1
else:
return 0
w,b = perceptron_algorithm(X,y)
This is the resulting line:
This is how the data looks:
Is there something wrong with my code ?
Here is the link to the data file:
https://drive.google.com/drive/folders/1TSug9tE6bljyBFv-u3mIGWW6F_3ZY2oa?usp=sharing
Edit: I have added the initial part of the code so it will be clear what I am trying to do.
Edit 2: I have added the data file and the "import pandas as pd" line of code
I'm trying to plot different paths at rviz
I'm using the following code to get a first approach (based on this repository: https://github.com/HaoQChen/show_trajectory/tree/master/src)
import rospy
import math
import numpy as np
from geometry_msgs.msg import PoseStamped
from nav_msgs.msg import Path, Odometry
from std_msgs.msg import Empty
class ProjectElement(object):
def __init__(self):
self.path_pub = rospy.Publisher('~path', Path, latch=True, queue_size=10)
self.circle_sub = rospy.Subscriber('~circle', Empty, self.circle_cb, queue_size=10)
self.line_sub = rospy.Subscriber('~line', Empty, self.line_cb, queue_size=10)
self.project_sub = rospy.Subscriber('~project', Empty, self.project_cb, queue_size=10)
self.paths = []
self.rate = rospy.Rate(50)
def circle_cb(self, msg):
path = Path()
centre_x = 1
centre_y = 1
R = 0.5
th = 0.0
delta_th = 0.1
while (th<2*math.pi):
x = centre_x + R * math.sin(th)
y = centre_y + R * math.cos(th)
th += delta_th
this_pose_stamped = PoseStamped()
this_pose_stamped.pose.position.x = x
this_pose_stamped.pose.position.y = y
this_pose_stamped.header.stamp = rospy.get_rostime()
this_pose_stamped.header.frame_id = "/my_cs"
path.poses.append(this_pose_stamped)
path.header.frame_id = "/my_cs"
path.header.stamp = rospy.get_rostime()
self.paths.append(path)
def line_cb(self, msg):
path = Path()
x_start = 0.0
y_start = 0.0
length = 2
angle = 45 * math.pi/180
th = 0.0
delta_th = 0.1
while (th<length):
x = x_start + th * math.cos(angle)
y = y_start + th * math.sin(angle)
th += delta_th
this_pose_stamped = PoseStamped()
this_pose_stamped.pose.position.x = x
this_pose_stamped.pose.position.y = y
this_pose_stamped.header.stamp = rospy.get_rostime()
this_pose_stamped.header.frame_id = "/my_cs"
path.poses.append(this_pose_stamped)
path.header.frame_id = "/my_cs"
path.header.stamp = rospy.get_rostime()
self.paths.append(path)
def project_cb(self, msg):
while(True):
for element in self.paths:
# element.header.stamp = rospy.get_rostime()
self.path_pub.publish(element)
if __name__ == '__main__':
rospy.init_node('path_simulate')
elements = ProjectElement()
rospy.spin()
I can visualize the paths at rviz, but I don't know how to plot both figures at the same time in this way.
line
circle
I would like to ask if this approach is the best way to address this issue or which could be the best way.
I finally solved with visualization_msgs/MarkerArray based on this question:
https://answers.ros.org/question/220898/how-to-use-rviz-to-show-multiple-planing-paths/?answer=220993#post-id-220993
I post the code used here in case anyone needs it
import rospy
import math
import numpy as np
from geometry_msgs.msg import Vector3, Point
from std_msgs.msg import Empty
from visualization_msgs.msg import Marker, MarkerArray
class ProjectElement(object):
def __init__(self):
self.marker_pub = rospy.Publisher('~marker', MarkerArray, latch=True, queue_size=10)
self.circle_sub = rospy.Subscriber('~circle', Empty, self.circle_cb, queue_size=10)
self.line_sub = rospy.Subscriber('~line', Empty, self.line_cb, queue_size=10)
self.project_sub = rospy.Subscriber('~project', Empty, self.project_cb, queue_size=10)
self.marker_array = MarkerArray()
def circle_cb(self, msg):
marker = Marker()
marker.type = Marker.LINE_STRIP
marker.action = Marker.ADD
marker.scale = Vector3(0.01, 0.01, 0)
marker.color.g = 1.0
marker.color.a = 1.0
centre_x = 1
centre_y = 1
R = 0.5
delta_th = 0.1
for th in np.arange(0.0, 2*math.pi+delta_th, delta_th):
x = centre_x + R * math.sin(th)
y = centre_y + R * math.cos(th)
point = Point()
point.x = x
point.y = y
marker.points.append(point)
marker.id = 0
marker.header.stamp = rospy.get_rostime()
marker.header.frame_id = "/my_cs"
self.marker_array.markers.append(marker)
def line_cb(self, msg):
marker = Marker()
marker.type = Marker.LINE_STRIP
marker.action = Marker.ADD
marker.scale = Vector3(0.01, 0.01, 0)
marker.color.g = 1.0
marker.color.a = 1.0
x_start = 0.0
y_start = 0.0
length = 2
angle = 45 * math.pi/180
delta_th = 0.1
for th in np.arange(0.0, length, delta_th):
x = x_start + th * math.cos(angle)
y = y_start + th * math.sin(angle)
point = Point()
point.x = x
point.y = y
marker.points.append(point)
marker.id = 1
marker.header.stamp = rospy.get_rostime()
marker.header.frame_id = "/my_cs"
self.marker_array.markers.append(marker)
def project_cb(self, msg):
self.marker_pub.publish(self.marker_array)
if __name__ == '__main__':
rospy.init_node('markers_simulate')
elements = ProjectElement()
rospy.spin()
And the result achieved
marker array display
I'm learning the Deep Reinforcement learning
framework Chainer.
I've followed a tutorial and gotten the following code:
def train_dddqn(env):
class Q_Network(chainer.Chain):
def __init__(self, input_size, hidden_size, output_size):
super(Q_Network, self).__init__(
fc1=L.Linear(input_size, hidden_size),
fc2=L.Linear(hidden_size, hidden_size),
fc3=L.Linear(hidden_size, hidden_size // 2),
fc4=L.Linear(hidden_size, hidden_size // 2),
state_value=L.Linear(hidden_size // 2, 1),
advantage_value=L.Linear(hidden_size // 2, output_size)
)
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
def __call__(self, x):
h = F.relu(self.fc1(x))
h = F.relu(self.fc2(h))
hs = F.relu(self.fc3(h))
ha = F.relu(self.fc4(h))
state_value = self.state_value(hs)
advantage_value = self.advantage_value(ha)
advantage_mean = (F.sum(advantage_value, axis=1) / float(self.output_size)).reshape(-1, 1)
q_value = F.concat([state_value for _ in range(self.output_size)], axis=1) + (
advantage_value - F.concat([advantage_mean for _ in range(self.output_size)], axis=1))
return q_value
def reset(self):
self.cleargrads()
Q = Q_Network(input_size=env.history_t + 1, hidden_size=100, output_size=3)
Q_ast = copy.deepcopy(Q)
optimizer = chainer.optimizers.Adam()
optimizer.setup(Q)
epoch_num = 50
step_max = len(env.data) - 1
memory_size = 200
batch_size = 50
epsilon = 1.0
epsilon_decrease = 1e-3
epsilon_min = 0.1
start_reduce_epsilon = 200
train_freq = 10
update_q_freq = 20
gamma = 0.97
show_log_freq = 5
memory = []
total_step = 0
total_rewards = []
total_losses = []
start = time.time()
for epoch in range(epoch_num):
pobs = env.reset()
step = 0
done = False
total_reward = 0
total_loss = 0
while not done and step < step_max:
# select act
pact = np.random.randint(3)
if np.random.rand() > epsilon:
pact = Q(np.array(pobs, dtype=np.float32).reshape(1, -1))
pact = np.argmax(pact.data)
# act
obs, reward, done = env.step(pact)
# add memory
memory.append((pobs, pact, reward, obs, done))
if len(memory) > memory_size:
memory.pop(0)
# train or update q
if len(memory) == memory_size:
if total_step % train_freq == 0:
shuffled_memory = np.random.permutation(memory)
memory_idx = range(len(shuffled_memory))
for i in memory_idx[::batch_size]:
batch = np.array(shuffled_memory[i:i + batch_size])
b_pobs = np.array(batch[:, 0].tolist(), dtype=np.float32).reshape(batch_size, -1)
b_pact = np.array(batch[:, 1].tolist(), dtype=np.int32)
b_reward = np.array(batch[:, 2].tolist(), dtype=np.int32)
b_obs = np.array(batch[:, 3].tolist(), dtype=np.float32).reshape(batch_size, -1)
b_done = np.array(batch[:, 4].tolist(), dtype=np.bool)
q = Q(b_pobs)
indices = np.argmax(q.data, axis=1)
maxqs = Q_ast(b_obs).data
target = copy.deepcopy(q.data)
for j in range(batch_size):
Q.reset()
loss = F.mean_squared_error(q, target)
total_loss += loss.data
loss.backward()
optimizer.update()
if total_step % update_q_freq == 0:
Q_ast = copy.deepcopy(Q)
# epsilon
if epsilon > epsilon_min and total_step > start_reduce_epsilon:
epsilon -= epsilon_decrease
# next step
total_reward += reward
pobs = obs
step += 1
total_step += 1
total_rewards.append(total_reward)
total_losses.append(total_loss)
if (epoch + 1) % show_log_freq == 0:
log_reward = sum(total_rewards[((epoch + 1) - show_log_freq):]) / show_log_freq
log_loss = sum(total_losses[((epoch + 1) - show_log_freq):]) / show_log_freq
elapsed_time = time.time() - start
print('\t'.join(map(str, [epoch + 1, epsilon, total_step, log_reward, log_loss, elapsed_time])))
start = time.time()
return Q, total_losses, total_rewards
Q, total_losses, total_rewards = train_dddqn(Environment1(train))
My question is how can I save and load this Model which has been train very well?I know Kreas has some function like: model.save and load_model.
So what's the specify code I need for this Chainer code?
You can use serializer module to save/load chainer's model's parameter (Chain class).
from chainer import serializers
Q = Q_Network(input_size=env.history_t + 1, hidden_size=100, output_size=3)
Q_ast = Q_Network(input_size=env.history_t + 1, hidden_size=100, output_size=3)
# --- train Q here... ---
# copy Q parameter into Q_ast by saving Q's parameter and load to Q_ast
serializers.save_npz('my.model', Q)
serializers.load_npz('my.model', Q_ast)
See official document for details:
http://docs.chainer.org/en/stable/guides/serializers.html
Also, you may refer chainerrl, which is a chainer library for reinforcement learning.
https://github.com/chainer/chainerrl
chainerrl have a util function copy_param to copy parameter from network source_link to target_link.
https://github.com/chainer/chainerrl/blob/master/chainerrl/misc/copy_param.py#L12-L30
I am using a machine which has 2 GPUs Titan Black to train my deep learning model which has 3 layers (3x3, 3x3 and 5x5).
The training runs pretty well but when I watch nvidia-smi (watch every 1 sec), I realized that my program uses only one GPU for computation, the second one always 0% even when the first one reach 100%.
I am trying to use tf.device to assign specific tasks for each of them but then they run one-by-one, not in parallel, and the total time was even increased, not reduced (I guess because 2 GPUs had to exchange values with each other)
Below is my program. It is quite messy, maybe you just need to pay attention at the graph where I use tf.device is enough...
Thank you so much!
import tensorflow as tf
import numpy as np
from six.moves import cPickle as pickle
import matplotlib.pyplot as plt
from os import listdir, sys
from os.path import isfile, join
from time import gmtime, strftime
import time
def validatePath(path):
path = path.replace("\\","/")
if (path[len(path)-1] != "/"):
path = path + "/"
return path
hidden_size_default = np.array([16, 32, 64, 32])
cnn1_default = 3
cnn2_default = 3
cnn3_default = 5
SIZE_BATCH_VALID = 200
input_path = 'ARCHIVES-sub-dataset'
output_path = 'ARCHIVES-model'
log_address = "trainlog.txt"
tf.app.flags.DEFINE_integer('h0', hidden_size_default[0], 'Size of hidden layer 0th')
tf.app.flags.DEFINE_integer('h1', hidden_size_default[1], 'Size of hidden layer 1st')
tf.app.flags.DEFINE_integer('h2', hidden_size_default[2], 'Size of hidden layer 2nd')
tf.app.flags.DEFINE_integer('h3', hidden_size_default[3], 'Size of hidden layer 3rd')
tf.app.flags.DEFINE_integer('k1', cnn1_default , 'Size of kernel 1st')
tf.app.flags.DEFINE_integer('k2', cnn2_default , 'Size of kernel 2nd')
tf.app.flags.DEFINE_integer('k3', cnn3_default , 'Size of kernel 3rd')
tf.app.flags.DEFINE_string('input_path', input_path, 'The parent directory which contains 2 directories: dataset and label')
tf.app.flags.DEFINE_string('output_path', output_path, 'The directory which will store models (you have to create)')
tf.app.flags.DEFINE_string('log_address', log_address, 'The file name which will store the log')
FLAGS = tf.app.flags.FLAGS
load_path = FLAGS.input_path
save_model_path = FLAGS.output_path
log_addr = FLAGS.log_address
load_path = validatePath(load_path)
save_model_path = validatePath(save_model_path)
cnn1 = FLAGS.k1
cnn2 = FLAGS.k2
cnn3 = FLAGS.k3
hidden_size = np.array([FLAGS.h0, FLAGS.h1, FLAGS.h2, FLAGS.h3])
# Shuffle the dataset and its label
def randomize(dataset, labels):
permutation = np.random.permutation(labels.shape[0])
shuffled_dataset = dataset[permutation,:]
shuffled_labels = labels[permutation]
return shuffled_dataset, shuffled_labels
def writemyfile(mystring):
with open(log_addr, "a") as myfile:
myfile.write(str(mystring + "\n"))
num_labels = 5
def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))/ predictions.shape[0])
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
def DivideSets(input_set):
length_set = input_set.shape[0]
index_70 = int(length_set*0.7)
index_90 = int(length_set*0.9)
set_train = input_set[0:index_70]
set_valid = input_set[index_70:index_90]
set_test = input_set[index_90:length_set]
return np.float32(set_train), np.float32(set_valid), np.float32(set_test)
# from 1-value labels to 5 values of (0 and 1)
def LabelReconstruct(label_set):
label_set = label_set.astype(int)
new_label_set = np.zeros(shape=(len(label_set),num_labels))
for i in range(len(label_set)):
new_label_set[i][label_set[i]] = 1
return new_label_set.astype(int)
def LoadDataSet(load_path):
list_data = [f for f in listdir(load_path + "dataset/") if isfile(join(load_path + "dataset/", f))]
list_label = [f for f in listdir(load_path + "label/") if isfile(join(load_path + "dataset/", f))]
if list_data.sort() == list_label.sort():
return list_data
else:
print("data and labels are not suitable")
return 0
# load, randomize, normalize images and reconstruct labels
def PrepareData(*arg):
filename = arg[0]
loaded_dataset = pickle.load( open( load_path + "dataset/" + filename, "rb" ))
loaded_labels = pickle.load( open( load_path + "label/" + filename, "rb" ))
if len(arg) == 1:
datasize = len(loaded_labels)
elif len(arg) == 2:
datasize = int(arg[1])
else:
print("not more than 2 arguments please!")
dataset_full,labels_full = randomize(loaded_dataset[0:datasize], loaded_labels[0:datasize])
return NormalizeData(dataset_full), LabelReconstruct(labels_full)
def NormalizeData(dataset):
dataset = dataset - (dataset.mean())
dataset = dataset / (dataset.std())
return dataset
### LOAD DATA
listfiles = LoadDataSet(load_path)
# divide
listfiles_train = listfiles[0:15]
listfiles_valid = listfiles[15:25]
listfiles_test = listfiles[25:len(listfiles)]
graphCNN = tf.Graph()
with graphCNN.as_default():
with tf.device('/gpu:0'):
x = tf.placeholder(tf.float32, shape=(None, 224,224,3)) # X
y_ = tf.placeholder(tf.float32, shape=(None, num_labels)) # Y_
dropout = tf.placeholder(tf.float32)
if dropout == 1.0:
keep_prob = tf.constant([0.2, 0.3, 0.5], dtype=tf.float32)
else:
keep_prob = tf.constant([1.0, 1.0, 1.0], dtype=tf.float32)
weights_1 = weight_variable([cnn1,cnn1,3, hidden_size[0]])
biases_1 = bias_variable([hidden_size[0]])
weights_2 = weight_variable([cnn2,cnn2,hidden_size[0], hidden_size[1]])
biases_2 = bias_variable([hidden_size[1]])
weights_3 = weight_variable([cnn3,cnn3,hidden_size[1], hidden_size[2]])
biases_3 = bias_variable([hidden_size[2]])
weights_4 = weight_variable([56 * 56 * hidden_size[2], hidden_size[3]])
biases_4 = bias_variable([hidden_size[3]])
weights_5 = weight_variable([hidden_size[3], num_labels])
biases_5 = bias_variable([num_labels])
def model(data):
with tf.device('/gpu:1'):
train_hidden_1 = tf.nn.relu(conv2d(data, weights_1) + biases_1)
train_hidden_2 = max_pool_2x2(tf.nn.relu(conv2d(train_hidden_1, weights_2) + biases_2))
train_hidden_2_drop = tf.nn.dropout(train_hidden_2, keep_prob[0])
train_hidden_3 = max_pool_2x2(tf.nn.relu(conv2d(train_hidden_2_drop, weights_3) + biases_3))
train_hidden_3_drop = tf.nn.dropout(train_hidden_3, keep_prob[1])
train_hidden_3_drop = tf.reshape(train_hidden_3_drop,[-1, 56 * 56 * hidden_size[2]])
train_hidden_4 = tf.nn.relu(tf.matmul(train_hidden_3_drop, weights_4) + biases_4)
train_hidden_4_drop = tf.nn.dropout(train_hidden_4, keep_prob[2])
logits = tf.matmul(train_hidden_4_drop, weights_5) + biases_5
return logits
t_train_labels = tf.argmax(y_, 1) # From one-hot (one and zeros) vectors to values
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=model(x), labels=t_train_labels))
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
y = tf.nn.softmax(model(x))
### RUNNING
print("log address: %s" % (log_addr))
#num_steps = 10001
times_repeat = 20 # number of epochs
batch_size = 100
with tf.Session(graph=graphCNN,config=tf.ConfigProto(log_device_placement=True)) as session:
tf.initialize_all_variables().run()
saver = tf.train.Saver(max_to_keep=0)
writemyfile("---ARCHIVES_M1----")
mytime = strftime("%Y-%m-%d %H:%M:%S", time.localtime())
writemyfile(str("\nTime: %s \nLayers: %d,%d,%d \epochs: %d" % (mytime,cnn1,cnn2,cnn3,times_repeat)))
writemyfile("Train files:" + str(listfiles_train))
writemyfile("Valid files:" + str(listfiles_valid))
writemyfile("Test files:" + str(listfiles_test))
print("Model will be saved in file: %s" % save_model_path)
writemyfile(str("Model will be saved in file: %s" % save_model_path))
### TRAINING & VALIDATION
valid_accuracies_epochs = np.array([])
for time_repeat in range(times_repeat):
print("- time_repeat:",time_repeat)
writemyfile("- time_repeat:"+str(time_repeat))
for file_train in listfiles_train:
file_train_id = int(file_train[0:len(file_train)-4])
time_start_this_file = time.time()
#LOAD DATA
print("- - file:",file_train_id, end=' ')
writemyfile("- - file:" + str(file_train_id))
Data_train, Label_train= PrepareData(file_train)
for step in range(0,len(Data_train)-batch_size,batch_size):
batch_data = Data_train[step:step+batch_size]
batch_labels = Label_train[step:step+batch_size]
feed_dict = {x : batch_data, y_ : batch_labels, dropout: 1.0}
opti, l, predictions = session.run([optimizer, loss, y], feed_dict=feed_dict)
train_accuracies = np.array([])
for index_tr_accu in range(0,len(Data_train)-SIZE_BATCH_VALID,SIZE_BATCH_VALID):
current_predictions = y.eval(feed_dict={x: Data_train[index_tr_accu:index_tr_accu+SIZE_BATCH_VALID],dropout: 0.0})
current_accuracy = accuracy(current_predictions, Label_train[index_tr_accu:index_tr_accu+SIZE_BATCH_VALID])
train_accuracies = np.r_[train_accuracies,current_accuracy]
train_accuracy = train_accuracies.mean()
print("batch accu: %.2f%%" %(train_accuracy),end=" | ")
writemyfile("batch accu: %.2f%%" %(train_accuracy))
time_done_this_file = time.time() - time_start_this_file
print("time: %.2fs" % (time_done_this_file))
writemyfile("time: %.2fs" % (time_done_this_file))
# save model
model_addr = save_model_path + "model335" + "-epoch-" + str(time_repeat) + ".ckpt"
save_path = saver.save(session, model_addr,) # max_to_keep default was 5
mytime = strftime("%Y-%m-%d %H:%M:%S", time.localtime())
print("epoch finished at %s \n model address: %s" % (mytime,model_addr))
writemyfile("epoch finished at %s \n model address: %s" % (mytime,model_addr))
# validation
valid_accuracies = np.array([])
for file_valid in listfiles_valid:
file_valid_id = int(file_valid[0:len(file_valid)-4])
Data_valid, Label_valid = PrepareData(file_valid)
for index_vl_accu in range(0,len(Data_valid)-SIZE_BATCH_VALID,SIZE_BATCH_VALID):
current_predictions = y.eval(feed_dict={x: Data_valid[index_vl_accu:index_vl_accu+SIZE_BATCH_VALID],dropout: 0.0})
current_accuracy = accuracy(current_predictions, Label_valid[index_vl_accu:index_vl_accu+SIZE_BATCH_VALID])
valid_accuracies = np.r_[valid_accuracies,current_accuracy]
valid_accuracy = valid_accuracies.mean()
print("epoch %d - valid accu: %.2f%%" %(time_repeat,valid_accuracy))
writemyfile("epoch %d - valid accu: %.2f%%" %(time_repeat,valid_accuracy))
valid_accuracies_epochs = np.hstack([valid_accuracies_epochs,valid_accuracy])
print('Done!!')
writemyfile(str('Done!!'))
session.close()
Update: I found cifar10_multi_gpu_train.py seems to be a good example for training with multi GPUs, but honestly I don't know how to apply on my case.
I think you need to change
def model(data):
with tf.device('/gpu:1'):
to:
def model(data):
for d in ['/gpu:0', '/gpu:1']:
with tf.device(d):
and ditch the line with tf.device('/gpu:0'):
Since at the first with tf.device... you are only doing initiation
of variables and then you are resetting your devices with the next with tf.device.
Let me know if this works since I can't test it.