April 21, 2019 Domain Adaptation
아래의 코드는 Unsupervised Domain Adaptation by Backpropagation의 내용을 코드로 정리한 내용입니다.
Code
import tensorflow as tf
import numpy as np
import random
import os
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1 import ImageGrid
from tensorflow.python.framework import ops
import pickle
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
WARNING:tensorflow:From <ipython-input-2-71e12f4bac70>:1: read_data_sets (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use alternatives such as official/mnist/dataset.py from tensorflow/models.
WARNING:tensorflow:From /home/donghwa/lab/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:260: maybe_download (from tensorflow.contrib.learn.python.learn.datasets.base) is deprecated and will be removed in a future version.
Instructions for updating:
Please write your own downloading logic.
WARNING:tensorflow:From /home/donghwa/lab/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:262: extract_images (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use tf.data to implement this functionality.
Extracting MNIST_data/train-images-idx3-ubyte.gz
WARNING:tensorflow:From /home/donghwa/lab/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:267: extract_labels (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use tf.data to implement this functionality.
Extracting MNIST_data/train-labels-idx1-ubyte.gz
WARNING:tensorflow:From /home/donghwa/lab/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:110: dense_to_one_hot (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use tf.one_hot on tensors.
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz
WARNING:tensorflow:From /home/donghwa/lab/lib/python3.5/site-packages/tensorflow/contrib/learn/python/learn/datasets/mnist.py:290: DataSet.__init__ (from tensorflow.contrib.learn.python.learn.datasets.mnist) is deprecated and will be removed in a future version.
Instructions for updating:
Please use alternatives such as official/mnist/dataset.py from tensorflow/models.
load mnist
mnist.train.images.shape
(55000, 784)
# Process MNIST (55000, 784)
mnist_train = (mnist.train.images > 0).reshape(55000, 28, 28, 1).astype(np.uint8) * 255
mnist_train = np.concatenate([mnist_train, mnist_train, mnist_train], 3)
mnist_test = (mnist.test.images > 0).reshape(10000, 28, 28, 1).astype(np.uint8) * 255
mnist_test = np.concatenate([mnist_test, mnist_test, mnist_test], 3)
mnist_train.shape
(55000, 28, 28, 3)
Load MNIST-M
- 여기서
mnistm_data.pkl를 다운 받을 수 있음
# Load MNIST-M
mnistm = pickle.load(open('mnistm_data.pkl', 'rb'))
mnistm_train = mnistm['train']
mnistm_test = mnistm['test']
mnistm_valid = mnistm['valid']
- mean pixel
# average of each RGB
pixel_mean = np.vstack([mnist_train, mnistm_train]).mean((0, 1, 2))
pixel_mean.shape
(3,)
# Create a mixed dataset for TSNE visualization
num_test = 500
combined_test_imgs = np.vstack([mnist_test[:num_test], mnistm_test[:num_test]])
combined_test_labels = np.vstack([mnist.test.labels[:num_test], mnist.test.labels[:num_test]])
combined_test_domain = np.vstack([np.tile([1., 0.], [num_test, 1]),
np.tile([0., 1.], [num_test, 1])])
def imshow_grid(images, shape=[2, 8]):
"""Plot images in a grid of a given shape."""
fig = plt.figure(1)
grid = ImageGrid(fig, 111, nrows_ncols=shape, axes_pad=0.05)
size = shape[0] * shape[1]
for i in range(size):
grid[i].axis('off')
grid[i].imshow(images[i]) # The AxesGrid object work as a list of axes.
plt.show()
imshow_grid(mnist_train)
imshow_grid(mnistm_train)
- Placeholders 정의
batch_size = 64
X = tf.placeholder(tf.uint8, [None, 28, 28, 3])
y = tf.placeholder(tf.float32, [None, 10])
domainID = tf.placeholder(tf.float32, [None, 2])
lb = tf.placeholder(tf.float32, []) # lambda
lr = tf.placeholder(tf.float32, []) # learning rate
is_training = tf.placeholder(tf.bool, [])
- Input normalize
X_input = (tf.cast(X, tf.float32) - pixel_mean) / 255.
X_input
<tf.Tensor 'truediv:0' shape=(?, 28, 28, 3) dtype=float32>
featureExtractor
- Conv block 정의
def conv_layer(inputs,
name,
kernelSize,
inChannel,
outChannel,
stride = 1):
with tf.variable_scope(name) as name:
convWeights = tf.get_variable(
"W",
shape= [kernelSize, kernelSize, inChannel, outChannel],
initializer=tf.initializers.truncated_normal(stddev=0.1)
)
convBiases = tf.get_variable(
"b",
shape=[outChannel],
initializer=tf.constant_initializer(0.1)
)
conv = tf.nn.conv2d(inputs, convWeights, [1,stride,stride,1], padding='SAME')
convBias = tf.nn.bias_add(conv, convBiases)
relu = tf.nn.relu(convBias)
return relu
with tf.variable_scope('feature_extractor') as scope:
# Conv1: n x 28(=(28-5+4)/1+1) x 28 x 32 ;
conv1 = conv_layer(X_input, kernelSize=5, name= "conv1", inChannel=3, outChannel=32)
# Pool1: n x 14(=(28-2)/2+1) x 14 x 32
pool1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool1')
# Conv2: n x 14(=(14-5+4)/1+1) x 14 x 48
conv2 = conv_layer(pool1, kernelSize=5, name= "conv2", inChannel=32, outChannel=48)
# Pool2: n x 7(=(14-2)/2+1) x 7 x 48
pool2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool2')
# domain-invariant feature
feature = tf.reshape(pool2, [-1, np.prod(pool2.shape.as_list()[1:])])
feature
<tf.Tensor 'feature_extractor/Reshape:0' shape=(?, 2352) dtype=float32>
labelPredictor
- Train할때는 source/target domain을 구분하여 학습하고, Test할때는 데이터를 전부 학습해야한다.
- tf.cond를 사용하여 ifelse조건에 따라 함수1(for train) or 함수2(for test)를 반환한다.
-
그 함수의 형태로 바꿔주는 간단한 방법은 lambda를 사용하는 것이다.
- when training
- tf.slice를 사용하여 soure데이터와 target데이터를 나눠준다.
[0, 0]: location[batch_size // 2, -1]: location으로부터batch_size // 2행만큼, `-1(2352) 열만큼 데이터를 잘라냄- 결국 주어진 데이터를 행으로 반으로 자르는 역할
# when training
source_features = lambda: tf.slice(feature, [0, 0], [batch_size // 2, -1])
source_labels = lambda: tf.slice(y, [0, 0], [batch_size // 2, -1])
# when testing
all_features = lambda: feature
all_labels = lambda: y
# features/labels
training_features = tf.cond(is_training, source_features, all_features)
true_labels = tf.cond(is_training, source_labels, all_labels)
def fc_layer(inputs,
name,
in_channel,
out_channel, act_output = True):
with tf.variable_scope(name) as name:
fcWeights = tf.get_variable(
"W",
shape= [in_channel, out_channel],
initializer=tf.initializers.truncated_normal(stddev=0.1)
)
fcBiases = tf.get_variable(
"b",
shape=[out_channel],
initializer=tf.constant_initializer(0.1)
)
logit = tf.nn.bias_add(tf.matmul(inputs, fcWeights), fcBiases)
if act_output==True:
return tf.nn.relu(logit)
return logit
- feature dimensions
feature.shape.as_list()[-1]
2352
with tf.variable_scope('label_predictor') as scope:
label_fc1_output= fc_layer(training_features, name ="fc1", in_channel= feature.shape.as_list()[-1], out_channel=100, act_output=True)
label_fc2_output= fc_layer(label_fc1_output, name ="fc2", in_channel= 100, out_channel=100, act_output=True)
label_fc3_output= fc_layer(label_fc2_output, name ="fc3", in_channel= 100, out_channel=10, act_output=False)
label_fc3_output
<tf.Tensor 'label_predictor/fc3/BiasAdd:0' shape=(?, 10) dtype=float32>
label_pred = tf.nn.softmax(label_fc3_output, axis = -1)
label_pred
<tf.Tensor 'Softmax:0' shape=(?, 10) dtype=float32>
# logit: label_fc3_output
pred_loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=label_fc3_output, labels=true_labels)
pred_loss
<tf.Tensor 'softmax_cross_entropy_with_logits/Reshape_2:0' shape=(?,) dtype=float32>
domainPredictor
- Gradient reverse layer를 구하는 방법을 살펴보자
- Forward propagation: identity에 의해서 같은 값이 나옴
- backward propagation: negative gradient(-lambda * gradient)이 계산됨
class FlipGradientBuilder(object):
def __init__(self):
self.num_calls = 0
def __call__(self, x, lb=1.0):
grad_name = "FlipGradient%d" % self.num_calls
@ops.RegisterGradient(grad_name)
def _flip_gradients(op, grad):
return [tf.negative(grad) * lb]
g = tf.get_default_graph()
with g.gradient_override_map({"Identity": grad_name}):
y = tf.identity(x) # copy for assign op
self.num_calls += 1
return y
- 위의
flip gradient를 적용하면 아래와 같다.lb(lambda)는 label loss의grad와 domain loss의-grad의 상대적 중요도를 의미한다.
flip_gradient = FlipGradientBuilder()
with tf.variable_scope('domain_predictor'):
# Flip the gradient when backpropagating through this operation
grad_reversed_feature = flip_gradient(feature, lb)
domain_fc1_output= fc_layer(grad_reversed_feature, name ="fc1", in_channel= feature.shape.as_list()[-1], out_channel=100, act_output=True)
domain_fc2_output= fc_layer(domain_fc1_output, name ="fc2", in_channel= 100, out_channel=2, act_output=False)
domain_pred = tf.nn.softmax(domain_fc2_output)
domain_loss = tf.nn.softmax_cross_entropy_with_logits_v2(logits=domain_fc2_output, labels=domainID)
domain_loss
<tf.Tensor 'softmax_cross_entropy_with_logits_1/Reshape_2:0' shape=(?,) dtype=float32>
Loss
# loss
predLoss = tf.reduce_mean(pred_loss)
domainLoss = tf.reduce_mean(domain_loss)
totalLoss = predLoss + domainLoss
# train op
# label_train_op = tf.train.AdamOptimizer(0.9).minimize(predLoss)
DA_train_op = tf.train.MomentumOptimizer(lr,0.9).minimize(totalLoss)
# Accuacy
correct_label_pred = tf.equal(tf.argmax(true_labels, 1), tf.argmax(label_pred, 1)) #boolean
label_acc = tf.reduce_mean(tf.cast(correct_label_pred, tf.float32)) #scalar
correct_domain_pred = tf.equal(tf.argmax(domainID, 1), tf.argmax(domain_pred, 1))
domain_acc = tf.reduce_mean(tf.cast(correct_domain_pred, tf.float32))
Session on
sess=tf.Session()
sess.run(tf.global_variables_initializer())
def shuffle_aligned_list(pairs):
"""Shuffle arrays in a list by shuffling each array identically."""
num = pairs[0].shape[0]
pm_idx = np.random.permutation(num)
return [pair[pm_idx] for pair in pairs]
def batch_generator(pairs, batch_size, shuffle=True):
"""Generate batches of data.
Given a list of array-like objects, generate batches of a given
size by yielding a list of array-like objects corresponding to the
same slice of each input.
"""
if shuffle:
pairs = shuffle_aligned_list(pairs)
batch_num = 0
while True:
if batch_num * batch_size + batch_size >= len(pairs[0]):
batch_num = 0
if shuffle:
pairs = shuffle_aligned_list(pairs)
start = batch_num * batch_size
end = start + batch_size
batch_num += 1
yield [pair[start:end] for pair in pairs]
- an half of source and target domain examples
- source/target domain examples을 반씩 잘라 붙여야 사전에 설정해 놓은 batch size랑 일치한다.
- [(32, 28, 28, 3), (32, 10)]
gen_source_batch = batch_generator(pairs=[mnist_train, mnist.train.labels], batch_size= batch_size // 2)
gen_target_batch = batch_generator(pairs=[mnistm_train, mnist.train.labels], batch_size= batch_size // 2)
- All of source/target domain examples
- [(64, 28, 28, 3), (64, 10)]
gen_source_only_batch = batch_generator(pairs=[mnist_train, mnist.train.labels],batch_size= batch_size)
gen_target_only_batch = batch_generator(pairs=[mnistm_train, mnist.train.labels], batch_size=batch_size)
- domain labels if source or target
domain_labels = np.vstack([np.tile([1., 0.], [batch_size // 2, 1]),
np.tile([0., 1.], [batch_size // 2, 1])])
# Training model
num_steps = 8600
for i in range(num_steps):
# Adaptation param and learning rate schedule as described in the paper
p = float(i) / num_steps # smaller over time
_lb = 2. / (1. + np.exp(-10. * p)) - 1 # smaller over time
_lr = 0.01 / (1. + 10 * p)**0.75
# Training step
XS, yS = next(gen_source_batch)
XT, yT = next(gen_target_batch)
X_feed = np.vstack([XS, XT])
y_feed = np.vstack([yS, yT])
_, batch_loss, dloss, ploss, d_acc, p_acc = sess.run(
[DA_train_op, totalLoss, domainLoss, predLoss, domain_acc, label_acc],
feed_dict = {X: X_feed,
y: y_feed,
domainID: domain_labels,
is_training: True,
lb: _lb,
lr: _lr})
if i % 100 == 0:
print('loss: {} d_acc: {} p_acc: {} p: {} lb: {} lr: {}'.format(
batch_loss, d_acc, p_acc, p, _lb, _lr))
loss: 4.825463771820068 d_acc: 0.5 p_acc: 0.0625 p: 0.0 lb: 0.0 lr: 0.01
loss: 0.40940043330192566 d_acc: 0.90625 p_acc: 0.9375 p: 0.011627906976744186 lb: 0.05807411547586794 lr: 0.009208108196143174
loss: 0.5818338394165039 d_acc: 0.890625 p_acc: 0.875 p: 0.023255813953488372 lb: 0.11575782577418603 lr: 0.008548589038554535
loss: 0.5852369666099548 d_acc: 0.796875 p_acc: 0.9375 p: 0.03488372093023256 lb: 0.1726711536624863 lr: 0.007989697680708598
loss: 0.5564760565757751 d_acc: 0.90625 p_acc: 0.9375 p: 0.046511627906976744 lb: 0.22845439143629886 lr: 0.0075092390442056635
loss: 0.35558179020881653 d_acc: 0.875 p_acc: 0.96875 p: 0.05813953488372093 lb: 0.28277682569099527 lr: 0.0070911987391239035
loss: 0.2507930397987366 d_acc: 0.9375 p_acc: 1.0 p: 0.06976744186046512 lb: 0.33534391863054624 lr: 0.006723713451795381
loss: 0.25439000129699707 d_acc: 0.953125 p_acc: 0.96875 p: 0.08139534883720931 lb: 0.3859026564904293 lr: 0.006397796023621206
loss: 0.5283358097076416 d_acc: 0.859375 p_acc: 0.9375 p: 0.09302325581395349 lb: 0.43424492823161986 lr: 0.006106505778508959
loss: 0.4712458550930023 d_acc: 0.859375 p_acc: 0.90625 p: 0.10465116279069768 lb: 0.4802089471455322 lr: 0.00584439199954653
loss: 0.4825291633605957 d_acc: 0.890625 p_acc: 0.96875 p: 0.11627906976744186 lb: 0.5236788585597902 lr: 0.0056071106987353935
loss: 0.19707489013671875 d_acc: 0.90625 p_acc: 1.0 p: 0.12790697674418605 lb: 0.5645827773148586 lr: 0.005391154585323163
loss: 0.4898703396320343 d_acc: 0.78125 p_acc: 1.0 p: 0.13953488372093023 lb: 0.6028895635619975 lr: 0.005193658897076379
loss: 0.3004189431667328 d_acc: 0.9375 p_acc: 1.0 p: 0.1511627906976744 lb: 0.6386046745549958 lr: 0.005012259239308981
loss: 0.4147544503211975 d_acc: 0.859375 p_acc: 0.96875 p: 0.16279069767441862 lb: 0.6717654275930591 lr: 0.004844985806046933
loss: 0.2904179096221924 d_acc: 0.875 p_acc: 0.96875 p: 0.1744186046511628 lb: 0.7024359819836457 lr: 0.004690183518397721
loss: 0.6488478183746338 d_acc: 0.84375 p_acc: 0.90625 p: 0.18604651162790697 lb: 0.730702303851418 lr: 0.0045464509301225125
loss: 0.2889407277107239 d_acc: 0.90625 p_acc: 0.96875 p: 0.19767441860465115 lb: 0.75666732465968 lr: 0.004412592926309448
loss: 0.34746885299682617 d_acc: 0.9375 p_acc: 0.90625 p: 0.20930232558139536 lb: 0.7804464491564882 lr: 0.004287583697558238
loss: 0.40334779024124146 d_acc: 0.84375 p_acc: 1.0 p: 0.22093023255813954 lb: 0.8021635162227865 lr: 0.004170537464590465
loss: 0.5859626531600952 d_acc: 0.953125 p_acc: 0.90625 p: 0.23255813953488372 lb: 0.8219472701703336 lr: 0.00406068511562926
loss: 0.5954075455665588 d_acc: 0.796875 p_acc: 0.9375 p: 0.2441860465116279 lb: 0.8399283622203215 lr: 0.003957355402193208
loss: 0.5475267171859741 d_acc: 0.8125 p_acc: 0.9375 p: 0.2558139534883721 lb: 0.8562368727213061 lr: 0.003859959683449113
loss: 0.8015364408493042 d_acc: 0.65625 p_acc: 0.9375 p: 0.26744186046511625 lb: 0.8710003237511923 lr: 0.0037679794579745665
loss: 0.9847149848937988 d_acc: 0.671875 p_acc: 0.9375 p: 0.27906976744186046 lb: 0.8843421381310939 lr: 0.0036809561034616242
loss: 0.5318312048912048 d_acc: 0.828125 p_acc: 0.90625 p: 0.29069767441860467 lb: 0.8963804933094455 lr: 0.0035984823790740587
loss: 1.4411591291427612 d_acc: 0.75 p_acc: 0.8125 p: 0.3023255813953488 lb: 0.907227515730854 lr: 0.0035201953452889123
loss: 1.3302115201950073 d_acc: 0.53125 p_acc: 0.875 p: 0.313953488372093 lb: 0.9169887619362538 lr: 0.0034457704314723612
loss: 0.5538606643676758 d_acc: 0.71875 p_acc: 1.0 p: 0.32558139534883723 lb: 0.9257629356552557 lr: 0.003374916438763681
loss: 0.8089284896850586 d_acc: 0.53125 p_acc: 0.9375 p: 0.3372093023255814 lb: 0.9336417946475963 lr: 0.003307371309778185
loss: 0.6464406847953796 d_acc: 0.65625 p_acc: 0.96875 p: 0.3488372093023256 lb: 0.9407102063254855 lr: 0.0032428985305833465
loss: 0.892132043838501 d_acc: 0.796875 p_acc: 0.875 p: 0.36046511627906974 lb: 0.9470463167220862 lr: 0.003181284056820732
loss: 0.5464159250259399 d_acc: 0.796875 p_acc: 0.96875 p: 0.37209302325581395 lb: 0.9527218027997246 lr: 0.0031223336765527007
loss: 0.8238319158554077 d_acc: 0.796875 p_acc: 0.96875 p: 0.38372093023255816 lb: 0.9578021831783621 lr: 0.003065870738750244
loss: 0.22796763479709625 d_acc: 0.890625 p_acc: 1.0 p: 0.3953488372093023 lb: 0.962347166972511 lr: 0.0030117341893097
loss: 0.7139698266983032 d_acc: 0.765625 p_acc: 0.96875 p: 0.4069767441860465 lb: 0.9664110244879942 lr: 0.002959776866846464
loss: 0.32875534892082214 d_acc: 0.84375 p_acc: 1.0 p: 0.4186046511627907 lb: 0.9700429670347053 lr: 0.002909864018835692
loss: 0.5904496908187866 d_acc: 0.875 p_acc: 0.90625 p: 0.43023255813953487 lb: 0.9732875260777021 lr: 0.002861872005390412
loss: 0.4229314923286438 d_acc: 0.875 p_acc: 1.0 p: 0.4418604651162791 lb: 0.9761849244171494 lr: 0.0028156871634225067
loss: 0.3996061086654663 d_acc: 0.796875 p_acc: 1.0 p: 0.45348837209302323 lb: 0.9787714341094471 lr: 0.0027712048083815533
loss: 0.4992428421974182 d_acc: 0.828125 p_acc: 0.96875 p: 0.46511627906976744 lb: 0.9810797174732149 lr: 0.002728328354412792
loss: 0.5148041248321533 d_acc: 0.828125 p_acc: 0.9375 p: 0.47674418604651164 lb: 0.9831391488202712 lr: 0.002686968536776859
loss: 0.5221948027610779 d_acc: 0.765625 p_acc: 0.96875 p: 0.4883720930232558 lb: 0.9849761155658179 lr: 0.0026470427228549257
loss: 0.6307872533798218 d_acc: 0.78125 p_acc: 0.90625 p: 0.5 lb: 0.9866142981514305 lr: 0.0026084743001221454
loss: 0.6382717490196228 d_acc: 0.828125 p_acc: 0.90625 p: 0.5116279069767442 lb: 0.9880749288014745 lr: 0.002571192131188139
loss: 0.4289991557598114 d_acc: 0.90625 p_acc: 0.96875 p: 0.5232558139534884 lb: 0.9893770295647741 lr: 0.002535130067438327
loss: 0.40435507893562317 d_acc: 0.890625 p_acc: 0.96875 p: 0.5348837209302325 lb: 0.9905376303999733 lr: 0.002500226514014457
loss: 0.6215903759002686 d_acc: 0.78125 p_acc: 0.96875 p: 0.5465116279069767 lb: 0.9915719682712341 lr: 0.0024664240398871913
loss: 0.5021497011184692 d_acc: 0.796875 p_acc: 1.0 p: 0.5581395348837209 lb: 0.9924936683523062 lr: 0.0024336690276309993
loss: 0.5572361946105957 d_acc: 0.75 p_acc: 0.96875 p: 0.5697674418604651 lb: 0.9933149085094493 lr: 0.0024019113582383557
loss: 0.5989366769790649 d_acc: 0.65625 p_acc: 1.0 p: 0.5813953488372093 lb: 0.9940465682614799 lr: 0.0023711041269283283
loss: 1.0323117971420288 d_acc: 0.6875 p_acc: 0.8125 p: 0.5930232558139535 lb: 0.9946983634099507 lr: 0.00234120338643174
loss: 0.9565933346748352 d_acc: 0.796875 p_acc: 0.90625 p: 0.6046511627906976 lb: 0.9952789675032998 lr: 0.002312167914685915
loss: 0.4626711308956146 d_acc: 0.796875 p_acc: 0.96875 p: 0.6162790697674418 lb: 0.9957961212529436 lr: 0.0022839590042587057
loss: 0.4873444736003876 d_acc: 0.8125 p_acc: 1.0 p: 0.627906976744186 lb: 0.9962567309623462 lr: 0.0022565402711539955
loss: 0.8458352088928223 d_acc: 0.75 p_acc: 0.90625 p: 0.6395348837209303 lb: 0.996666956966364 lr: 0.0022298774809375384
loss: 0.7572154402732849 d_acc: 0.796875 p_acc: 0.90625 p: 0.6511627906976745 lb: 0.9970322930109379 lr: 0.0022039383903697716
loss: 0.45236918330192566 d_acc: 0.859375 p_acc: 0.96875 p: 0.6627906976744186 lb: 0.9973576374349025 lr: 0.0021786926029468733
loss: 0.35806888341903687 d_acc: 0.890625 p_acc: 0.96875 p: 0.6744186046511628 lb: 0.9976473569480822 lr: 0.0021541114369377466
loss: 0.435672402381897 d_acc: 0.84375 p_acc: 0.96875 p: 0.686046511627907 lb: 0.9979053437342209 lr: 0.002130167804666827
loss: 0.5702878832817078 d_acc: 0.71875 p_acc: 1.0 p: 0.6976744186046512 lb: 0.9981350665445223 lr: 0.0021068361019340996
loss: 0.7712510228157043 d_acc: 0.65625 p_acc: 0.96875 p: 0.7093023255813954 lb: 0.998339616388191 lr: 0.002084092106587389
loss: 0.45937344431877136 d_acc: 0.859375 p_acc: 0.96875 p: 0.7209302325581395 lb: 0.9985217473707251 lr: 0.002061912885370307
loss: 0.8559670448303223 d_acc: 0.640625 p_acc: 0.96875 p: 0.7325581395348837 lb: 0.9986839131789393 lr: 0.0020402767082642755
loss: 0.7581861019134521 d_acc: 0.71875 p_acc: 0.9375 p: 0.7441860465116279 lb: 0.9988282996638527 lr: 0.0020191629696266573
loss: 0.8087642192840576 d_acc: 0.71875 p_acc: 0.9375 p: 0.7558139534883721 lb: 0.9989568539285347 lr: 0.001998552115500608
loss: 0.7783613204956055 d_acc: 0.578125 p_acc: 0.96875 p: 0.7674418604651163 lb: 0.9990713102877162 lr: 0.0019784255765372813
loss: 0.5618288516998291 d_acc: 0.59375 p_acc: 1.0 p: 0.7790697674418605 lb: 0.9991732134291553 lr: 0.0019587657060284253
loss: 0.5146697163581848 d_acc: 0.640625 p_acc: 1.0 p: 0.7906976744186046 lb: 0.9992639390733074 lr: 0.0019395557225983058
loss: 0.6340583562850952 d_acc: 0.625 p_acc: 1.0 p: 0.8023255813953488 lb: 0.9993447123974792 lr: 0.0019207796571489855
loss: 0.7205818891525269 d_acc: 0.640625 p_acc: 0.9375 p: 0.813953488372093 lb: 0.999416624463169 lr: 0.0019024223036931153
loss: 0.634271502494812 d_acc: 0.625 p_acc: 0.96875 p: 0.8255813953488372 lb: 0.9994806468604833 lr: 0.0018844691737440408
loss: 0.6444191336631775 d_acc: 0.578125 p_acc: 0.96875 p: 0.8372093023255814 lb: 0.9995376447611286 lr: 0.0018669064539648453
loss: 0.7390989065170288 d_acc: 0.59375 p_acc: 0.96875 p: 0.8488372093023255 lb: 0.9995883885513366 lr: 0.0018497209668063268
loss: 0.5952032804489136 d_acc: 0.6875 p_acc: 1.0 p: 0.8604651162790697 lb: 0.9996335641979495 lr: 0.0018329001338892768
loss: 0.5313435792922974 d_acc: 0.75 p_acc: 1.0 p: 0.872093023255814 lb: 0.9996737824846389 lr: 0.0018164319419091504
loss: 0.9486538767814636 d_acc: 0.75 p_acc: 0.875 p: 0.8837209302325582 lb: 0.9997095872406112 lr: 0.001800304910861562
loss: 0.5704761147499084 d_acc: 0.640625 p_acc: 1.0 p: 0.8953488372093024 lb: 0.9997414626710781 lr: 0.0017845080644053356
loss: 0.7692083120346069 d_acc: 0.640625 p_acc: 0.96875 p: 0.9069767441860465 lb: 0.9997698398870507 lr: 0.0017690309021962457
loss: 0.6062610745429993 d_acc: 0.71875 p_acc: 0.96875 p: 0.9186046511627907 lb: 0.9997951027215122 lr: 0.0017538633740393809
loss: 0.6653850078582764 d_acc: 0.609375 p_acc: 1.0 p: 0.9302325581395349 lb: 0.9998175929096487 lr: 0.0017389958557213823
loss: 0.6266434788703918 d_acc: 0.640625 p_acc: 1.0 p: 0.9418604651162791 lb: 0.9998376147024313 lr: 0.0017244191263958181
loss: 0.5251486897468567 d_acc: 0.75 p_acc: 1.0 p: 0.9534883720930233 lb: 0.9998554389753282 lr: 0.0017101243474058232
loss: 0.6981379985809326 d_acc: 0.59375 p_acc: 0.96875 p: 0.9651162790697675 lb: 0.9998713068872478 lr: 0.0016961030424379434
loss: 0.718218207359314 d_acc: 0.640625 p_acc: 0.96875 p: 0.9767441860465116 lb: 0.999885433138822 lr: 0.001682347078910014
loss: 0.7044693827629089 d_acc: 0.6875 p_acc: 0.90625 p: 0.9883720930232558 lb: 0.9998980088738034 lr: 0.0016688486505039575
Evaluation
- 학습을 할때는 source/target데이터가 모두 필요하다.
- 평가할때 source 데이터 성능을 볼땐 source만, target 데이터 성능을 볼땐 target만 사용할 수 있다.
- 어떤 domain인지를 평가할때는 당연히 두개의 source/target 데이터가 필요하다.
# Compute final evaluation on test data
source_acc = sess.run(label_acc,
feed_dict={X: mnist_test, y: mnist.test.labels,
is_training: False})
target_acc = sess.run(label_acc,
feed_dict={X: mnistm_test, y: mnist.test.labels,
is_training: False})
test_domain_acc = sess.run(domain_acc,
feed_dict={X: combined_test_imgs,
domainID: combined_test_domain, lb: 1.0})
test_emb = sess.run(feature, feed_dict={X: combined_test_imgs})
test_emb.shape
(1000, 2352)
print('\nDomain adaptation training')
print('Source (MNIST) accuracy:', source_acc)
print('Target (MNIST-M) accuracy:', target_acc)
print('Domain accuracy:', test_domain_acc)
Domain adaptation training
Source (MNIST) accuracy: 0.9807
Target (MNIST-M) accuracy: 0.7198
Domain accuracy: 0.669
- 시각화를 위해 따로 임베딩 vector, domain, label를 뽑아 두었다.
if not os.path.exists('./results'):
os.mkdir('./results')
# images(source: 500, target: 500)
np.save('./results/SourceTargetImages.npy',combined_test_imgs)
# 2352차원을 가진 1000 vector(source: 500, target: 500)
np.save('./results/SourceTargetEmbed.npy', test_emb)
# source or target
np.save('./results/SourceTargetDomain.npy', combined_test_domain)
# mnist label
np.save('./results/SourceTargetLabels.npy', combined_test_labels)
Reference
https://github.com/TGISer/tf-dann
