def test(mnist):
with tf.Graph().as_default() as g:
x = tf.placeholder(tf.float32, [None, fwp.INPUT_NODE])
y_ = tf.placeholder(tf.float32, [None, fwp.OUTPUT_NODE])
y = fwp.fw_propagation(x, None)
ema = tf.train.ExponentialMovingAverage(bkp.EMA_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
correct_predicition = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_predicition, tf.float32))
while True:
with tf.Session() as testSess:
ckpt = tf.train.get_checkpoint_state(bkp.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(testSess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split(
'/')[-1].split('-')[-1]
accuracy_score = testSess.run(accuracy,
feed_dict={
x: mnist.test.images,
y_: mnist.test.labels
})
print "After %s steps, test accuracy is %g." % (
global_step, accuracy_score)
else:
print "No checkpoint data found."
return
time.sleep(TEST_INTVAL_IN_SECS)
def test(mnist):
with tf.Graph().as_default() as g:
x = tf.placeholder(tf.float32, [None, fwp.INPUT_NODE])
y_ = tf.placeholder(tf.float32, [None, fwp.OUTPUT_NODE])
y = fwp.fw_propagation(x, None)
ema = tf.train.ExponentialMovingAverage(bkp.EMA_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
correct_predicition = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_predicition, tf.float32))
img_batch, label_batch = gends.get_tfRecords(TOTAL_TEST_EXAMPLES,
isTrain=False)
while True:
with tf.Session() as testSess:
ckpt = tf.train.get_checkpoint_state(bkp.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(testSess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split(
'/')[-1].split('-')[-1]
#multi-threads start
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=testSess,
coord=coord)
xs, y_s = testSess.run([img_batch, label_batch])
#accuracy_score = testSess.run(accuracy, feed_dict = {x: mnist.test.images, y_: mnist.test.labels})
accuracy_score = testSess.run(accuracy,
feed_dict={
x: xs,
y_: y_s
})
print "After %s steps, test accuracy is %g." % (
global_step, accuracy_score)
#multi-threads end
coord.request_stop()
coord.join(threads)
else:
print "No checkpoint data found."
return
time.sleep(TEST_INTVAL_IN_SECS)
def test(mnist):
with tf.Graph().as_default() as g:
#x = tf.placeholder(tf.float32, [None, fwp.INPUT_NODE])
#同样，调整x的维度
print mnist.test.num_examples
x = tf.placeholder(tf.float32, [
MNIST_MODF_TEST_CNT, fwp.IMAGE_XY_RES, fwp.IMAGE_XY_RES,
fwp.IMAGE_CHANS
])
y_ = tf.placeholder(tf.float32, [None, fwp.OUTPUT_NODE])
y = fwp.fw_propagation(x, False, None)
ema = tf.train.ExponentialMovingAverage(bkp.EMA_DECAY)
ema_restore = ema.variables_to_restore()
saver = tf.train.Saver(ema_restore)
correct_predicition = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_predicition, tf.float32))
while True:
with tf.Session() as testSess:
ckpt = tf.train.get_checkpoint_state(bkp.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(testSess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split(
'/')[-1].split('-')[-1]
test_imgs = mnist.test.next_batch(MNIST_MODF_TEST_CNT)
mnist_test_images_reshaped = np.reshape(
test_imgs[0], [
MNIST_MODF_TEST_CNT, fwp.IMAGE_XY_RES,
fwp.IMAGE_XY_RES, fwp.IMAGE_CHANS
])
accuracy_score = testSess.run(
accuracy,
feed_dict={
x: mnist_test_images_reshaped,
y_: test_imgs[1]
})
print "After %s steps, test accuracy is %g." % (
global_step, accuracy_score)
else:
print "No checkpoint data found."
return
time.sleep(TEST_INTVAL_IN_SECS)
def FCN_restore_model(imgarr_pped):
with tf.Graph().as_default() as tg:
x = tf.placeholder(tf.float32, [1, fwp.IMAGE_XY_RES, fwp.IMAGE_XY_RES, fwp.IMAGE_CHANS])
y = fwp.fw_propagation(x, False, None)
resultV = tf.argmax(y, 1)
var_avg = tf.train.ExponentialMovingAverage(bkp.EMA_DECAY)
var_to_restore = var_avg.variables_to_restore()
saver = tf.train.Saver(var_to_restore)
with tf.Session() as fcnSess:
ckpt = tf.train.get_checkpoint_state(bkp.MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(fcnSess, ckpt.model_checkpoint_path)
resultV = fcnSess.run(resultV, feed_dict = {x:imgarr_pped})
return resultV
else:
print "No checkpoint data found."
return -1
def bk_propagation(mnist):
#x = tf.placeholder(tf.float32, shape = (None, fwp.INPUT_NODE))
#在lenet5卷积网络中，输入x是一个四阶张量：[图片个数，图片X分辨率，图片Y分辨率，通道数]
x = tf.placeholder(tf.float32, [BATCH_SIZE, fwp.IMAGE_XY_RES, fwp.IMAGE_XY_RES, fwp.IMAGE_CHANS])
y_ = tf.placeholder(tf.float32, shape = (None, fwp.OUTPUT_NODE))
#X, Y_, Y_c = gends.generateds()
y = fwp.fw_propagation(x, True, REGULARIZER)
global_step = tf.Variable(0, trainable = False)
#loss_mse = tf.reduce_mean(tf.square(y - y_))
#loss = loss_mse
ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits = y, labels = tf.argmax(y_, 1))
loss_cem = tf.reduce_mean(ce)
loss = loss_cem
loss_total = loss + tf.add_n(tf.get_collection("losses"))
learning_rate = tf.train.exponential_decay(
LR_BASE,
global_step,
mnist.train.num_examples / BATCH_SIZE,
LR_DECAY,
staircase = True)
#train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss_total)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss_total, global_step = global_step)
ema = tf.train.ExponentialMovingAverage(EMA_DECAY, global_step)
ema_op = ema.apply(tf.trainable_variables())
with tf.control_dependencies([train_step, ema_op]):
train_op = tf.no_op(name = "train")
saver = tf.train.Saver()
with tf.Session() as fcSess:
init_op = tf.global_variables_initializer()
fcSess.run(init_op)
ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(fcSess, ckpt.model_checkpoint_path)
xs, ys = mnist.train.next_batch(BATCH_SIZE)
xs = np.reshape(xs, (BATCH_SIZE, fwp.IMAGE_XY_RES, fwp.IMAGE_XY_RES, fwp.IMAGE_CHANS))
_, loss_v, step = fcSess.run([train_op, loss, global_step], feed_dict = {x: xs, y_: ys})
print "Restored session at %d steps, loss is %g." %(step, loss_v)
else:
step = 0
loss_v = float("inf")
for i in range(step, TOTAL_STEPS):
if i % 100 == 0:
print "After %d steps, loss is %g." %(step, loss_v)
saver.save(fcSess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step = global_step)
xs, ys = mnist.train.next_batch(BATCH_SIZE)
xs = np.reshape(xs, (BATCH_SIZE, fwp.IMAGE_XY_RES, fwp.IMAGE_XY_RES, fwp.IMAGE_CHANS))
_, loss_v, step = fcSess.run([train_op, loss, global_step], feed_dict = {x: xs, y_: ys})
print "After %d steps, loss is %g." %(step, loss_v)
saver.save(fcSess, os.path.join(MODEL_SAVE_PATH, MODEL_NAME), global_step = global_step)
def bk_propagation(mnist):
x = tf.placeholder(tf.float32, shape=(None, fwp.INPUT_NODE))
y_ = tf.placeholder(tf.float32, shape=(None, fwp.OUTPUT_NODE))
#X, Y_, Y_c = gends.generateds()
y = fwp.fw_propagation(x, REGULARIZER)
global_step = tf.Variable(0, trainable=False)
#loss_mse = tf.reduce_mean(tf.square(y - y_))
#loss = loss_mse
ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y,
labels=tf.argmax(
y_, 1))
loss_cem = tf.reduce_mean(ce)
loss = loss_cem
loss_total = loss + tf.add_n(tf.get_collection("losses"))
learning_rate = tf.train.exponential_decay(LR_BASE,
global_step,
TOTAL_NUM_EXAMPLES / BATCH_SIZE,
LR_DECAY,
staircase=True)
#train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss_total)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
loss_total, global_step=global_step)
ema = tf.train.ExponentialMovingAverage(EMA_DECAY, global_step)
ema_op = ema.apply(tf.trainable_variables())
with tf.control_dependencies([train_step, ema_op]):
train_op = tf.no_op(name="train")
saver = tf.train.Saver()
img_batch, label_batch = gends.get_tfRecords(BATCH_SIZE)
with tf.Session() as fcSess:
init_op = tf.global_variables_initializer()
fcSess.run(init_op)
#multi-threads start
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=fcSess, coord=coord)
ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(fcSess, ckpt.model_checkpoint_path)
#xs, ys = mnist.train.next_batch(BATCH_SIZE)
xs, ys = fcSess.run([img_batch, label_batch])
_, loss_v, step = fcSess.run([train_op, loss, global_step],
feed_dict={
x: xs,
y_: ys
})
print "Restored session at %d steps, loss is %g." % (step, loss_v)
else:
step = 0
loss_v = float("inf")
for i in range(step, TOTAL_STEPS):
if i % 1000 == 0:
print "After %d steps, loss is %g." % (step, loss_v)
saver.save(fcSess,
os.path.join(MODEL_SAVE_PATH, MODEL_NAME),
global_step=global_step)
#xs, ys = mnist.train.next_batch(BATCH_SIZE)
xs, ys = fcSess.run([img_batch, label_batch])
_, loss_v, step = fcSess.run([train_op, loss, global_step],
feed_dict={
x: xs,
y_: ys
})
print "After %d steps, loss is %g." % (step, loss_v)
saver.save(fcSess,
os.path.join(MODEL_SAVE_PATH, MODEL_NAME),
global_step=global_step)
#multi-threads end
coord.request_stop()
coord.join(threads)