max_iter = 200000 # 600000
# True = start data generator client, False = use augmented dataset file (deprecated)
use_client_gen = True
WEIGHTS_BEST = "weights.best.h5"
TRAINING_LOG = "training.csv"
LOGS_DIR = "./logs"
def get_last_epoch():
data = pandas.read_csv(TRAINING_LOG)
return max(data['epoch'].values)
model = get_training_model(weight_decay, vgg_norm=True)
from_vgg = dict()
from_vgg['conv1_1'] = 'block1_conv1'
from_vgg['conv1_2'] = 'block1_conv2'
from_vgg['conv2_1'] = 'block2_conv1'
from_vgg['conv2_2'] = 'block2_conv2'
from_vgg['conv3_1'] = 'block3_conv1'
from_vgg['conv3_2'] = 'block3_conv2'
from_vgg['conv3_3'] = 'block3_conv3'
from_vgg['conv3_4'] = 'block3_conv4'
from_vgg['conv4_1'] = 'block4_conv1'
from_vgg['conv4_2'] = 'block4_conv2'
# load previous weights or vgg19 if this is the first run
if os.path.exists(WEIGHTS_BEST):
def letter_probs_to_code(letter_probs):
return "".join(common.CHARS[i] for i in numpy.argmax(letter_probs, axis=1))
if __name__ == "__main__":
weights_cnn = 'weights_20180708_171720_feedback.npz'
weights_yolo = 'model_data/yolo_19000.h5'
start_time = time.time()
prYellow('Carregando Yolov3...')
yolo = yolo.YOLO(weights_yolo)
prGreen('Yolov3: modelo, âncoras e classes carregadas.')
prYellow('Carregando Rede Neural Convolucional (CNN)...')
f = numpy.load(weights_cnn)
param_vals = [f[n] for n in sorted(f.files, key=lambda s: int(s[4:]))]
x, y, params = model.get_training_model()
prGreen('CNN: Modelo e pesos carregados.')
time.sleep(2)
prYellow('Inicializando sessão do Tensorflow...')
with tf.Session(config=tf.ConfigProto()) as sess:
feed_dict = {x: numpy.zeros([1, 64, 128])}
feed_dict.update(dict(zip(params, param_vals)))
y_val = sess.run(y, feed_dict=feed_dict)
out_boxes, out_probs, out_labels = yolo.get_detections(
Image.fromarray(numpy.zeros([100, 100, 3]).astype('uint8')))
prGreen('Sessão do Tensorflow inicializada.')
prYellow('Inicializando câmera...')
camera = PiCamera()
camera.resolution = (1024, 768)
camera.framerate = 16 #original 32
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
digits_loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(
digits_loss) ###ORIGINAL LOSS
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.global_variables_initializer()
#init = tf.initialize_all_variables()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([best, correct, digits_loss],
feed_dict={
x: test_xs,
y_: test_ys
})
#num_correct = numpy.sum(numpy.all(r[0] == r[1], axis=1))
num_correct = numpy.sum(r[0] == r[1])
r_short = (r[0][:30], r[1][:30])
#for b, c in zip(*r_short):
# print("{} <-> {}".format(vec_to_plate(c),vec_to_plate(b)))
#num_p_correct = numpy.sum(r[2] == r[3])
#train_writer.add_summary(mrg, batch_idx)
print("B{:3d} {:2.02f}% loss: {} ".format(
batch_idx, 100. * num_correct / (7 * len(r[0])), r[2]))
#if (100. * num_correct / len(r[0]) > 48.1):
# last_weights = [p.eval() for p in params]
# timestr = time.strftime("%Y%m%d_%H%M%S")
# numpy.savez("weights_{}_result.npz".format(timestr), *last_weights)
def do_batch():
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.80)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
#tf.summary.scalar('loss', loss)
#tf.summary.scalar('digits', digits_loss)
#tf.summary.scalar('presence', presence_loss)
#merged = tf.summary.merge_all()
#train_writer = tf.summary.FileWriter('./logs',sess.graph)
test_xs, test_ys = unzip(
list(read_data("dataset_210618_test/*.png"))[:114])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print("time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx)))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
timestr = time.strftime("%Y%m%d_%H%M%S")
numpy.savez("weights_{}.npz".format(timestr), *last_weights)
return last_weights
stepsize = 136106 #68053 // after each stepsize iterations update learning rate: lr=lr*gamma
max_iter = 200000 # 600000
# True = start data generator client, False = use augmented dataset file (deprecated)
use_client_gen = True
WEIGHTS_BEST = "weights.best.h5"
TRAINING_LOG = "training.csv"
LOGS_DIR = "./logs"
def get_last_epoch():
data = pandas.read_csv(TRAINING_LOG)
return max(data['epoch'].values)
model = get_training_model(weight_decay)
from_vgg = dict()
from_vgg['conv1_1'] = 'block1_conv1'
from_vgg['conv1_2'] = 'block1_conv2'
from_vgg['conv2_1'] = 'block2_conv1'
from_vgg['conv2_2'] = 'block2_conv2'
from_vgg['conv3_1'] = 'block3_conv1'
from_vgg['conv3_2'] = 'block3_conv2'
from_vgg['conv3_3'] = 'block3_conv3'
from_vgg['conv3_4'] = 'block3_conv4'
from_vgg['conv4_1'] = 'block4_conv1'
from_vgg['conv4_2'] = 'block4_conv2'
# load previous weights or vgg19 if this is the first run
if os.path.exists(WEIGHTS_BEST):
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
global params
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
digits_loss, presence_loss, loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([best,
correct,
tf.greater(y[:, 0], 0),
y_[:, 0],
digits_loss,
presence_loss,
loss],
feed_dict={x: test_xs, y_: test_ys})
num_correct = numpy.sum(
numpy.logical_or(
numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5,
r[3] < 0.5)))
recall = numpy.sum(numpy.logical_and(numpy.all(r[0] == r[1], axis=1) , r[3] > 0.5))
end = 300
r_short = (r[0][:end], r[1][:end], r[2][:end], r[3][:end])
#set_trace()
for i,v in enumerate(zip(*r_short)):
print "{} {} {} <-> {} {}".format(i, vec_to_plate(v[1]), v[3],
vec_to_plate(v[0]), float(v[2]))
num_p_correct = numpy.sum(r[2] == r[3])
error_str = ""
for i, v in enumerate(zip(*r_short)):
if v[2] != v[3]:
error_str += "{}:(presence) ".format(i)
elif not numpy.array_equal(v[0], v[1]) and v[3] > 0.5:
error_letters = v[0] != v[1]
for b,c in zip(v[0][error_letters], v[1][error_letters]):
error_str += "{}:{}->{} ".format(i, common.CHARS[c], common.CHARS[b])
print ("B{:3d} num:{:2.02f}% recall:{:2.02f}% presence:{:02.02f}% loss: {} "
"(digits: {}, presence: {}) |{}|").format(
batch_idx,
100. * num_correct / (len(r[0])),
100. * recall / numpy.sum(r[3] > 0.5),
100. * num_p_correct / len(r[2]),
r[6]/len(r[0]),
r[4]/len(r[0]),
r[5]/len(r[0]),
error_str)
# "".join("X "[numpy.array_equal(v[0], v[1]) or (not v[2] and not v[3])]
# for i, v in enumerate(zip(*r_short))))
# " ".join(str(i) for i, v in enumerate(zip(*r_short))
# if not(numpy.array_equal(v[0], v[1]) or (not v[2] and not v[3]))))
def do_batch():
sess.run(train_step,
feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("test/*.png"))[:300])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def get_last_epoch_and_weights_file():
os.makedirs(WEIGHT_DIR, exist_ok=True)
files = [file for file in glob(WEIGHT_DIR + '/weights.*.h5')]
files = [file.split('/')[-1] for file in files]
epochs = [file.split('.')[1] for file in files if file]
epochs = [int(epoch) for epoch in epochs if epoch.isdigit()]
if len(epochs) == 0:
if 'weights.best.h5' in files:
return -1, WEIGHT_DIR + '/weights.best.h5'
else:
ep = max([int(epoch) for epoch in epochs])
return ep, WEIGHT_DIR + '/' + WEIGHTS_SAVE.format(epoch=ep)
return None, None
model = get_training_model(weight_decay, gpus=use_multiple_gpus)
from_vgg = dict()
from_vgg['conv1_1'] = 'block1_conv1'
from_vgg['conv1_2'] = 'block1_conv2'
from_vgg['conv2_1'] = 'block2_conv1'
from_vgg['conv2_2'] = 'block2_conv2'
from_vgg['conv3_1'] = 'block3_conv1'
from_vgg['conv3_2'] = 'block3_conv2'
from_vgg['conv3_3'] = 'block3_conv3'
from_vgg['conv3_4'] = 'block3_conv4'
from_vgg['conv4_1'] = 'block4_conv1'
from_vgg['conv4_2'] = 'block4_conv2'
# load previous weights or vgg19 if this is the first run
last_epoch, wfile = get_last_epoch_and_weights_file()
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 1 + len(model.CLASSES)])
class_loss, presence_loss, loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
best = tf.argmax(y[:, 1:], 1)
correct = tf.argmax(y_[:, 1:], 1)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.global_variables_initializer()
def do_report():
r = sess.run([
best, correct,
tf.greater(y[:, 0], 0), y_[:, 0], class_loss, presence_loss, loss
],
feed_dict={
x: test_xs,
y_: test_ys
})
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print "{} {} <-> {} {}".format(model.CLASSES[c], pc,
model.CLASSES[b], float(pb))
print("B{:3d} loss: {} ").format(batch_idx, r[6])
def do_batch():
#writer = tf.summary.FileWriter("./log", sess.graph)
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("test/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 8 * len(common.CHARS) + 1])
digits_loss, presence_loss, loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 8, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 8, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.global_variables_initializer()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([
best, correct,
tf.greater(y[:, 0], 0), y_[:, 0], digits_loss, presence_loss, loss
],
feed_dict={
x: test_xs,
y_: test_ys
})
num_correct = numpy.sum(
numpy.logical_or(numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5, r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print "{} {} <-> {} {}".format(vec_to_plate(c), pc,
vec_to_plate(b), float(pb))
num_p_correct = numpy.sum(r[2] == r[3])
print(
"B{:3d} {:2.02f}% {:02.02f}% loss: {} "
"(digits: {}, presence: {}) |{}|").format(
batch_idx, 100. * num_correct / (len(r[0])),
100. * num_p_correct / len(r[2]), r[6], r[4], r[5],
"".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
for b, c, pb, pc in zip(*r_short)))
def do_batch():
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
#Revised by Juyi on 2017/10/24
# gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.333)
# with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC'
with tf.Session(config=config) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
#modify by joying original is 50
test_xs, test_ys = unzip(list(read_data("test8/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights8.npz", *last_weights)
return last_weights
def detect(initial_weights=None):
"""
Test the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param initial_weights:
(Optional.) Weights to initialize the network with.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
digits_loss, presence_loss, loss = get_loss(y, y_)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
#init = tf.initialize_all_variables() # deprecated
init = tf.global_variables_initializer()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([
best, correct,
tf.greater(y[:, 0], 0), y_[:, 0], digits_loss, presence_loss, loss
],
feed_dict={
x: test_xs,
y_: test_ys
})
num_correct = numpy.sum(
numpy.logical_or(numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5, r[3] < 0.5)))
num = numpy.equal(r[0], r[1])
num_correct_chars = numpy.sum(num)
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print("{} {} <-> {} {}".format(vec_to_plate(c), pc,
vec_to_plate(b), float(pb)))
print(
"Num plates: {} corrChars: {:2.02f}% corrLPs: {:2.02f}% loss: {} (digits: {}, presence: {})"
.format(len(r[0]), 100. * num_correct_chars / (len(r[0]) * 7),
100. * num_correct / (len(r[0])), r[6], r[4], r[5]))
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("pickering_test/*.png"))[:])
do_report()
stepsize = 136106 #68053 // after each stepsize iterations update learning rate: lr=lr*gamma
max_iter = 200000 # 600000
# True = start data generator client, False = use augmented dataset file (deprecated)
use_client_gen = True
WEIGHTS_BEST = "weights.best.h5"
TRAINING_LOG = "training.csv"
LOGS_DIR = "./logs"
def get_last_epoch():
data = pandas.read_csv(TRAINING_LOG)
return max(data['epoch'].values)
model = get_training_model(weight_decay)
from_vgg = dict()
from_vgg['conv1_1'] = 'block1_conv1'
from_vgg['conv1_2'] = 'block1_conv2'
from_vgg['conv2_1'] = 'block2_conv1'
from_vgg['conv2_2'] = 'block2_conv2'
from_vgg['conv3_1'] = 'block3_conv1'
from_vgg['conv3_2'] = 'block3_conv2'
from_vgg['conv3_3'] = 'block3_conv3'
from_vgg['conv3_4'] = 'block3_conv4'
from_vgg['conv4_1'] = 'block4_conv1'
from_vgg['conv4_2'] = 'block4_conv2'
# load previous weights or vgg19 if this is the first run
if os.path.exists(WEIGHTS_BEST):
def train(learn_rate, report_steps, batch_size, initial_weights=None):
'''
plt_epoch = []
plt_loss = []
'''
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
digits_loss, presence_loss, loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.global_variables_initializer()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([
best, correct,
tf.greater(y[:, 0], 0), y_[:, 0], digits_loss, presence_loss, loss
],
feed_dict={
x: test_xs,
y_: test_ys
})
num_correct = numpy.sum(
numpy.logical_or(numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5, r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
num_p_correct = numpy.sum(r[2] == r[3])
print("Batch:", batch_idx, "Digit loss:", r[4], "Presence_loss", r[5],
"Loss:", r[6])
'''
plt_epoch.append(batch_idx)
plt_loss.append(r[6])
'''
def do_batch():
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(
list(read_data("E:/project data/test/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
'''
plt.plot(plt_epoch, plt_loss)
plt.xlabel('epoch')
plt.ylabel('loss')
plt.title('epoch-loss graph')
fig = plt.gcf()
plt.show()
fig.savefig('loss_graph.pdf')
'''
last_weights = [p.eval() for p in params]
numpy.savez("weights0417ver_mini_train.npz", *last_weights)
return last_weights
def train(report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
global_step = tf.Variable(0, trainable=False)
batch_idx = tf.placeholder(tf.int32)
learning_rate = tf.train.exponential_decay(
common.INITIAL_LEARNING_RATE,
global_step,
common.DECAY_STEPS,
common.LEARNING_RATE_DECAY_FACTOR,
staircase=True)
# y is the predicted value
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, common.LENGTH * len(common.CHARS)])
digits_loss = tf.nn.softmax_cross_entropy_with_logits(
tf.reshape(y, [-1, len(common.CHARS)]),
tf.reshape(y_, [-1, len(common.CHARS)]))
cross_entropy = tf.reduce_sum(digits_loss)
train_step = tf.train.AdamOptimizer(learning_rate).minimize(
cross_entropy, global_step=global_step)
predict = tf.argmax(tf.reshape(
y, [-1, common.LENGTH, len(common.CHARS)]), 2)
real_value = tf.argmax(
tf.reshape(y_,
[-1, common.LENGTH, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run(
[predict, real_value, cross_entropy, learning_rate, global_step],
feed_dict={
x: test_xs,
y_: test_ys
})
num_correct = numpy.sum(numpy.all(r[0] == r[1], axis=1))
r_short = (r[0][:common.TEST_SIZE], r[1][:common.TEST_SIZE])
print "{} <--> {} ".format("real_value", "predict_value")
for pred, real in zip(*r_short):
print "{} <--> {} ".format(vec_to_plate(real), vec_to_plate(pred))
print(
"batch:{:3d}, hit_rate:{:2.02f}%,cross_entropy:{}, learning_rate:{},global_step:{} "
).format(batch_idx, 100. * num_correct / (len(r[0])), r[2], r[3], r[4])
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
def do_batch():
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(
list(read_data("test/*.png"))[:common.TEST_SIZE])
# print "test_xs.shape:{}".format(test_xs.shape)
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
# print "batch_ys.shape():{}".format(batch_ys.shape)
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}s".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model(
) # 获取训练网络(3个卷积层+2个全连接层)的输入,输出以及所有参数
y_ = tf.placeholder(
tf.float32,
[None, 1 + 7 * len(common.CHARS)]) # 给定网络输入对应的真输出[n, 1+7*36]
digits_loss, presence_loss, loss = get_loss(y, y_) # 获取损失函数
train_step = tf.train.AdamOptimizer(learn_rate).minimize(
loss) # 采用Adam优化器对损失函数进行优化
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]),
2) # 7位车牌号的每一位对应概率最大的字符的下标
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables() # 初始化所有变量
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run(
[
best, # 提供测试数据集,执行测试操作
correct,
tf.greater(y[:, 0], 0),
y_[:, 0],
digits_loss,
presence_loss,
loss
],
feed_dict={
x: test_xs,
y_: test_ys
})
num_correct = numpy.sum( # 车牌存在的情况下,车牌号识别正确的个数
numpy.logical_or(numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5, r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print "{} {} <-> {} {}".format(
vec_to_plate(c),
pc, # 输出：给定的7位车牌号 给定的车牌存在与否标志位
vec_to_plate(b),
float(pb)) # <-> 测试得到的7位车牌号 测试得到的车牌存在概率
num_p_correct = numpy.sum(r[2] == r[3]) # 车牌存在与否判断正确的个数
print(
"B{:3d} {:2.02f}% {:02.02f}% loss: {} " # 输出：B第n批次 车牌号识别准确率 车牌存在与否判断准确率 loss：总偏差
"(digits: {}, presence: {}) |{}|"
).format( # (digits：车牌号识别偏差, presence：车牌存在与否判断偏差)
batch_idx, # |batch_size个2类字符| 识别错误显示字符'X',识别正确显示字符' '
100. * num_correct / (len(r[0])),
100. * num_p_correct / len(r[2]),
r[6],
r[4],
r[5],
"".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
for b, c, pb, pc in zip(*r_short)))
def do_batch():
sess.run(
train_step, # 批量提供训练数据集,执行训练操作
feed_dict={
x: batch_xs,
y_: batch_ys
})
if batch_idx % report_steps == 0: # 训练1个回合进行1次测试，并显示相关指标
do_report() # 显示训练效果
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options)) as sess: # 配置GPU资源,建立会话
#with tf.device("/gpu:1"):
sess.run(init) # 执行变量初始化
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(
read_data("syndata/*.png"))[:50]) # 取前50个样本作为测试集
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(
read_batches(batch_size)) # 枚举出batch_size个训练数据集和对应下标
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch() # 批处理
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}".format( # 计算训练60批次(1批batch_size张图片)的用时
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt: # 出现键盘中断异常,则保存此刻的模型参数值到weights.npz文件
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def train(learn_rate, report_steps, batch_size, initial_weights=None):
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(constants.CHARS) + 1])
digits_loss = tf.nn.softmax_cross_entropy_with_logits(
tf.reshape(y[:, 1:], [-1, len(constants.CHARS)]),
tf.reshape(y_[:, 1:], [-1, len(constants.CHARS)]))
digits_loss = tf.reduce_sum(digits_loss)
presence_loss = 10. * tf.nn.sigmoid_cross_entropy_with_logits(
y[:, :1], y_[:, :1])
presence_loss = tf.reduce_sum(presence_loss)
cross_entropy = digits_loss + presence_loss
train_step = tf.train.AdamOptimizer(learn_rate).minimize(cross_entropy)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(constants.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(constants.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_number(v):
return "".join(constants.CHARS[i] for i in v)
def do_report():
r = sess.run([best, correct, tf.greater(y[:, 0], 0),
y_[:, 0], digits_loss, presence_loss, cross_entropy],
feed_dict={x: test_xs, y_: test_ys})
num_correct = numpy.sum(numpy.logical_or(
numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5, r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print("{} {} <-> {} {}".format(vec_to_number(c), pc,
vec_to_number(b), float(pb)))
num_p_correct = numpy.sum(r[2] == r[3])
print("B{:3d} {:2.02f}% {:02.02f}% loss: {} "
"(digits: {}, presence: {}) |{}|".format(
batch_idx, 100. * num_correct / (len(r[0])),
100. * num_p_correct / len(r[2]), r[6], r[4], r[5],
"".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
for b, c, pb, pc in zip(*r_short))))
def do_batch():
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
with tf.Session as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("test/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print("time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx)))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def train(starter_learning_rate, report_steps, batch_size, initial_weights):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32,
[None, 10 * len(CHARS) + 1]) #7 changed to 10
#y_ = tf.placeholder(tf.float32, [None, 729 ]) #7 changed to 10
#y_ = tf.placeholder(tf.int32, [None, 10 * len(CHARS) + 1]) #7 changed to 10
#y_ = tf.one_hot( y_ , 621 )
#y_ = tf.squeeze( y_ , 0)
digits_loss, presence_loss, loss = get_loss(y, y_)
#train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
global_step = tf.Variable(0, trainable=False)
#starter_learning_rate = 0.1
learn_rate = tf.train.exponential_decay(starter_learning_rate,
global_step,
1000,
0.96,
staircase=True)
train_step = (tf.train.AdamOptimizer(learn_rate).minimize(
loss, global_step=global_step))
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 10, len(CHARS)]),
2) # 7 changed to 10
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 10, len(CHARS)]),
2) #7 changed to 10
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_plate(v):
return "".join(CHARS[i] for i in v)
def do_report():
r = sess.run([
best, correct,
tf.greater(y[:, 0], 0), y_[:, 0], digits_loss, presence_loss, loss
],
feed_dict={
x: test_xs,
y_: test_ys
})
num_correct = numpy.sum(
numpy.logical_or(numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5, r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
#print "{} {} <-> {} {}".format(vec_to_plate(c), pc,
#vec_to_plate(b), float(pb)) # print command needs to be changed according to python3
print("{} {} <-> {} {}".format(vec_to_plate(c), pc,
vec_to_plate(b), float(pb)))
num_p_correct = numpy.sum(r[2] == r[3])
print("B{:3d} {:2.02f}% {:02.02f}% loss: {} "
"(digits: {}, presence: {}) |{}|".format(
batch_idx, 100. * num_correct / (len(r[0])),
100. * num_p_correct / len(r[2]), r[6], r[4], r[5],
"".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
for b, c, pb, pc in zip(*r_short))))
def do_batch():
#batch_xs = numpy.reshape(batch_xs, (None, 621))
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("test/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print("time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx)))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def train(learn_rate, report_steps, batch_size, initial_weights=None):
x,y,params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, len(common.OUTCOMES)])
loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
best = tf.argmax(tf.reshape(y[:,:], [-1, 1, len(common.OUTCOMES)]), 2)
correct = tf.argmax(tf.reshape(y_[:,1:], [-1, 1, len(common.OUTCOMES)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_result(v):
return "".join(common.OUTCOMES[i] for i in v)
def do_report():
#r = sess.run([best,
# correct,
# tf.greater(y[:, 0], 0),
# y_,
# loss],
# feed_dict = {x: test_xs, y_: test_ys})
r = sess.run(y, feed_dict = {x: test_xs, y_: test_ys})
print('Outcome for testing images: {}'.format(r))
#num_correct = numpy.sum(r[0] == r[1], axis=0)
#r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
#for b, c, pb, pc in zip(*r_short):
# print("{} {} <-> {} {}".format(vec_to_result(c), pc,
# vec_to_result(b), float(pb)))
#num_p_correct = numpy.sum(r[2] == r[3])
#print ('B{:3d} {:2.02f}% {:02.02f}% loss: {} (digits: {}, presence: {}) |{}|').format(
# batch_idx,
# 100. * num_correct / (len(r[0])),
# 100. * num_p_correct / len(r[2]),
# r[6],
# r[4],
# r[5],
# "".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
# for b, c, pb, pc in zip(*r_short)))
def do_batch():
# print(batch_xs.shape)
# print(batch_ys.shape)
sess.run(train_step,
feed_dict={x: batch_xs, y_:batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("Nodules/Positive/*.jpg", "Nodules/Negative/*.jpg"))[:common.TEST_NUM]) #First some images for testing
try:
while True:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_training_batch(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print("time for 60 batches {}".format(
60*(last_batch_time - batch_time) /
(last_batch_idx - batch_idx)))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def prepare(config,
config_name,
exp_id,
train_samples,
val_samples,
batch_size,
epoch=None):
metrics_id = config_name + "_" + exp_id if exp_id is not None else config_name
weights_id = config_name + "/" + exp_id if exp_id is not None else config_name
WEIGHT_DIR = "./" + weights_id
WEIGHTS_SAVE = 'weights.{epoch:04d}.h5'
TRAINING_LOG = "./" + metrics_id + ".csv"
LOGS_DIR = "./logs"
model = get_training_model(weight_decay,
np_branch1=config.paf_layers,
np_branch2=config.heat_layers + 1)
lr_mult = get_lrmult(model)
# load previous weights or vgg19 if this is the first run
last_epoch, wfile = get_last_epoch_and_weights_file(
WEIGHT_DIR, WEIGHTS_SAVE, epoch)
print("last_epoch:", last_epoch)
if wfile is not None:
print("Loading %s ..." % wfile)
model.load_weights(wfile)
else:
print("Loading vgg19 weights...")
vgg_model = VGG19(include_top=False, weights='imagenet')
from_vgg = dict()
from_vgg['conv1_1'] = 'block1_conv1'
from_vgg['conv1_2'] = 'block1_conv2'
from_vgg['conv2_1'] = 'block2_conv1'
from_vgg['conv2_2'] = 'block2_conv2'
from_vgg['conv3_1'] = 'block3_conv1'
from_vgg['conv3_2'] = 'block3_conv2'
from_vgg['conv3_3'] = 'block3_conv3'
from_vgg['conv3_4'] = 'block3_conv4'
from_vgg['conv4_1'] = 'block4_conv1'
from_vgg['conv4_2'] = 'block4_conv2'
for layer in model.layers:
if layer.name in from_vgg:
vgg_layer_name = from_vgg[layer.name]
layer.set_weights(
vgg_model.get_layer(vgg_layer_name).get_weights())
print("Loaded VGG19 layer: " + vgg_layer_name)
last_epoch = 0
# euclidean loss as implemented in caffe https://github.com/BVLC/caffe/blob/master/src/caffe/layers/euclidean_loss_layer.cpp
def eucl_loss(x, y):
l = K.sum(K.square(x - y)) / batch_size / 2
return l
# learning rate schedule - equivalent of caffe lr_policy = "step"
iterations_per_epoch = train_samples // batch_size
def step_decay(epoch):
steps = epoch * iterations_per_epoch * batch_size
lrate = base_lr * math.pow(gamma, math.floor(steps / stepsize))
print("Epoch:", epoch, "Learning rate:", lrate)
return lrate
print("Weight decay policy...")
for i in range(1, 100, 5):
step_decay(i)
# configure callbacks
lrate = LearningRateScheduler(step_decay)
checkpoint = ModelCheckpoint(WEIGHT_DIR + '/' + WEIGHTS_SAVE,
monitor='loss',
verbose=0,
save_best_only=False,
save_weights_only=True,
mode='min',
period=1)
csv_logger = CSVLogger(TRAINING_LOG, append=True)
tb = TensorBoard(log_dir=LOGS_DIR,
histogram_freq=0,
write_graph=True,
write_images=False)
tnan = TerminateOnNaN()
#coco_eval = CocoEval(train_client, val_client)
callbacks_list = [lrate, checkpoint, csv_logger, tb, tnan]
# sgd optimizer with lr multipliers
multisgd = MultiSGD(lr=base_lr,
momentum=momentum,
decay=0.0,
nesterov=False,
lr_mult=lr_mult)
# start training
model.compile(loss=eucl_loss, optimizer=multisgd)
return model, iterations_per_epoch, val_samples // batch_size, last_epoch, metrics_id, callbacks_list
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
Train the network.
The function operates interactively: Progress is reported on stdout, and
training ceases upon `KeyboardInterrupt` at which point the learned weights
are saved to `weights.npz`, and also returned.
:param learn_rate:
Learning rate to use.
:param report_steps:
Every `report_steps` batches a progress report is printed.
:param batch_size:
The size of the batches used for training.
:param initial_weights:
(Optional.) Weights to initialize the network with.
:return:
The learned network weights.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
digits_loss = tf.nn.softmax_cross_entropy_with_logits(
tf.reshape(y[:, 1:],
[-1, len(common.CHARS)]),
tf.reshape(y_[:, 1:],
[-1, len(common.CHARS)]))
digits_loss = tf.reduce_sum(digits_loss)
presence_loss = 10. * tf.nn.sigmoid_cross_entropy_with_logits(
y[:, :1], y_[:, :1])
presence_loss = tf.reduce_sum(presence_loss)
cross_entropy = digits_loss + presence_loss
train_step = tf.train.AdamOptimizer(learn_rate).minimize(cross_entropy)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([best,
correct,
tf.greater(y[:, 0], 0),
y_[:, 0],
digits_loss,
presence_loss,
cross_entropy],
feed_dict={x: test_xs, y_: test_ys})
num_correct = numpy.sum(
numpy.logical_or(
numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5,
r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print "{} {} <-> {} {}".format(vec_to_plate(c), pc,
vec_to_plate(b), float(pb))
num_p_correct = numpy.sum(r[2] == r[3])
print ("B{:3d} {:2.02f}% {:02.02f}% loss: {} "
"(digits: {}, presence: {}) |{}|").format(
batch_idx,
100. * num_correct / (len(r[0])),
100. * num_p_correct / len(r[2]),
r[6],
r[4],
r[5],
"".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
for b, c, pb, pc in zip(*r_short)))
def do_batch():
sess.run(train_step,
feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("test/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights
def train(learn_rate, report_steps, batch_size, initial_weights=None):
"""
1 interactively
2 `KeyboardInterrupt`
3 learned weights are saved to `weights.npz`, and also returned.
"""
x, y, params = model.get_training_model()
y_ = tf.placeholder(tf.float32, [None, 7 * len(common.CHARS) + 1])
digits_loss, presence_loss, loss = get_loss(y, y_)
train_step = tf.train.AdamOptimizer(learn_rate).minimize(loss)
best = tf.argmax(tf.reshape(y[:, 1:], [-1, 7, len(common.CHARS)]), 2)
correct = tf.argmax(tf.reshape(y_[:, 1:], [-1, 7, len(common.CHARS)]), 2)
if initial_weights is not None:
assert len(params) == len(initial_weights)
assign_ops = [w.assign(v) for w, v in zip(params, initial_weights)]
init = tf.initialize_all_variables()
def vec_to_plate(v):
return "".join(common.CHARS[i] for i in v)
def do_report():
r = sess.run([best,
correct,
tf.greater(y[:, 0], 0),
y_[:, 0],
digits_loss,
presence_loss,
loss],
feed_dict={x: test_xs, y_: test_ys})
num_correct = numpy.sum(
numpy.logical_or(
numpy.all(r[0] == r[1], axis=1),
numpy.logical_and(r[2] < 0.5,
r[3] < 0.5)))
r_short = (r[0][:190], r[1][:190], r[2][:190], r[3][:190])
for b, c, pb, pc in zip(*r_short):
print "{} {} <-> {} {}".format(vec_to_plate(c), pc,
vec_to_plate(b), float(pb))
num_p_correct = numpy.sum(r[2] == r[3])
print ("B{:3d} {:2.02f}% {:02.02f}% loss: {} "
"(digits: {}, presence: {}) |{}|").format(
batch_idx,
100. * num_correct / (len(r[0])),
100. * num_p_correct / len(r[2]),
r[6],
r[4],
r[5],
"".join("X "[numpy.array_equal(b, c) or (not pb and not pc)]
for b, c, pb, pc in zip(*r_short)))
def do_batch():
sess.run(train_step,
feed_dict={x: batch_xs, y_: batch_ys})
if batch_idx % report_steps == 0:
do_report()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init)
if initial_weights is not None:
sess.run(assign_ops)
test_xs, test_ys = unzip(list(read_data("test/*.png"))[:50])
try:
last_batch_idx = 0
last_batch_time = time.time()
batch_iter = enumerate(read_batches(batch_size))
for batch_idx, (batch_xs, batch_ys) in batch_iter:
do_batch()
if batch_idx % report_steps == 0:
batch_time = time.time()
if last_batch_idx != batch_idx:
print "time for 60 batches {}".format(
60 * (last_batch_time - batch_time) /
(last_batch_idx - batch_idx))
last_batch_idx = batch_idx
last_batch_time = batch_time
except KeyboardInterrupt:
last_weights = [p.eval() for p in params]
numpy.savez("weights.npz", *last_weights)
return last_weights