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, common.LENGTH * 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, common.LENGTH, len(common.CHARS)]), 2)
correct = tf.argmax(
tf.reshape(y_[:, 1:],
[-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([
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
# Builds the graph
input_placeholder = tf.placeholder(tf.float32, [
embedding_batch_size, sequence_length, height, width, channels
])
embeddings = create_id3_embedding(
preprocess(input_placeholder, (224, 224)))
reference_embeddings_placeholder = tf.placeholder(
tf.float32, [sequences, 400])
generated_embeddings_placeholder = tf.placeholder(
tf.float32, [sequences, 400])
fvd = calculate_fvd(reference_embeddings_placeholder,
generated_embeddings_placeholder)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.3)
with tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options)) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
# Computes the embeddings
reference_embeddings = extract_embeddings(
embeddings,
input_placeholder,
reference,
sess,
batch_size=embedding_batch_size)
generated_embeddings = extract_embeddings(
embeddings,
input_placeholder,
def __call__(self, reference_dataloader, generated_dataloader) -> float:
'''
Computes the FVD between the reference and the generated observations
:param reference_dataloader: dataloader for reference observations
:param generated_dataloader: dataloader for generated observations
:return: The FVD between the two distributions
'''
# Multiples of 16 must be fes to the embeddings network
embedding_batch_size = 16
# Extracts information about the batch
sample_observations = next(iter(reference_dataloader)).to_tuple()[0]
_, sequence_length, channels, height, width = list(
sample_observations.size())
with tf.Graph().as_default():
# Builds the embedding computation part of the graph
input_placeholder = tf.placeholder(tf.float32, [
embedding_batch_size, sequence_length, height, width, channels
])
embeddings = create_id3_embedding(
preprocess(input_placeholder, (224, 224)))
# Computes all embeddings
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.3)
with tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options)) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
reference_embeddings = self.extract_all_embeddings(
reference_dataloader, embeddings, input_placeholder, sess,
embedding_batch_size)
generated_embeddings = self.extract_all_embeddings(
generated_dataloader, embeddings, input_placeholder, sess,
embedding_batch_size)
if reference_embeddings.shape[0] != generated_embeddings.shape[
0]:
raise Exception(
f"Size of reference ({reference_embeddings.shape[0]}) and generated embeddings ({generated_embeddings.shape[0]}) differ"
)
sequences = reference_embeddings.shape[0]
# Builds the fvd computation part of the graph
reference_embeddings_placeholder = tf.placeholder(
tf.float32, [sequences, 400])
generated_embeddings_placeholder = tf.placeholder(
tf.float32, [sequences, 400])
fvd = calculate_fvd(reference_embeddings_placeholder,
generated_embeddings_placeholder)
# Computes the fvd
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.3)
with tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options)) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
# Computes the fvd
fvd_np = sess.run(fvd,
feed_dict={
reference_embeddings_placeholder:
reference_embeddings,
generated_embeddings_placeholder:
generated_embeddings
})
return float(fvd_np)
def transfrom_data(NUM_TEST_PER_EPOCH):
encoded_data = []
original_data = []
train_loader = get_data(BATCH_SIZE)
trainiter = iter(train_loader)
for i in range(NUM_ITERATION):
features2, labels = next(trainiter)
features2 = np.array(
tf.keras.applications.resnet.preprocess_input(
features2.data.numpy().transpose(0, 2, 3, 1) * 255))
features = np.zeros(
[len(features2), img_size[0], img_size[1], img_size[2]])
features[:, :, :, 0] = (features2[:, :, :, 2])
features = np.clip(features + 123.68, 0, 255) / 255
original_images = np.copy(features)
for k in range(len(features)):
features[k] = add_gaussian_noise(features[k],
sd=np.random.uniform(
NOISE_STD_RANGE[0],
NOISE_STD_RANGE[1]))
encoded_data.append(np.copy(features))
original_data.append(np.copy(original_images))
gpu_options = tf.GPUOptions(allow_growth=True,
visible_device_list=str(GPU_ID))
config = tf.ConfigProto(gpu_options=gpu_options)
gx1 = tf.Graph()
with gx1.as_default():
with tf.Session(config=config) as sess1:
sargan_model1 = SARGAN(img_size, BATCH_SIZE, img_channel=1)
sargan_saver1 = tf.train.Saver()
sargan_saver1 = tf.train.import_meta_graph(trained_model_path1 +
'/sargan_mnist.meta')
sargan_saver1.restore(
sess1, tf.train.latest_checkpoint(trained_model_path1))
for ibatch in range(NUM_TEST_PER_EPOCH):
processed_batch = sess1.run(sargan_model1.gen_img,
feed_dict={
sargan_model1.image:
encoded_data[ibatch],
sargan_model1.cond:
encoded_data[ibatch]
})
encoded_data[ibatch] = (np.copy(processed_batch))
gx2 = tf.Graph()
with gx2.as_default():
with tf.Session(config=config) as sess2:
sargan_model2 = SARGAN(img_size, BATCH_SIZE, img_channel=1)
sargan_saver2 = tf.train.Saver()
sargan_saver2 = tf.train.import_meta_graph(trained_model_path2 +
'/sargan_mnist.meta')
sargan_saver2.restore(
sess2, tf.train.latest_checkpoint(trained_model_path2))
for ibatch in range(NUM_TEST_PER_EPOCH):
processed_batch = sess2.run(sargan_model2.gen_img,
feed_dict={
sargan_model2.image:
encoded_data[ibatch],
sargan_model2.cond:
encoded_data[ibatch]
})
encoded_data[ibatch] = (np.copy(processed_batch))
gx3 = tf.Graph()
with gx3.as_default():
with tf.Session(config=config) as sess3:
sargan_model3 = SARGAN(img_size, BATCH_SIZE, img_channel=1)
sargan_saver3 = tf.train.Saver()
sargan_saver3 = tf.train.import_meta_graph(trained_model_path3 +
'/sargan_mnist.meta')
sargan_saver3.restore(
sess3, tf.train.latest_checkpoint(trained_model_path3))
for ibatch in range(NUM_TEST_PER_EPOCH):
processed_batch = sess3.run(sargan_model3.gen_img,
feed_dict={
sargan_model3.image:
encoded_data[ibatch],
sargan_model3.cond:
encoded_data[ibatch]
})
encoded_data[ibatch] = (np.copy(processed_batch))
return encoded_data, original_data
def Recognize(idList):
with tf.Graph().as_default():
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.6)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False))
with sess.as_default():
pnet, rnet, onet = detect_face.create_mtcnn(sess, npy)
minsize = 30 # minimum size of face
threshold = [0.7, 0.8, 0.8] # three steps's threshold
factor = 0.709 # scale factor
margin = 44
batch_size = 100 #1000
image_size = 182
input_image_size = 160
HumanNames = os.listdir(train_img)
HumanNames.sort()
print('Loading Model')
facenet.load_model(modeldir)
images_placeholder = tf.get_default_graph().get_tensor_by_name(
"input:0")
embeddings = tf.get_default_graph().get_tensor_by_name(
"embeddings:0")
phase_train_placeholder = tf.get_default_graph(
).get_tensor_by_name("phase_train:0")
embedding_size = embeddings.get_shape()[1]
classifier_filename_exp = os.path.expanduser(classifier_filename)
with open(classifier_filename_exp, 'rb') as infile:
(model, class_names) = pickle.load(infile, encoding='latin1')
video_capture = cv2.VideoCapture(video)
print('Start Recognition')
while True:
ret, frame = video_capture.read()
#frame = cv2.resize(frame, (0,0), fx=0.5, fy=0.5) #resize frame (optional)
timer = time.time()
if frame.ndim == 2:
frame = facenet.to_rgb(frame)
bounding_boxes, _ = detect_face.detect_face(
frame, minsize, pnet, rnet, onet, threshold, factor)
faceNum = bounding_boxes.shape[0]
if faceNum > 0:
det = bounding_boxes[:, 0:4]
img_size = np.asarray(frame.shape)[0:2]
cropped = []
scaled = []
scaled_reshape = []
for i in range(faceNum):
emb_array = np.zeros((1, embedding_size))
xmin = int(det[i][0])
ymin = int(det[i][1])
xmax = int(det[i][2])
ymax = int(det[i][3])
try:
# inner exception
if xmin <= 0 or ymin <= 0 or xmax >= len(
frame[0]) or ymax >= len(frame):
print('Face is very close!')
continue
cropped.append(frame[ymin:ymax, xmin:xmax, :])
cropped[i] = facenet.flip(cropped[i], False)
scaled.append(
np.array(
Image.fromarray(cropped[i]).resize(
(image_size, image_size))))
scaled[i] = cv2.resize(
scaled[i],
(input_image_size, input_image_size),
interpolation=cv2.INTER_CUBIC)
scaled[i] = facenet.prewhiten(scaled[i])
scaled_reshape.append(scaled[i].reshape(
-1, input_image_size, input_image_size, 3))
feed_dict = {
images_placeholder: scaled_reshape[i],
phase_train_placeholder: False
}
emb_array[0, :] = sess.run(embeddings,
feed_dict=feed_dict)
predictions = model.predict_proba(emb_array)
best_class_indices = np.argmax(predictions, axis=1)
best_class_probabilities = predictions[
np.arange(len(best_class_indices)),
best_class_indices]
if best_class_probabilities > 0.87:
cv2.rectangle(frame, (xmin, ymin),
(xmax, ymax), (0, 255, 0),
2) #boxing face
for H_i in HumanNames:
if HumanNames[
best_class_indices[0]] == H_i:
result_ids = HumanNames[
best_class_indices[0]]
result_names = "?"
profile = GetProfile(result_ids)
if (profile != None):
result_names = profile[1]
if int(result_ids) not in idList:
idList.append(int(result_ids))
print(
"Predictions : [ name: {} , accuracy: {:.3f} ]"
.format(
HumanNames[
best_class_indices[0]],
best_class_probabilities[0]))
cv2.rectangle(frame, (xmin, ymin - 20),
(xmax, ymin - 2),
(0, 255, 255), -1)
cv2.putText(
frame,
result_names, (xmin, ymin - 5),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
1, (0, 0, 0),
thickness=1,
lineType=1)
else:
cv2.rectangle(frame, (xmin, ymin),
(xmax, ymax), (0, 255, 0), 2)
cv2.rectangle(frame, (xmin, ymin - 20),
(xmax, ymin - 2), (0, 255, 255),
-1)
cv2.putText(frame,
"?", (xmin, ymin - 5),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
1, (0, 0, 0),
thickness=1,
lineType=1)
except:
print("error")
endtimer = time.time()
fps = 1 / (endtimer - timer)
cv2.rectangle(frame, (15, 30), (135, 60), (0, 255, 255), -1)
cv2.putText(frame, "fps: {:.2f}".format(fps), (20, 50),
cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 0, 0), 2)
cv2.imshow('Face Recognition', frame)
key = cv2.waitKey(1)
if key == 113: # "q"
break
video_capture.release()
cv2.destroyAllWindows()
def __init__(self, net_factory, model_path):
#create a graph
graph = tf.Graph()
with graph.as_default():
#define tensor and op in graph(-1,1)
self.image_op = tf.placeholder(tf.float32, name='input_image')
self.width_op = tf.placeholder(tf.int32, name='image_width')
self.height_op = tf.placeholder(tf.int32, name='image_height')
image_reshape = tf.reshape(self.image_op, [1, self.height_op, self.width_op, 3])
#self.cls_prob batch*2
#self.bbox_pred batch*4
#construct model here
#self.cls_prob, self.bbox_pred = net_factory(image_reshape, training=False)
#contains landmark
self.cls_prob, self.bbox_pred, _ = net_factory(image_reshape, training=False)
#allow
self.sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, gpu_options=tf.GPUOptions(allow_growth=True)))
saver = tf.train.Saver()
#check whether the dictionary is valid
model_dict = '/'.join(model_path.split('/')[:-1])
ckpt = tf.train.get_checkpoint_state(model_dict)
print(model_path)
readstate = ckpt and ckpt.model_checkpoint_path
assert readstate, "the params dictionary is not valid"
print("restore models' param")
saver.restore(self.sess, model_path)
def initialize_globals():
c = AttrDict()
# Set default dropout rates
if FLAGS.dropout_rate2 < 0:
FLAGS.dropout_rate2 = FLAGS.dropout_rate
if FLAGS.dropout_rate3 < 0:
FLAGS.dropout_rate3 = FLAGS.dropout_rate
if FLAGS.dropout_rate6 < 0:
FLAGS.dropout_rate6 = FLAGS.dropout_rate
# Set default checkpoint dir
if not FLAGS.checkpoint_dir:
FLAGS.checkpoint_dir = xdg.save_data_path(
os.path.join('deepspeech', 'checkpoints'))
if FLAGS.load not in ['last', 'best', 'init', 'auto']:
FLAGS.load = 'auto'
# Set default summary dir
if not FLAGS.summary_dir:
FLAGS.summary_dir = xdg.save_data_path(
os.path.join('deepspeech', 'summaries'))
# Standard session configuration that'll be used for all new sessions.
c.session_config = tfv1.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_placement,
inter_op_parallelism_threads=FLAGS.inter_op_parallelism_threads,
intra_op_parallelism_threads=FLAGS.intra_op_parallelism_threads,
gpu_options=tfv1.GPUOptions(allow_growth=FLAGS.use_allow_growth))
# CPU device
c.cpu_device = '/cpu:0'
# Available GPU devices
c.available_devices = get_available_gpus(c.session_config)
print("*******device**********", c.available_devices)
# If there is no GPU available, we fall back to CPU based operation
if not c.available_devices:
c.available_devices = [c.cpu_device]
if FLAGS.utf8:
c.alphabet = UTF8Alphabet()
else:
c.alphabet = Alphabet(os.path.abspath(FLAGS.alphabet_config_path))
# Geometric Constants
# ===================
# For an explanation of the meaning of the geometric constants, please refer to
# doc/Geometry.md
# Number of MFCC features
c.n_input = 13 # TODO: Determine this programmatically from the sample rate
# The number of frames in the context
c.n_context = 10 # TODO: Determine the optimal value using a validation data set
# Number of units in hidden layers
c.n_hidden = FLAGS.n_hidden
c.n_hidden_1 = c.n_hidden
c.n_hidden_2 = c.n_hidden
c.n_hidden_5 = c.n_hidden
# LSTM cell state dimension
c.n_cell_dim = c.n_hidden
# The number of units in the third layer, which feeds in to the LSTM
c.n_hidden_3 = c.n_cell_dim
# Units in the sixth layer = number of characters in the target language plus one
c.n_hidden_6 = c.alphabet.size() + 1 # +1 for CTC blank label
# Size of audio window in samples
if (FLAGS.feature_win_len * FLAGS.audio_sample_rate) % 1000 != 0:
log_error(
'--feature_win_len value ({}) in milliseconds ({}) multiplied '
'by --audio_sample_rate value ({}) must be an integer value. Adjust '
'your --feature_win_len value or resample your audio accordingly.'
''.format(FLAGS.feature_win_len, FLAGS.feature_win_len / 1000,
FLAGS.audio_sample_rate))
sys.exit(1)
c.audio_window_samples = FLAGS.audio_sample_rate * (FLAGS.feature_win_len /
1000)
# Stride for feature computations in samples
if (FLAGS.feature_win_step * FLAGS.audio_sample_rate) % 1000 != 0:
log_error(
'--feature_win_step value ({}) in milliseconds ({}) multiplied '
'by --audio_sample_rate value ({}) must be an integer value. Adjust '
'your --feature_win_step value or resample your audio accordingly.'
''.format(FLAGS.feature_win_step, FLAGS.feature_win_step / 1000,
FLAGS.audio_sample_rate))
sys.exit(1)
c.audio_step_samples = FLAGS.audio_sample_rate * (FLAGS.feature_win_step /
1000)
if FLAGS.one_shot_infer:
if not os.path.exists(FLAGS.one_shot_infer):
log_error(
'Path specified in --one_shot_infer is not a valid file.')
sys.exit(1)
ConfigSingleton._config = c # pylint: disable=protected-access
def collect_data(self):
output_dir = os.path.expanduser(self.output_datadir)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
dataset = facenet.get_dataset(self.input_datadir)
with tf.Graph().as_default():
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
sess = tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options, log_device_placement=False))
with sess.as_default():
pnet, rnet, onet = detect_face.create_mtcnn(sess, './npy')
minsize = 20 # minimum size of face
threshold = [0.6, 0.7, 0.7] # three steps's threshold
factor = 0.709 # scale factor
margin = 44
image_size = 182
# Add a random key to the filename to allow alignment using multiple processes
random_key = np.random.randint(0, high=99999)
bounding_boxes_filename = os.path.join(
output_dir, 'bounding_boxes_%05d.txt' % random_key)
with open(bounding_boxes_filename, "w") as text_file:
nrof_images_total = 0
nrof_successfully_aligned = 0
for cls in dataset:
output_class_dir = os.path.join(output_dir, cls.name)
if not os.path.exists(output_class_dir):
os.makedirs(output_class_dir)
for image_path in cls.image_paths:
nrof_images_total += 1
filename = os.path.splitext(
os.path.split(image_path)[1])[0]
output_filename = os.path.join(output_class_dir,
filename + '.png')
print("Image: %s" % image_path)
if not os.path.exists(output_filename):
try:
img = misc.imread(image_path)
except (IOError, ValueError, IndexError) as e:
errorMessage = '{}: {}'.format(image_path, e)
print(errorMessage)
else:
if img.ndim < 2:
print('Unable to align "%s"' % image_path)
text_file.write('%s\n' % (output_filename))
continue
if img.ndim == 2:
img = facenet.to_rgb(img)
print('to_rgb data dimension: ', img.ndim)
img = img[:, :, 0:3]
bounding_boxes, _ = detect_face.detect_face(
img, minsize, pnet, rnet, onet, threshold,
factor)
nrof_faces = bounding_boxes.shape[0]
print('No of Detected Face: %d' % nrof_faces)
if nrof_faces > 0:
det = bounding_boxes[:, 0:4]
img_size = np.asarray(img.shape)[0:2]
if nrof_faces > 1:
bounding_box_size = (
det[:, 2] - det[:, 0]) * (det[:, 3] -
det[:, 1])
img_center = img_size / 2
offsets = np.vstack([
(det[:, 0] + det[:, 2]) / 2 -
img_center[1],
(det[:, 1] + det[:, 3]) / 2 -
img_center[0]
])
offset_dist_squared = np.sum(
np.power(offsets, 2.0), 0)
index = np.argmax(
bounding_box_size -
offset_dist_squared * 2.0
) # some extra weight on the centering
det = det[index, :]
det = np.squeeze(det)
bb_temp = np.zeros(4, dtype=np.int32)
bb_temp[0] = det[0]
bb_temp[1] = det[1]
bb_temp[2] = det[2]
bb_temp[3] = det[3]
cropped_temp = img[bb_temp[1]:bb_temp[3],
bb_temp[0]:bb_temp[2], :]
scaled_temp = misc.imresize(
cropped_temp, (image_size, image_size),
interp='bilinear')
nrof_successfully_aligned += 1
misc.imsave(output_filename, scaled_temp)
text_file.write(
'%s %d %d %d %d\n' %
(output_filename, bb_temp[0], bb_temp[1],
bb_temp[2], bb_temp[3]))
else:
print('Unable to align "%s"' % image_path)
text_file.write('%s\n' % (output_filename))
return (nrof_images_total, nrof_successfully_aligned)
def export(self,
output_dir: Text,
tflite_path: Text = None,
tensorrt: Text = None):
"""Export a saved model, frozen graph, and potential tflite/tensorrt model.
Args:
output_dir: the output folder for saved model.
tflite_path: the path for saved tflite file.
tensorrt: If not None, must be {'FP32', 'FP16', 'INT8'}.
"""
signitures = self.signitures
signature_def_map = {
'serving_default':
tf.saved_model.predict_signature_def(
{signitures['image_arrays'].name: signitures['image_arrays']},
{signitures['prediction'].name: signitures['prediction']}),
}
b = tf.saved_model.Builder(output_dir)
b.add_meta_graph_and_variables(self.sess,
tags=['serve'],
signature_def_map=signature_def_map,
assets_collection=tf.get_collection(
tf.GraphKeys.ASSET_FILEPATHS),
clear_devices=True)
b.save()
logging.info('Model saved at %s', output_dir)
# also save freeze pb file.
graphdef = self.freeze()
pb_path = os.path.join(output_dir, self.model_name + '_frozen.pb')
tf.io.gfile.GFile(pb_path, 'wb').write(graphdef.SerializeToString())
logging.info('Frozen graph saved at %s', pb_path)
if tflite_path:
height, width = utils.parse_image_size(self.params['image_size'])
input_name = signitures['image_arrays'].op.name
input_shapes = {input_name: [None, height, width, 3]}
converter = tf.lite.TFLiteConverter.from_saved_model(
output_dir,
input_arrays=[input_name],
input_shapes=input_shapes,
output_arrays=[signitures['prediction'].op.name])
converter.target_spec.supported_ops = [
tf.lite.OpsSet.TFLITE_BUILTINS
]
tflite_model = converter.convert()
tf.io.gfile.GFile(tflite_path, 'wb').write(tflite_model)
logging.info('TFLite is saved at %s', tflite_path)
if tensorrt:
from tensorflow.python.compiler.tensorrt import trt # pylint: disable=g-direct-tensorflow-import,g-import-not-at-top
sess_config = tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True))
trt_path = os.path.join(output_dir, 'tensorrt_' + tensorrt.lower())
trt.create_inference_graph(None,
None,
precision_mode=tensorrt,
input_saved_model_dir=output_dir,
output_saved_model_dir=trt_path,
session_config=sess_config)
logging.info('TensorRT model is saved at %s', trt_path)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--path',
help='Path of the video you want to test on.',
default=0)
args = parser.parse_args()
MINSIZE = 20
THRESHOLD = [0.6, 0.7, 0.7]
FACTOR = 0.709
IMAGE_SIZE = 182
INPUT_IMAGE_SIZE = 160
CLASSIFIER_PATH = 'Models/facemodel_rework.pkl'
VIDEO_PATH = args.path
FACENET_MODEL_PATH = 'Models/20180402-114759.pb'
# Load The Custom Classifier
with open(CLASSIFIER_PATH, 'rb') as file:
model, class_names = pickle.load(file)
print("Custom Classifier, Successfully loaded")
with tf.Graph().as_default():
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.6)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options,
log_device_placement=False))
with sess.as_default():
# Load the model
print('Loading feature extraction model')
facenet.load_model(FACENET_MODEL_PATH)
# Get input and output tensors
images_placeholder = tf.get_default_graph().get_tensor_by_name(
"input:0")
embeddings = tf.get_default_graph().get_tensor_by_name(
"embeddings:0")
phase_train_placeholder = tf.get_default_graph(
).get_tensor_by_name("phase_train:0")
embedding_size = embeddings.get_shape()[1]
pnet, rnet, onet = align.detect_face.create_mtcnn(sess, "align/")
people_detected = set()
person_detected = collections.Counter()
cap = VideoStream(src=0).start()
while (True):
frame = cap.read()
frame = imutils.resize(frame, width=800)
frame = cv2.flip(frame, 1)
bounding_boxes, _ = align.detect_face.detect_face(
frame, MINSIZE, pnet, rnet, onet, THRESHOLD, FACTOR)
faces_found = bounding_boxes.shape[0]
try:
# if faces_found > 1:
# cv2.putText(frame, "Only one face", (0, 100), cv2.FONT_HERSHEY_COMPLEX_SMALL,
# 1, (255, 255, 255), thickness=1, lineType=2)
if faces_found > 0:
det = bounding_boxes[:, 0:4]
bb = np.zeros((faces_found, 4), dtype=np.int32)
for i in range(faces_found):
bb[i][0] = det[i][0]
bb[i][1] = det[i][1]
bb[i][2] = det[i][2]
bb[i][3] = det[i][3]
print(bb[i][3] - bb[i][1])
print(frame.shape[0])
print((bb[i][3] - bb[i][1]) / frame.shape[0])
if (bb[i][3] - bb[i][1]) / frame.shape[0] > 0.25:
cropped = frame[bb[i][1]:bb[i][3],
bb[i][0]:bb[i][2], :]
scaled = cv2.resize(
cropped,
(INPUT_IMAGE_SIZE, INPUT_IMAGE_SIZE),
interpolation=cv2.INTER_CUBIC)
scaled = facenet.prewhiten(scaled)
scaled_reshape = scaled.reshape(
-1, INPUT_IMAGE_SIZE, INPUT_IMAGE_SIZE, 3)
feed_dict = {
images_placeholder: scaled_reshape,
phase_train_placeholder: False
}
emb_array = sess.run(embeddings,
feed_dict=feed_dict)
predictions = model.predict_proba(emb_array)
best_class_indices = np.argmax(predictions,
axis=1)
best_class_probabilities = predictions[
np.arange(len(best_class_indices)),
best_class_indices]
best_name = class_names[best_class_indices[0]]
print("Name: {}, Probability: {}".format(
best_name, best_class_probabilities))
cv2.rectangle(frame, (bb[i][0], bb[i][1]),
(bb[i][2], bb[i][3]),
(0, 255, 0), 2)
text_x = bb[i][0]
text_y = bb[i][3] + 20
if best_class_probabilities > 0.07:
name = class_names[best_class_indices[0]]
cv2.putText(frame,
name, (text_x, text_y),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
1, (0, 255, 0),
thickness=1,
lineType=2)
cv2.putText(
frame,
str(
round(best_class_probabilities[0],
3)), (text_x, text_y + 17),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
1, (0, 255, 0),
thickness=1,
lineType=2)
person_detected[best_name] += 1
print(person_detected)
print("most_common",
person_detected.most_common(1)[0][0])
if person_detected.most_common(
1)[0][1] > 50:
person_detected = collections.Counter()
#print("best_detect",best_detect)
else:
name = "Unknown"
cv2.putText(frame,
name, (text_x, text_y),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
1, (0, 255, 0),
thickness=1,
lineType=2)
cv2.putText(
frame,
str(
round(best_class_probabilities[0],
3)), (text_x, text_y + 17),
cv2.FONT_HERSHEY_COMPLEX_SMALL,
1, (0, 255, 0),
thickness=1,
lineType=2)
except:
pass
cv2.imshow('Face Recognition', frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
cap.release()
cv2.destroyAllWindows()
def train(self):
start_time = time.time()
# Config GPU options
if self.FLAGS.per_process_gpu_memory_fraction == 0.0:
gpu_options = tf.GPUOptions(allow_growth=True)
elif self.FLAGS.per_process_gpu_memory_fraction == 1.0:
gpu_options = tf.GPUOptions()
else:
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=self.
FLAGS.per_process_gpu_memory_fraction,
allow_growth=True)
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ['CUDA_VISIBLE_DEVICES'] = self.FLAGS.cuda_visible_devices
self.sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
if not tf.test.gpu_device_name():
self.logger.warning("No GPU is found")
else:
self.logger.info(tf.test.gpu_device_name())
global_step_lr = tf.Variable(0, trainable=False)
lr1 = tf.train.exponential_decay(
learning_rate=self.FLAGS.learning_rate,
global_step=global_step_lr,
decay_steps=1000,
decay_rate=0.995,
staircase=True)
lr2 = tf.train.exponential_decay(learning_rate=0.001,
global_step=global_step_lr,
decay_steps=1000,
decay_rate=0.995,
staircase=True)
with self.sess.as_default():
if self.FLAGS.experiment_type == "MSRTP":
self.model = MSRTP(self.FLAGS, self.emb, self.sess)
self.logger.info('Init finish.\tCost time: %.2fs' %
(time.time() - start_time))
#AUC暂时不看
# test_auc = self.model.metrics(sess=self.sess,
# batch_data=self.test_set,
# global_step=self.global_step,
# name='test auc')
# Eval init AUC
# self.logger.info('Init AUC: %.4f' % test_auc)
test_start = time.time()
self.hr_1, self.ndcg_1, self.hr_5, self.ndcg_5, self.hr_10, self.ndcg_10, self.hr_30, self.ndcg_30, self.hr_50, self.ndcg_50 = \
0,0,0,0,0,0,0,0,0,0
self.mse = 1000000
self.best_result_hr = []
self.best_result_ndcg = []
def eval_mse():
max_step = 0.
mse_lst = []
for step_i, batch_data in DataInput(
self.test_set, self.FLAGS.test_batch_size):
max_step = 1 + max_step
step_mse = self.model.metrics_mse(
sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
mse_lst.extend(list(step_mse[0]))
mse_val = np.mean(mse_lst)
if mse_val < self.mse:
self.mse = mse_val
print('----test mse: %.5f-----' % mse_val)
print('----MIN mse: %.5f-----' % (self.mse))
def eval_topk():
sum_hr_1, sum_ndcg_1, sum_hr_5, sum_ndcg_5, sum_hr_10, sum_ndcg_10, sum_hr_30, sum_ndcg_30, sum_hr_50, sum_ndcg_50 = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
result_list_hr_all = []
result_list_ndcg_all = []
max_step = 0
for step_i, batch_data in DataInput(
self.test_set, self.FLAGS.test_batch_size):
max_step = 1 + max_step
if self.FLAGS.experiment_type == "NARM" or \
self.FLAGS.experiment_type == "NARM+" or \
self.FLAGS.experiment_type == "NARM++": # or \
#self.FLAGS.experiment_type == "MTAM" :
hr_1, ndcg_1, hr_5, ndcg_5, hr_10, ndcg_10, hr_30, ndcg_30, hr_50, ndcg_50, \
result_list_hr, result_list_ndcg= \
self.model.metrics_topK_concat(sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
else:
hr_1, ndcg_1, hr_5, ndcg_5, hr_10, ndcg_10, hr_30, ndcg_30, hr_50, ndcg_50, \
result_list_hr, result_list_ndcg = \
self.model.metrics_topK(sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
sum_hr_1 = sum_hr_1 + hr_1
sum_ndcg_1 = sum_ndcg_1 + ndcg_1
sum_hr_5 = sum_hr_5 + hr_5
sum_ndcg_5 = sum_ndcg_5 + ndcg_5
sum_hr_10 = sum_hr_10 + hr_10
sum_ndcg_10 = sum_ndcg_10 + ndcg_10
sum_hr_30 = sum_hr_30 + hr_30
sum_ndcg_30 = sum_ndcg_30 + ndcg_30
sum_hr_50 = sum_hr_50 + hr_50
sum_ndcg_50 = sum_ndcg_50 + ndcg_50
result_list_hr_all = result_list_hr_all + result_list_hr
result_list_ndcg_all = result_list_ndcg_all + result_list_ndcg
sum_hr_1 /= max_step
sum_ndcg_1 /= max_step
sum_hr_5 /= max_step
sum_ndcg_5 /= max_step
sum_hr_10 /= max_step
sum_ndcg_10 /= max_step
sum_hr_30 /= max_step
sum_ndcg_30 /= max_step
sum_hr_50 /= max_step
sum_ndcg_50 /= max_step
if sum_hr_1 > self.hr_1 and sum_ndcg_1 > self.ndcg_1:
self.hr_1, self.ndcg_1 = sum_hr_1, sum_ndcg_1
if sum_hr_5 > self.hr_5 and sum_ndcg_5 > self.ndcg_5:
self.hr_5, self.ndcg_5 = sum_hr_5, sum_ndcg_5
if sum_hr_10 > self.hr_10 and sum_ndcg_10 > self.ndcg_10:
self.hr_10, self.ndcg_10 = sum_hr_10, sum_ndcg_10
self.best_result_hr = result_list_hr_all
self.best_result_ndcg = result_list_ndcg_all
if sum_hr_30 > self.hr_30 and sum_ndcg_30 > self.ndcg_30:
self.hr_30, self.ndcg_30 = sum_hr_30, sum_ndcg_30
if sum_hr_50 > self.hr_50 and sum_ndcg_50 > self.ndcg_50:
self.hr_50, self.ndcg_50 = sum_hr_50, sum_ndcg_50
def summery(k, hr, ndcg):
tag_recall = 'recall@' + str(k)
tag_ndcg = 'ndgc@' + str(k)
summary_recall_rate = tf.Summary(value=[
tf.Summary.Value(tag=tag_recall, simple_value=hr)
])
self.model.train_writer.add_summary(
summary_recall_rate, global_step=self.global_step)
summary_avg_ndcg = tf.Summary(value=[
tf.Summary.Value(tag=tag_ndcg, simple_value=ndcg)
])
self.model.train_writer.add_summary(
summary_avg_ndcg, global_step=self.global_step)
self.logger.info(
'Test recall rate @ %d : %.4f ndcg @ %d: %.4f' %
(k, hr, k, ndcg))
summery(1, sum_hr_1, sum_ndcg_1)
summery(5, sum_hr_5, sum_ndcg_5)
summery(10, sum_hr_10, sum_ndcg_10)
summery(30, sum_hr_30, sum_ndcg_30)
summery(50, sum_hr_50, sum_ndcg_50)
eval_topk()
eval_mse()
self.logger.info('End test. \tTest Cost time: %.2fs' %
(time.time() - test_start))
# Start training
self.logger.info(
'Training....\tmax_epochs:%d\tepoch_size:%d' %
(self.FLAGS.max_epochs, self.FLAGS.train_batch_size))
start_time, avg_loss, self.best_auc, self.best_recall, self.best_ndcg = time.time(
), 0.0, 0.0, 0.0, 0.0
loss_origin = []
loss_time = []
for epoch in range(self.FLAGS.max_epochs):
#if epoch > 2:
#lr = lr/1.5
random.shuffle(self.train_set)
self.logger.info('tain_set:%d' % len(self.train_set))
epoch_start_time = time.time()
learning_rate = self.FLAGS.learning_rate
for step_i, train_batch_data in DataInput(
self.train_set, self.FLAGS.train_batch_size):
try:
#print(self.sess.run(global_step_lr))
if learning_rate > 0.001:
learning_rate = self.sess.run(
lr1,
feed_dict={global_step_lr: self.global_step})
else:
learning_rate = self.sess.run(
lr2,
feed_dict={global_step_lr: self.global_step})
#print(learning_rate)
add_summary = bool(self.global_step %
self.FLAGS.display_freq == 0)
step_loss, step_loss_origin, step_loss_time, merge = self.model.train(
self.sess, train_batch_data, learning_rate,
add_summary, self.global_step, epoch)
self.sess.graph.finalize()
self.model.train_writer.add_summary(
merge, self.global_step)
avg_loss = avg_loss + step_loss
loss_origin.extend(step_loss_origin)
loss_time.extend(step_loss_time)
self.global_step = self.global_step + 1
self.one_epoch_step = self.one_epoch_step + 1
#evaluate for eval steps
if self.global_step % self.FLAGS.eval_freq == 0:
print(learning_rate)
self.logger.info("Epoch step is " +
str(self.one_epoch_step))
self.logger.info("Global step is " +
str(self.global_step))
self.logger.info("Train_loss is " +
str(avg_loss /
self.FLAGS.eval_freq))
self.logger.info(
"Cross Entropy Loss is " +
str(np.mean(np.array(loss_origin))))
self.logger.info("Time Loss is " +
str(np.mean(np.array(loss_time))))
# train_auc = self.model.metrics(sess=self.sess, batch_data=train_batch_data,
# global_step=self.global_step,name='train auc')
# self.logger.info('Batch Train AUC: %.4f' % train_auc)
# self.test_auc = self.model.metrics(sess=self.sess, batch_data=self.test_set,
# global_step=self.global_step,name='test auc')
# self.logger.info('Test AUC: %.4f' % self.test_auc)
eval_topk()
eval_mse()
avg_loss = 0
loss_origin = []
loss_time = []
loss_reconsume = []
self.save_model()
if self.FLAGS.draw_pic == True:
self.save_fig()
except Exception as e:
self.logger.info("Error！！！！！！！！！！！！")
self.logger.info(e)
traceback.print_exc()
self.logger.info('one epoch Cost time: %.2f' %
(time.time() - epoch_start_time))
self.logger.info("Epoch step is " + str(self.one_epoch_step))
self.logger.info("Global step is " + str(self.global_step))
self.logger.info("Train_loss is " + str(step_loss))
self.logger.info("Cross Entropy Loss is " +
str(np.mean(np.array(loss_origin))))
self.logger.info("Time Loss is " +
str(np.mean(np.array(loss_time))))
eval_topk()
eval_mse()
with open('best_result_hr_' + self.FLAGS.version, 'w+') as f:
f.write(str(self.best_result_hr))
with open('best_result_ndcg' + self.FLAGS.version, 'w+') as f:
f.write(str(self.best_result_ndcg))
self.logger.info('Max recall rate @ 1: %.4f ndcg @ 1: %.4f' %
(self.hr_1, self.ndcg_1))
self.logger.info('Max recall rate @ 5: %.4f ndcg @ 5: %.4f' %
(self.hr_5, self.ndcg_5))
self.logger.info(
'Max recall rate @ 10: %.4f ndcg @ 10: %.4f' %
(self.hr_10, self.ndcg_10))
self.logger.info(
'Max recall rate @ 30: %.4f ndcg @ 30: %.4f' %
(self.hr_30, self.ndcg_30))
self.logger.info(
'Max recall rate @ 50: %.4f ndcg @ 50: %.4f' %
(self.hr_50, self.ndcg_50))
self.one_epoch_step = 0
#if self.global_step > 1000:
#lr = lr / 2
#if lr < 0.0005:
#lr = lr * 0.99
#elif self.FLAGS.type == "tmall":
#lr = lr * 0.5
#else:
#lr = lr * 0.98
self.logger.info('Epoch %d DONE\tCost time: %.2f' %
(self.now_epoch, time.time() - start_time))
self.now_epoch = self.now_epoch + 1
self.one_epoch_step = 0
self.model.save(self.sess, self.global_step)
self.logger.info('best test_auc: ' + str(self.best_auc))
self.logger.info('best recall: ' + str(self.best_recall))
self.logger.info('Finished')
def set_threshold(self):
dataset = facenet.get_dataset(self.input_datadir)
with tf.Graph().as_default():
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.5)
sess = tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options, log_device_placement=False))
with sess.as_default():
pnet, rnet, onet = detect_face.create_mtcnn(
sess, './Model/data')
image_size = 180
input_image_size = 160
facenet.load_model(self.model)
images_placeholder = tf.get_default_graph().get_tensor_by_name(
"input:0")
embeddings = tf.get_default_graph().get_tensor_by_name(
"embeddings:0")
phase_train_placeholder = tf.get_default_graph(
).get_tensor_by_name("phase_train:0")
embedding_size = embeddings.get_shape()[1]
emb_array = np.zeros((1, embedding_size))
self.classifier_exp = os.path.expanduser(self.classifier)
with open(self.classifier_exp, 'rb') as infile:
(model, class_names) = pickle.load(infile)
threshold_dict = {}
index = 0
for data in dataset:
total = 0
for path in data.image_paths:
frame = cv2.imread(path, 0)
if frame.ndim == 2:
frame = facenet.to_rgb(frame)
cropped = []
scaled = []
scaled_reshape = []
cropped = facenet.flip(frame, False)
scaled.append(
cv2.resize(cropped, (image_size, image_size),
interpolation=cv2.INTER_LINEAR))
scaled = cv2.resize(scaled[0],
(input_image_size, input_image_size),
interpolation=cv2.INTER_CUBIC)
scaled = facenet.prewhiten(scaled)
scaled_reshape.append(
scaled.reshape(-1, input_image_size, input_image_size,
3))
feed_dict = {
images_placeholder: scaled_reshape[0],
phase_train_placeholder: False
}
emb_array[0, :] = sess.run(embeddings, feed_dict=feed_dict)
predictions = model.predict_proba(emb_array)
best_class_indices = np.argmax(predictions, axis=1)
best_class_probabilities = predictions[
np.arange(len(best_class_indices)), best_class_indices]
if best_class_indices[0] != index:
total -= predictions[0][best_class_indices[0]]
else:
total += predictions[0][index]
try:
threshold_dict[str(
data.name)] = total / len(data.image_paths)
except ZeroDivisionError:
pass
index += 1
try:
with open('./Model/prediction_threshold/threshold.pickle',
'wb') as handle:
pickle.dump(threshold_dict,
handle,
protocol=pickle.HIGHEST_PROTOCOL)
except FileNotFoundError:
os.mkdir('./Model/prediction_threshold')
with open('./Model/prediction_threshold/threshold.pickle',
'wb') as handle:
pickle.dump(threshold_dict,
handle,
protocol=pickle.HIGHEST_PROTOCOL)
def benchmark(self, ckpt_dir, outer_steps=100, inner_steps=1000):
"""Run repeatedly on dummy data to benchmark inference."""
# Turn off Grappler optimizations.
options = {"disable_meta_optimizer": True}
tf.config.optimizer.set_experimental_options(options)
# Create the model outside the loop body.
hparams = registry.hparams(self.hparams_set)
hparams_lib.add_problem_hparams(hparams, self.problem_name)
model_cls = registry.model(self.model_name)
model = model_cls(hparams, tf.estimator.ModeKeys.EVAL)
# Run only the model body (no data pipeline) on device.
feature_shape = [
hparams.batch_size, 3 * self.image_size * self.image_size
]
features = {"targets": tf.zeros(feature_shape, dtype=tf.int32)}
# Call the model once to initialize the variables. Note that
# this should never execute.
with tf.variable_scope(self.model_name) as vso:
transformed_features = model.bottom(features)
with tf.variable_scope("body") as vsi:
body_out = model.body(transformed_features)
logits = model.top(body_out, features)
model.loss(logits, features)
def call_model(features):
with tf.variable_scope(vso, reuse=tf.AUTO_REUSE):
transformed_features = model.bottom(features)
with tf.variable_scope(vsi, reuse=tf.AUTO_REUSE):
body_out = model.body(transformed_features)
logits = model.top(body_out, features)
return model.loss(logits, features)
# Run the function body in a loop to amortize session overhead.
loop_index = tf.zeros([], dtype=tf.int32)
initial_loss = (tf.zeros([]), tf.zeros([]))
def loop_cond(idx, _):
return tf.less(idx, tf.constant(inner_steps, dtype=tf.int32))
def loop_body(idx, _):
return idx + 1, call_model(features)
benchmark_op = tf.while_loop(loop_cond,
loop_body, [loop_index, initial_loss],
parallel_iterations=1,
back_prop=False)
session_config = tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=False, per_process_gpu_memory_fraction=0.95))
run_metadata = tf.RunMetadata()
with tf.Session(config=session_config) as sess:
self.restore_model(sess, ckpt_dir)
tps = []
for idx in range(outer_steps):
start_time = time.time()
sess.run(benchmark_op, run_metadata=run_metadata)
elapsed_time = time.time() - start_time
tps.append(inner_steps * hparams.batch_size * (64 * 64 * 3) /
elapsed_time)
logging.error("Iterations %d processed %f TPS.", idx, tps[-1])
# Skip the first iteration where all the setup and allocation happens.
tps = np.asarray(tps[1:])
logging.error("Mean/Std/Max/Min throughput = %f / %f / %f / %f",
np.mean(tps), np.std(tps), tps.max(), tps.min())
self.train()
time.sleep(0.01)
if __name__ == '__main__':
FPS = 60
# For stats
ep_rewards = [-200]
# For more repetitive results
random.seed(1)
np.random.seed(1)
tf.set_random_seed(1)
# Memory fraction, used mostly when trai8ning multiple agents
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=MEMORY_FRACTION)
backend.set_session(
tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)))
# Create models folder
if not os.path.isdir('models'):
os.makedirs('models')
# Create agent and environment
agent = DQNAgent()
env = CarEnv()
# Start training thread and wait for training to be initialized
trainer_thread = Thread(target=agent.train_in_loop, daemon=True)
trainer_thread.start()
while not agent.training_initialized:
def main(args):
sleep(random.random())
output_dir = os.path.expanduser(args.output_dir)
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# Store some git revision info in a text file in the log directory
src_path,_ = os.path.split(os.path.realpath(__file__))
facenet.store_revision_info(src_path, output_dir, ' '.join(sys.argv))
dataset = facenet.get_dataset(args.input_dir)
print('Creating networks and loading parameters')
with tf.Graph().as_default():
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=args.gpu_memory_fraction)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
with sess.as_default():
pnet, rnet, onet = align.detect_face.create_mtcnn(sess, None)
minsize = 20 # minimum size of face
threshold = [ 0.6, 0.7, 0.7 ] # three steps's threshold
factor = 0.709 # scale factor
# Add a random key to the filename to allow alignment using multiple processes
random_key = np.random.randint(0, high=99999)
bounding_boxes_filename = os.path.join(output_dir, 'bounding_boxes_%05d.txt' % random_key)
with open(bounding_boxes_filename, "w") as text_file:
nrof_images_total = 0
nrof_successfully_aligned = 0
if args.random_order:
random.shuffle(dataset)
for cls in dataset:
output_class_dir = os.path.join(output_dir, cls.name)
if not os.path.exists(output_class_dir):
os.makedirs(output_class_dir)
if args.random_order:
random.shuffle(cls.image_paths)
for image_path in cls.image_paths:
nrof_images_total += 1
filename = os.path.splitext(os.path.split(image_path)[1])[0]
output_filename = os.path.join(output_class_dir, filename+'.png')
print(image_path)
if not os.path.exists(output_filename):
try:
img = misc.imread(image_path)
except (IOError, ValueError, IndexError) as e:
errorMessage = '{}: {}'.format(image_path, e)
print(errorMessage)
else:
if img.ndim<2:
print('Unable to align "%s"' % image_path)
text_file.write('%s\n' % (output_filename))
continue
if img.ndim == 2:
img = facenet.to_rgb(img)
img = img[:,:,0:3]
bounding_boxes, _ = align.detect_face.detect_face(img, minsize, pnet, rnet, onet, threshold, factor)
nrof_faces = bounding_boxes.shape[0]
if nrof_faces>0:
det = bounding_boxes[:,0:4]
det_arr = []
img_size = np.asarray(img.shape)[0:2]
if nrof_faces>1:
if args.detect_multiple_faces:
for i in range(nrof_faces):
det_arr.append(np.squeeze(det[i]))
else:
bounding_box_size = (det[:,2]-det[:,0])*(det[:,3]-det[:,1])
img_center = img_size / 2
offsets = np.vstack([ (det[:,0]+det[:,2])/2-img_center[1], (det[:,1]+det[:,3])/2-img_center[0] ])
offset_dist_squared = np.sum(np.power(offsets,2.0),0)
index = np.argmax(bounding_box_size-offset_dist_squared*2.0) # some extra weight on the centering
det_arr.append(det[index,:])
else:
det_arr.append(np.squeeze(det))
for i, det in enumerate(det_arr):
det = np.squeeze(det)
bb = np.zeros(4, dtype=np.int32)
bb[0] = np.maximum(det[0]-args.margin/2, 0)
bb[1] = np.maximum(det[1]-args.margin/2, 0)
bb[2] = np.minimum(det[2]+args.margin/2, img_size[1])
bb[3] = np.minimum(det[3]+args.margin/2, img_size[0])
cropped = img[bb[1]:bb[3],bb[0]:bb[2],:]
scaled = misc.imresize(cropped, (args.image_size, args.image_size), interp='bilinear')
nrof_successfully_aligned += 1
filename_base, file_extension = os.path.splitext(output_filename)
if args.detect_multiple_faces:
output_filename_n = "{}_{}{}".format(filename_base, i, file_extension)
else:
output_filename_n = "{}{}".format(filename_base, file_extension)
misc.imsave(output_filename_n, scaled)
text_file.write('%s %d %d %d %d\n' % (output_filename_n, bb[0], bb[1], bb[2], bb[3]))
else:
print('Unable to align "%s"' % image_path)
text_file.write('%s\n' % (output_filename))
print('Total number of images: %d' % nrof_images_total)
print('Number of successfully aligned images: %d' % nrof_successfully_aligned)
def set_session(device_count=None, seed=0):
gpu_options = tf.GPUOptions(allow_growth=True)
config = tf.ConfigProto(gpu_options=gpu_options, device_count=device_count)
def tf_execute_helper(data_cache, epoch, partitions, checkpoint_path, input_paths,
input_fn_string, model_fn_string, train_fn_string, mst, train=True):
"""
:param data_cache:
:param epoch:
:param partitions:
:param checkpoint_path:
:param input_paths:
:param input_fn_string:
:param model_fn_string:
:param train_fn_string:
:param mst:
:param train:
:return:
"""
begin_time = time.time()
tf.reset_default_graph()
input_fn = dill.loads(base64.b64decode(input_fn_string))
model_fn = dill.loads(base64.b64decode(model_fn_string))
train_fn = dill.loads(base64.b64decode(train_fn_string))
losses = []
errors = []
message = ""
for partition, input_path in zip(partitions, input_paths):
tf.reset_default_graph()
if input_path in data_cache:
# already cached
data = data_cache[input_path]
else:
data = input_fn(input_path)
data_cache[input_path] = data
opt, loss, error = model_fn(data, mst)
train_step = opt.minimize(loss, colocate_gradients_with_ops=True)
saver = tf.train.Saver()
gpu_options = tf.GPUOptions(allow_growth = True)
sess = tf.Session(
config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False, gpu_options=gpu_options))
if os.path.exists(checkpoint_path + ".index"):
saver.restore(sess, checkpoint_path)
else:
sess.run(tf.global_variables_initializer())
message += "\nEPOCH: %d, PARTITION: %d, Model Building and Session Initialization Time: %d\n" % (
epoch, partition, time.time() - begin_time)
time_begin = time.time()
loss_val, error_val = train_fn(sess, train_step, loss, error, train=train)
losses.append(loss_val)
errors.append(error_val)
elapsed_time = time.time() - time_begin
mode = "TRAIN" if train else "VALID"
message += "EPOCH: %d, PARTITION: %d, %s LOSS: %f, ERROR: %f, Time: %f\n" % (epoch, partition, mode, loss_val, error_val, elapsed_time)
if train:
begin_time = time.time()
saver.save(sess, checkpoint_path)
message += "EPOCH: %d, PARTITION: %d, Checkpoint Save Time: %d\n" % (epoch, partition, time.time() - begin_time)
sess.close()
return {'loss': losses, 'error': errors, 'message': message[1:]}
def main(_):
tf.logging.set_verbosity(tf.logging.INFO)
# If we have ps_hosts, then we'll assume that this is going to be a
# distributed training run. Configure the cluster appropriately. Otherwise,
# we just do everything in-process.
if FLAGS.ps_hosts:
cluster = tf.train.ClusterSpec({
'ps': FLAGS.ps_hosts.split(','),
'worker': FLAGS.worker_hosts.split(','),
})
if FLAGS.job_name == 'ps':
# Ignore the GPU if we're the parameter server. This let's the PS run on
# the same machine as a worker.
config = tf.ConfigProto(device_count={'GPU': 0})
elif FLAGS.job_name == 'worker':
config = tf.ConfigProto(gpu_options=tf.GPUOptions(
visible_device_list='%d' % FLAGS.gpu_device,
allow_growth=True))
else:
raise ValueError('unknown job name "%s"' % FLAGS.job_name)
server = tf.train.Server(
cluster,
job_name=FLAGS.job_name,
task_index=FLAGS.task_index,
config=config)
if FLAGS.job_name == 'ps':
return server.join()
device_setter = tf.train.replica_device_setter(
worker_device='/job:worker/task:%d' % FLAGS.task_index,
cluster=cluster)
else:
server = None
device_setter = tf.train.replica_device_setter(0)
# Build the graph.
with tf.Graph().as_default():
with tf.device(device_setter):
model = Model(FLAGS.input_base_path, FLAGS)
# If an eval path is present, then create eval operators and set up scalar
# summaries to report on the results. Run the evals on the CPU since
# the analogy eval requires a fairly enormous tensor to be allocated to
# do the nearest neighbor search.
if FLAGS.eval_base_path:
wordsim_filenames = glob.glob(
os.path.join(FLAGS.eval_base_path, '*.ws.tab'))
for filename in wordsim_filenames:
name = os.path.basename(filename).split('.')[0]
with tf.device(tf.DeviceSpec(device_type='CPU')):
op = model.wordsim_eval_op(filename, FLAGS.eval_word_prefix)
tf.summary.scalar(name, op)
analogy_filenames = glob.glob(
os.path.join(FLAGS.eval_base_path, '*.an.tab'))
for filename in analogy_filenames:
name = os.path.basename(filename).split('.')[0]
with tf.device(tf.DeviceSpec(device_type='CPU')):
op = model.analogy_eval_op(filename)
tf.summary.scalar(name, op)
tf.summary.scalar('loss', model.loss_op)
tf.summary.scalar('l2_loss', model.l2_loss_op)
tf.summary.scalar('sigmoid_loss', model.sigmoid_loss_op)
tf.summary.scalar('vector_regul_loss', model.vector_regul_loss_op)
# Train on, soldier.
supervisor = tf.train.Supervisor(
logdir=FLAGS.output_base_path,
is_chief=(FLAGS.task_index == 0),
save_summaries_secs=60,
recovery_wait_secs=5)
max_step = FLAGS.num_epochs * model.steps_per_epoch
master = server.target if server else ''
with supervisor.managed_session(master) as session:
local_step = 0
global_step = session.run(model.global_step)
while not supervisor.should_stop() and global_step < max_step:
global_step, loss, _ = session.run([
model.global_step, model.loss_op, model.train_op])
if not np.isfinite(loss):
raise ValueError('non-finite cost at step %d' % global_step)
local_step += 1
if local_step % 10 == 0:
tf.logging.info(
'local_step=%d global_step=%d loss=%.1f, %.1f%% complete',
local_step, global_step, loss, 100.0 * global_step / max_step)
if FLAGS.task_index == 0:
supervisor.saver.save(
session, supervisor.save_path, global_step=global_step)
model.write_embeddings(FLAGS, session)
def main(args):
img_mean = np.array([134.10714722, 102.52040863, 87.15436554])
img_stddev = np.sqrt(np.array([3941.30175781, 2856.94287109, 2519.35791016]))
vae_def = importlib.import_module(args.vae_def)
vae = vae_def.Vae(args.latent_var_size)
gen_image_size = vae.get_image_size()
with tf.Graph().as_default():
tf.set_random_seed(args.seed)
images = tf.placeholder(tf.float32, shape=(None,gen_image_size,gen_image_size,3), name='input')
# Normalize
images_norm = (images-img_mean) / img_stddev
# Resize to appropriate size for the encoder
images_norm_resize = tf.image.resize_images(images_norm, (gen_image_size,gen_image_size))
# Create encoder network
mean, log_variance = vae.encoder(images_norm_resize, True)
epsilon = tf.random_normal((tf.shape(mean)[0], args.latent_var_size))
std = tf.exp(log_variance/2)
latent_var = mean + epsilon * std
# Create decoder
reconstructed_norm = vae.decoder(latent_var, False)
# Un-normalize
reconstructed = (reconstructed_norm*img_stddev) + img_mean
# Create a saver
saver = tf.train.Saver(tf.trainable_variables(), max_to_keep=3)
# Start running operations on the Graph
gpu_memory_fraction = 1.0
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=gpu_memory_fraction)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options, log_device_placement=False))
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
tf.train.start_queue_runners(coord=coord, sess=sess)
with sess.as_default():
vae_checkpoint = os.path.expanduser(args.vae_checkpoint)
print('Restoring VAE checkpoint: %s' % vae_checkpoint)
saver.restore(sess, vae_checkpoint)
filename = os.path.expanduser(args.attributes_filename)
with h5py.File(filename,'r') as f:
latent_vars = np.array(f.get('latent_vars'))
attributes = np.array(f.get('attributes'))
#fields = np.array(f.get('fields'))
attribute_vectors = np.array(f.get('attribute_vectors'))
# Reconstruct faces while adding varying amount of the selected attribute vector
attribute_index = 31 # 31: 'Smiling'
image_indices = [8,11,13,18,19,26,31,39,47,54,56,57,58,59,60,73]
nrof_images = len(image_indices)
nrof_interp_steps = 10
sweep_latent_var = np.zeros((nrof_interp_steps*nrof_images, args.latent_var_size), np.float32)
for j in range(nrof_images):
image_index = image_indices[j]
idx = np.argwhere(attributes[:,attribute_index]==-1)[image_index,0]
for i in range(nrof_interp_steps):
sweep_latent_var[i+nrof_interp_steps*j,:] = latent_vars[idx,:] + 5.0*i/nrof_interp_steps*attribute_vectors[attribute_index,:]
recon = sess.run(reconstructed, feed_dict={latent_var:sweep_latent_var})
img = facenet.put_images_on_grid(recon, shape=(nrof_interp_steps*2,int(math.ceil(nrof_images/2))))
image_filename = os.path.expanduser(args.output_image_filename)
print('Writing generated image to %s' % image_filename)
misc.imsave(image_filename, img)
import numpy as np
import tensorflow.compat.v1 as tf
import vgg16
import tmp_model
import time
img1 = tmp_model.load_image("./test_img/tiger.jpeg")
img2 = tmp_model.load_image("./test_img/puzzle.jpeg")
batch1 = img1.reshape((1, 128, 128, 3))
batch2 = img2.reshape((1, 128, 128, 3))
batch = np.concatenate((batch1, batch2), 0)
with tf.Session(config=tf.ConfigProto(gpu_options=(tf.GPUOptions(
per_process_gpu_memory_fraction=0.7)))) as sess:
images = tf.placeholder("float", [2, 128, 128, 3], name="pc_img")
feed_dict = {images: batch}
noise = np.random.random([2, 128, 128, 3])
vgg = vgg16.Vgg16("/tmp/vgg16/vgg16.npy")
with tf.name_scope("content_vgg"):
vgg.build(images)
t1 = time.time() * 1000
prob = sess.run(vgg.prob, feed_dict={images: noise})
prob2 = sess.run(vgg.prob, feed_dict={images: batch})
print(prob)
print(prob2)
t2 = time.time() * 1000
print(t2 - t1)
def main(args):
model = SARGAN(img_size, BATCH_SIZE, img_channel=img_size[2])
with tf.variable_scope("d_opt", reuse=tf.AUTO_REUSE):
d_opt = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE).minimize(
model.d_loss, var_list=model.d_vars)
with tf.variable_scope("g_opt", reuse=tf.AUTO_REUSE):
g_opt = tf.train.AdamOptimizer(learning_rate=LEARNING_RATE).minimize(
model.g_loss, var_list=model.g_vars)
if (retrain == 0):
saver = tf.train.Saver(max_to_keep=20)
else:
saver = tf.train.import_meta_graph(trained_model_path +
'/sargan_mnist.meta')
gpu_options = tf.GPUOptions(allow_growth=True,
visible_device_list=str(GPU_ID))
config = tf.ConfigProto(gpu_options=gpu_options)
progress_bar = tqdm(range(MAX_EPOCH), unit="epoch")
#list of loss values each item is the loss value of one ieteration
train_d_loss_values = []
train_g_loss_values = []
#test_imgs, test_classes = get_data(test_filename)
#imgs, classes = get_data(train_filename)
with tf.Session(config=config) as sess:
if (retrain == 0):
sess.run(tf.global_variables_initializer())
else:
sess.run(tf.global_variables_initializer())
saver.restore(sess, tf.train.latest_checkpoint(trained_model_path))
#test_copies = test_imgs.astype('float32')
for epoch in progress_bar:
NUM_TEST_PER_EPOCH = 1
counter = 0
epoch_start_time = time.time()
encoded_data, original_data = transfrom_data(NUM_TEST_PER_EPOCH)
#shuffle(copies)
#divide the images into equal sized batches
#image_batches = np.array(list(chunks(copies, BATCH_SIZE)))
for i in range(NUM_ITERATION):
#getting a batch from the training data
#one_batch_of_imgs = image_batches[i]
#copy the batch
features = original_data[i]
#corrupt the images
corrupted_batch = encoded_data[i]
_, m = sess.run([d_opt, model.d_loss],
feed_dict={
model.image: features,
model.cond: corrupted_batch
})
_, M = sess.run([g_opt, model.g_loss],
feed_dict={
model.image: features,
model.cond: corrupted_batch
})
train_d_loss_values.append(m)
train_g_loss_values.append(M)
#print some notifications
counter += 1
if counter % 25 == 0:
print("\rEpoch [%d], Iteration [%d]: time: %4.4f, d_loss: %.8f, g_loss: %.8f" \
% (epoch, counter, time.time() - epoch_start_time, m, M))
# save the trained network
if epoch % SAVE_EVERY_EPOCH == 0:
save_path = saver.save(sess,
(trained_model_path + "/sargan_mnist"))
print("\n\nModel saved in file: %s\n\n" % save_path)
import numpy as np
import matplotlib.pyplot as plt
dataset = BinaryDbReader(mode='training',
batch_size=8,
shuffle=True,
hand_crop=True,
use_wrist_coord=False,
sigma=10,
coord_uv_noise=False,
crop_center_noise=False,
crop_offset_noise=False,
crop_scale_noise=False)
data = dataset.get()
print(data)
# Start TF
gpu_options = tf.GPUOptions(allow_growth=True, )
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
tf.train.start_queue_runners(sess=sess)
sess.run(tf.global_variables_initializer())
while 1:
d = sess.run(data)
# print(d)
hm_sum = np.sum(d["scoremap"][0], axis=2)
plt.figure(figsize=(12, 12))
plt.subplot(121)
plt.imshow(hm_sum)
plt.subplot(122)
img = d["image_crop"][0] + 0.5
img[:, :, 0] += hm_sum
plt.imshow(img)
plt.show()
x_tf = tf.identity(x_in)
model = nn_dict[model_type](is_train, hps)
cleverhans_model = CallableModelWrapper(forward_pass, 'logits')
# Separate forward pass graph for Cleverhans wrapper (for C&W and PGD attacks) placed on the last GPU
logits = forward_pass(x_tf)
# Error rate
incorrect_prediction = tf.not_equal(tf.argmax(logits, 1),
tf.argmax(y_in, 1))
err_rate = tf.reduce_mean(tf.cast(incorrect_prediction, tf.float32))
# GPU settings
gpu_options = tf.GPUOptions(visible_device_list=str(hps.gpus)[1:-1],
per_process_gpu_memory_fraction=hps.gpu_memory)
config = tf.ConfigProto(allow_soft_placement=True, gpu_options=gpu_options)
# --------- Pytorch part for the Kolter-Wong model ----------
device = torch.device('cuda:' + str(hps.gpus[-1]))
torch.cuda.set_device(hps.gpus[-1])
model_torch = kolter_wong.models.select_model(hps.nn_type, hps.n_in,
hps.n_out).to(device)
for var in model_torch.parameters():
var.requires_grad = False
# ---------- end ----------
with tf.Session(graph=graph, config=config) as sess:
with graph.as_default(), tf.device('/gpu:0'):
pgd_ae_tensor = ae.pgd_attack(x_tf, y_in, cleverhans_model, hps.p, eps,
def initialize_globals():
c = AttrDict()
# Augmentations
c.augmentations = parse_augmentations(FLAGS.augment)
if len(c.augmentations
) > 0 and FLAGS.feature_cache and FLAGS.cache_for_epochs == 0:
log_warn(
'Due to current feature-cache settings the exact same sample augmentations of the first '
'epoch will be repeated on all following epochs. This could lead to unintended over-fitting. '
'You could use --cache_for_epochs <n_epochs> to invalidate the cache after a given number of epochs.'
)
# Caching
if FLAGS.cache_for_epochs == 1:
log_warn(
'--cache_for_epochs == 1 is (re-)creating the feature cache on every epoch but will never use it.'
)
# Read-buffer
FLAGS.read_buffer = parse_file_size(FLAGS.read_buffer)
# Set default dropout rates
if FLAGS.dropout_rate2 < 0:
FLAGS.dropout_rate2 = FLAGS.dropout_rate
if FLAGS.dropout_rate3 < 0:
FLAGS.dropout_rate3 = FLAGS.dropout_rate
if FLAGS.dropout_rate6 < 0:
FLAGS.dropout_rate6 = FLAGS.dropout_rate
# Set default checkpoint dir
if not FLAGS.checkpoint_dir:
FLAGS.checkpoint_dir = xdg.save_data_path(
os.path.join('deepspeech', 'checkpoints'))
if FLAGS.load_train not in ['last', 'best', 'init', 'auto']:
FLAGS.load_train = 'auto'
if FLAGS.load_evaluate not in ['last', 'best', 'auto']:
FLAGS.load_evaluate = 'auto'
# Set default summary dir
if not FLAGS.summary_dir:
FLAGS.summary_dir = xdg.save_data_path(
os.path.join('deepspeech', 'summaries'))
# Standard session configuration that'll be used for all new sessions.
c.session_config = tfv1.ConfigProto(
allow_soft_placement=True,
log_device_placement=FLAGS.log_placement,
inter_op_parallelism_threads=FLAGS.inter_op_parallelism_threads,
intra_op_parallelism_threads=FLAGS.intra_op_parallelism_threads,
gpu_options=tfv1.GPUOptions(allow_growth=FLAGS.use_allow_growth))
# CPU device
c.cpu_device = '/cpu:0'
# Available GPU devices
c.available_devices = get_available_gpus(c.session_config)
# If there is no GPU available, we fall back to CPU based operation
if not c.available_devices:
c.available_devices = [c.cpu_device]
if FLAGS.bytes_output_mode:
c.alphabet = UTF8Alphabet()
else:
c.alphabet = Alphabet(os.path.abspath(FLAGS.alphabet_config_path))
# Geometric Constants
# ===================
# For an explanation of the meaning of the geometric constants, please refer to
# doc/Geometry.md
# Number of MFCC features
c.n_input = 26 # TODO: Determine this programmatically from the sample rate
# The number of frames in the context
c.n_context = 9 # TODO: Determine the optimal value using a validation data set
# Number of units in hidden layers
c.n_hidden = FLAGS.n_hidden
c.n_hidden_1 = c.n_hidden
c.n_hidden_2 = c.n_hidden
c.n_hidden_5 = c.n_hidden
# LSTM cell state dimension
c.n_cell_dim = c.n_hidden
# The number of units in the third layer, which feeds in to the LSTM
c.n_hidden_3 = c.n_cell_dim
# Units in the sixth layer = number of characters in the target language plus one
c.n_hidden_6 = c.alphabet.GetSize() + 1 # +1 for CTC blank label
# Size of audio window in samples
if (FLAGS.feature_win_len * FLAGS.audio_sample_rate) % 1000 != 0:
log_error(
'--feature_win_len value ({}) in milliseconds ({}) multiplied '
'by --audio_sample_rate value ({}) must be an integer value. Adjust '
'your --feature_win_len value or resample your audio accordingly.'
''.format(FLAGS.feature_win_len, FLAGS.feature_win_len / 1000,
FLAGS.audio_sample_rate))
sys.exit(1)
c.audio_window_samples = FLAGS.audio_sample_rate * (FLAGS.feature_win_len /
1000)
# Stride for feature computations in samples
if (FLAGS.feature_win_step * FLAGS.audio_sample_rate) % 1000 != 0:
log_error(
'--feature_win_step value ({}) in milliseconds ({}) multiplied '
'by --audio_sample_rate value ({}) must be an integer value. Adjust '
'your --feature_win_step value or resample your audio accordingly.'
''.format(FLAGS.feature_win_step, FLAGS.feature_win_step / 1000,
FLAGS.audio_sample_rate))
sys.exit(1)
c.audio_step_samples = FLAGS.audio_sample_rate * (FLAGS.feature_win_step /
1000)
if FLAGS.one_shot_infer:
if not path_exists_remote(FLAGS.one_shot_infer):
log_error(
'Path specified in --one_shot_infer is not a valid file.')
sys.exit(1)
if FLAGS.train_cudnn and FLAGS.load_cudnn:
log_error('Trying to use --train_cudnn, but --load_cudnn '
'was also specified. The --load_cudnn flag is only '
'needed when converting a CuDNN RNN checkpoint to '
'a CPU-capable graph. If your system is capable of '
'using CuDNN RNN, you can just specify the CuDNN RNN '
'checkpoint normally with --save_checkpoint_dir.')
sys.exit(1)
# If separate save and load flags were not specified, default to load and save
# from the same dir.
if not FLAGS.save_checkpoint_dir:
FLAGS.save_checkpoint_dir = FLAGS.checkpoint_dir
if not FLAGS.load_checkpoint_dir:
FLAGS.load_checkpoint_dir = FLAGS.checkpoint_dir
ConfigSingleton._config = c # pylint: disable=protected-access
def __init__(self,
hidden_size=100,
out_size=100,
batch_size=300,
n_node=None,
lr=None,
l2=None,
step=1,
decay=None,
lr_dc=0.1,
nonhybrid=False):
super(GGNN, self).__init__(hidden_size, out_size, batch_size,
nonhybrid)
self.embedding = tf.get_variable(
shape=[n_node, hidden_size],
name='embedding',
dtype=tf.float32,
initializer=tf.random_uniform_initializer(-self.stdv, self.stdv))
self.adj_in = tf.placeholder(dtype=tf.float32,
shape=[self.batch_size, None, None])
self.adj_out = tf.placeholder(dtype=tf.float32,
shape=[self.batch_size, None, None])
self.n_node = n_node
self.L2 = l2
self.step = step
self.nonhybrid = nonhybrid
self.W_in = tf.get_variable('W_in',
shape=[self.out_size, self.out_size],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(
-self.stdv, self.stdv))
self.b_in = tf.get_variable('b_in', [self.out_size],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(
-self.stdv, self.stdv))
self.W_out = tf.get_variable('W_out', [self.out_size, self.out_size],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(
-self.stdv, self.stdv))
self.b_out = tf.get_variable('b_out', [self.out_size],
dtype=tf.float32,
initializer=tf.random_uniform_initializer(
-self.stdv, self.stdv))
with tf.variable_scope('ggnn_model', reuse=None):
self.loss_train, _ = self.forward(self.ggnn())
with tf.variable_scope('ggnn_model', reuse=True):
self.loss_test, self.score_test = self.forward(self.ggnn(),
train=False)
self.global_step = tf.Variable(0)
self.learning_rate = tf.train.exponential_decay(
lr,
global_step=self.global_step,
decay_steps=decay,
decay_rate=lr_dc,
staircase=True)
self.opt = tf.train.AdamOptimizer(self.learning_rate).minimize(
self.loss_train, global_step=self.global_step)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
config = tf.ConfigProto(gpu_options=gpu_options)
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
self.sess.run(tf.global_variables_initializer())
# But many TF deprecation warnings in 1.14.0, e.g.:
# "The name tf.GPUOptions is deprecated. Please use tf.compat.v1.GPUOptions
# instead". See tf_export.py
if g.tf_ver >= parse_version('1.14.0'):
import tensorflow.compat.v1 as tf
else:
import tensorflow as tf
# TODO(KGF): above, builder.py (bug workaround), mpi_launch_tensorflow.py,
# and runner.py are the only files that import tensorflow directly
from keras.backend.tensorflow_backend import set_session
# KGF: next 3 lines dump many TensorFlow diagnostics to stderr.
# All MPI ranks first "Successfully opened dynamic library libcuda"
# then, one by one: ID GPU, libcudart, libcublas, libcufft, ...
# Finally, "Device interconnect StreamExecutor with strength 1 edge matrix"
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.95,
allow_growth=True)
config = tf.ConfigProto(gpu_options=gpu_options)
set_session(tf.Session(config=config))
g.flush_all_inorder()
else:
os.environ['KERAS_BACKEND'] = 'theano'
base_compile_dir = '{}/tmp/{}-{}'.format(conf['paths']['output_path'],
socket.gethostname(),
g.task_index)
os.environ['THEANO_FLAGS'] = (
'device=gpu{},floatX=float32,base_compiledir={}'.format(
g.MY_GPU, base_compile_dir)) # ,mode=NanGuardMode'
# import theano
# import keras
for i in range(g.num_workers):
g.comm.Barrier()
def __call__(self, reference_observations: np.ndarray,
generated_observations: np.ndarray) -> float:
'''
Computes the FVD between the reference and the generated observations
:param reference_observations: (bs, observations_count, channels, height, width) tensor with reference observations
:param generated_observations: (bs, observations_count, channels, height, width) tensor with generated observations
:return: The FVD between the two distributions
'''
# Multiples of 16 must be fes to the embeddings network
embedding_batch_size = 16
# Puts the observations in the expected range [0, 255]
reference_observations = reference_observations * 255
generated_observations = generated_observations * 255
# Converts dimensions to tensorflow format by moving channels
reference_observations = np.moveaxis(reference_observations, 2, -1)
generated_observations = np.moveaxis(generated_observations, 2, -1)
# Cuts the sequences to multiples of the batch size
reference_observations = cut_to_multiple_size(reference_observations,
embedding_batch_size)
generated_observations = cut_to_multiple_size(generated_observations,
embedding_batch_size)
sequences, sequence_length, height, width, channels = reference_observations.shape
with tf.Graph().as_default():
# Builds the graph
input_placeholder = tf.placeholder(tf.float32, [
embedding_batch_size, sequence_length, height, width, channels
])
embeddings = create_id3_embedding(
preprocess(input_placeholder, (224, 224)))
reference_embeddings_placeholder = tf.placeholder(
tf.float32, [sequences, 400])
generated_embeddings_placeholder = tf.placeholder(
tf.float32, [sequences, 400])
fvd = calculate_fvd(reference_embeddings_placeholder,
generated_embeddings_placeholder)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.3)
with tf.Session(config=tf.ConfigProto(
gpu_options=gpu_options)) as sess:
sess.run(tf.global_variables_initializer())
sess.run(tf.tables_initializer())
# Computes the embeddings
reference_embeddings = extract_embeddings(
embeddings,
input_placeholder,
reference_observations,
sess,
batch_size=embedding_batch_size)
generated_embeddings = extract_embeddings(
embeddings,
input_placeholder,
generated_observations,
sess,
batch_size=embedding_batch_size)
# Computes the fvd
fvd_np = sess.run(fvd,
feed_dict={
reference_embeddings_placeholder:
reference_embeddings,
generated_embeddings_placeholder:
generated_embeddings
})
return float(fvd_np)
def train(self):
start_time = time.time()
# Config GPU options
if self.FLAGS.per_process_gpu_memory_fraction == 0.0:
gpu_options = tf.GPUOptions(allow_growth=True)
elif self.FLAGS.per_process_gpu_memory_fraction == 1.0:
gpu_options = tf.GPUOptions()
else:
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=self.
FLAGS.per_process_gpu_memory_fraction,
allow_growth=True)
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ['CUDA_VISIBLE_DEVICES'] = self.FLAGS.cuda_visible_devices
self.sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
if not tf.test.gpu_device_name():
self.logger.warning("No GPU is found")
else:
self.logger.info(tf.test.gpu_device_name())
global_step_lr = tf.Variable(0, trainable=False)
lr1 = tf.train.exponential_decay(
learning_rate=self.FLAGS.learning_rate,
global_step=global_step_lr,
decay_steps=1000,
decay_rate=0.995,
staircase=True)
lr2 = tf.train.exponential_decay(learning_rate=0.001,
global_step=global_step_lr,
decay_steps=1000,
decay_rate=0.995,
staircase=True)
with self.sess.as_default():
if self.FLAGS.experiment_type == "SR_GNN":
self.model = SR_GNN(self.FLAGS, self.emb, self.sess)
if self.FLAGS.experiment_type == "GC_SAN":
self.model = GC_SAN(self.FLAGS, self.emb, self.sess)
elif self.FLAGS.experiment_type == "GNN_T_Att":
self.model = GNN_T_Att(self.FLAGS, self.emb, self.sess)
elif self.FLAGS.experiment_type == "TimeAwareSR_GNN":
self.model = TimeAwareSR_GNN(self.FLAGS, self.emb, self.sess)
elif self.FLAGS.experiment_type == "GatedGrnnRec":
self.model = GatedGrnnRec(self.FLAGS, self.emb, self.sess)
elif self.FLAGS.experiment_type == "OrderedGatedGrnnRec":
self.model = OrderedGatedGrnnRec(self.FLAGS, self.emb,
self.sess)
elif self.FLAGS.experiment_type == "ModifiedGatedGrnnRec":
self.model = ModifiedGatedGrnnRec(self.FLAGS, self.emb,
self.sess)
elif self.FLAGS.experiment_type == "FGNN":
self.model = FGNN(self.FLAGS, self.emb, self.sess)
self.logger.info('Init finish.\tCost time: %.2fs' %
(time.time() - start_time))
#AUC暂时不看
# test_auc = self.model.metrics(sess=self.sess,
# batch_data=self.test_set,
# global_step=self.global_step,
# name='test auc')
# Eval init AUC
# self.logger.info('Init AUC: %.4f' % test_auc)
test_start = time.time()
self.hr_1, self.ndcg_1, self.hr_5, self.ndcg_5, self.hr_10, self.ndcg_10, self.hr_20, self.ndcg_20, self.hr_50, self.ndcg_50 = \
0,0,0,0,0,0,0,0,0,0
self.best_dev_hr_10 = 0
self.best_dev_ndcg_10 = 0
self.best_result_hr = []
self.best_result_ndcg = []
self.max_stay_count = 0
self.last_dev_hr_10 = 0
self.last_dev_ndcg_10 = 0
def eval_topk():
dev_sum_hr_1, dev_sum_ndcg_1, dev_sum_mrr_1, dev_sum_hr_5, dev_sum_ndcg_5, dev_sum_mrr_5, \
dev_sum_hr_10, dev_sum_ndcg_10, dev_sum_mrr_10, dev_sum_hr_20, dev_sum_ndcg_20, dev_sum_mrr_20, \
dev_sum_hr_50, dev_sum_ndcg_50, dev_sum_mrr_50 = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
result_list_hr_all = []
result_list_ndcg_all = []
max_step = 0
for step_i, batch_data in DataInput(
self.dev_set, self.FLAGS.test_batch_size):
max_step = 1 + max_step
if self.FLAGS.experiment_type == "NARM" or \
self.FLAGS.experiment_type == "NARM+" or \
self.FLAGS.experiment_type == "NARM++" or \
self.FLAGS.experiment_type == "SR_GNN":
hr_1, ndcg_1, mrr_1, hr_5, ndcg_5, mrr_5, \
hr_10, ndcg_10, mrr_10, hr_20, ndcg_20, mrr_20, \
hr_50, ndcg_50, mrr_50, \
result_list_hr, result_list_ndcg = \
self.model.metrics_topK_concat(sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
else:
hr_1, ndcg_1, mrr_1, hr_5, ndcg_5, mrr_5, \
hr_10, ndcg_10, mrr_10, hr_20, ndcg_20, mrr_20, \
hr_50, ndcg_50, mrr_50, \
result_list_hr, result_list_ndcg = \
self.model.metrics_topK(sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
dev_sum_hr_1 = dev_sum_hr_1 + hr_1
dev_sum_ndcg_1 = dev_sum_ndcg_1 + ndcg_1
dev_sum_mrr_1 = dev_sum_mrr_1 + mrr_1
dev_sum_hr_5 = dev_sum_hr_5 + hr_5
dev_sum_ndcg_5 = dev_sum_ndcg_5 + ndcg_5
dev_sum_mrr_5 = dev_sum_mrr_5 + mrr_5
dev_sum_hr_10 = dev_sum_hr_10 + hr_10
dev_sum_ndcg_10 = dev_sum_ndcg_10 + ndcg_10
dev_sum_mrr_10 = dev_sum_mrr_10 + mrr_10
dev_sum_hr_20 = dev_sum_hr_20 + hr_20
dev_sum_ndcg_20 = dev_sum_ndcg_20 + ndcg_20
dev_sum_mrr_20 = dev_sum_mrr_20 + mrr_20
dev_sum_hr_50 = dev_sum_hr_50 + hr_50
dev_sum_ndcg_50 = dev_sum_ndcg_50 + ndcg_50
dev_sum_mrr_50 = dev_sum_mrr_50 + mrr_50
dev_sum_hr_1 /= max_step
dev_sum_ndcg_1 /= max_step
dev_sum_mrr_1 /= max_step
dev_sum_hr_5 /= max_step
dev_sum_ndcg_5 /= max_step
dev_sum_mrr_5 /= max_step
dev_sum_hr_10 /= max_step
dev_sum_ndcg_10 /= max_step
dev_sum_mrr_10 /= max_step
dev_sum_hr_20 /= max_step
dev_sum_ndcg_20 /= max_step
dev_sum_mrr_20 /= max_step
dev_sum_hr_50 /= max_step
dev_sum_ndcg_50 /= max_step
dev_sum_mrr_50 /= max_step
sum_hr_1, sum_ndcg_1, sum_mrr_1, sum_hr_5, sum_ndcg_5, sum_mrr_5, \
sum_hr_10, sum_ndcg_10, sum_mrr_10, sum_hr_20, sum_ndcg_20, sum_mrr_20, \
sum_hr_50, sum_ndcg_50, sum_mrr_50 = 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
result_list_hr_all = []
result_list_ndcg_all = []
max_step = 0
for step_i, batch_data in DataInput(
self.test_set, self.FLAGS.test_batch_size):
max_step = 1 + max_step
if self.FLAGS.experiment_type == "NARM" or \
self.FLAGS.experiment_type == "NARM+" or \
self.FLAGS.experiment_type == "NARM++" or \
self.FLAGS.experiment_type == "SR_GNN":
hr_1, ndcg_1, mrr_1, hr_5, ndcg_5, mrr_5, \
hr_10, ndcg_10, mrr_10, hr_20, ndcg_20, mrr_20, \
hr_50, ndcg_50, mrr_50, \
result_list_hr, result_list_ndcg = \
self.model.metrics_topK_concat(sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
else:
hr_1, ndcg_1, mrr_1, hr_5, ndcg_5, mrr_5, \
hr_10, ndcg_10, mrr_10, hr_20, ndcg_20, mrr_20, \
hr_50, ndcg_50, mrr_50, \
result_list_hr, result_list_ndcg = \
self.model.metrics_topK(sess=self.sess,
batch_data=batch_data,
global_step=self.global_step,
topk=self.FLAGS.top_k)
sum_hr_1 = sum_hr_1 + hr_1
sum_ndcg_1 = sum_ndcg_1 + ndcg_1
sum_mrr_1 = sum_mrr_1 + mrr_1
sum_hr_5 = sum_hr_5 + hr_5
sum_ndcg_5 = sum_ndcg_5 + ndcg_5
sum_mrr_5 = sum_mrr_5 + mrr_5
sum_hr_10 = sum_hr_10 + hr_10
sum_ndcg_10 = sum_ndcg_10 + ndcg_10
sum_mrr_10 = sum_mrr_10 + mrr_10
sum_hr_20 = sum_hr_20 + hr_20
sum_ndcg_20 = sum_ndcg_20 + ndcg_20
sum_mrr_20 = sum_mrr_20 + mrr_20
sum_hr_50 = sum_hr_50 + hr_50
sum_ndcg_50 = sum_ndcg_50 + ndcg_50
sum_mrr_50 = sum_mrr_50 + mrr_50
result_list_hr_all = result_list_hr_all + result_list_hr
result_list_ndcg_all = result_list_ndcg_all + result_list_ndcg
sum_hr_1 /= max_step
sum_ndcg_1 /= max_step
sum_mrr_1 /= max_step
sum_hr_5 /= max_step
sum_ndcg_5 /= max_step
sum_mrr_5 /= max_step
sum_hr_10 /= max_step
sum_ndcg_10 /= max_step
sum_mrr_10 /= max_step
sum_hr_20 /= max_step
sum_ndcg_20 /= max_step
sum_mrr_20 /= max_step
sum_hr_50 /= max_step
sum_ndcg_50 /= max_step
sum_mrr_50 /= max_step
if dev_sum_hr_10 > self.best_dev_hr_10 and dev_sum_ndcg_10 > self.best_dev_ndcg_10:
self.best_dev_hr_10 = dev_sum_hr_10
self.best_dev_ndcg_10 = dev_sum_ndcg_10
self.hr_1, self.ndcg_1, self.mrr_1 = sum_hr_1, sum_ndcg_1, sum_mrr_1
self.hr_5, self.ndcg_5, self.mrr_5 = sum_hr_5, sum_ndcg_5, sum_mrr_5
self.hr_10, self.ndcg_10, self.mrr_10 = sum_hr_10, sum_ndcg_10, sum_mrr_10
self.best_result_hr = result_list_hr_all
self.best_result_ndcg = result_list_ndcg_all
self.hr_20, self.ndcg_20, self.mrr_20 = sum_hr_20, sum_ndcg_20, sum_mrr_20
self.hr_50, self.ndcg_50, self.mrr_50 = sum_hr_50, sum_ndcg_50, sum_mrr_50
def dev_log(k, hr, ndcg, mrr):
self.logger.info(
'Dev recall rate @ %d : %.4f ndcg @ %d: %.4f' %
(k, hr, k, ndcg))
dev_log(1, dev_sum_hr_1, dev_sum_ndcg_1, dev_sum_mrr_1)
dev_log(5, dev_sum_hr_5, dev_sum_ndcg_5, dev_sum_mrr_5)
dev_log(10, dev_sum_hr_10, dev_sum_ndcg_10, dev_sum_mrr_10)
dev_log(20, dev_sum_hr_20, dev_sum_ndcg_20, dev_sum_mrr_20)
dev_log(50, dev_sum_hr_50, dev_sum_ndcg_50, dev_sum_mrr_50)
def summery(k, hr, ndcg, mrr):
tag_recall = 'test recall@' + str(k)
tag_ndcg = 'test ndgc@' + str(k)
summary_recall_rate = tf.Summary(value=[
tf.Summary.Value(tag=tag_recall, simple_value=hr)
])
self.model.train_writer.add_summary(
summary_recall_rate, global_step=self.global_step)
summary_avg_ndcg = tf.Summary(value=[
tf.Summary.Value(tag=tag_ndcg, simple_value=ndcg)
])
self.model.train_writer.add_summary(
summary_avg_ndcg, global_step=self.global_step)
self.logger.info(
'Test recall rate @ %d : %.4f ndcg @ %d: %.4f mrr @ %d: %.4f'
% (k, hr, k, ndcg, k, mrr))
summery(1, sum_hr_1, sum_ndcg_1, sum_mrr_1)
summery(5, sum_hr_5, sum_ndcg_5, sum_mrr_5)
summery(10, sum_hr_10, sum_ndcg_10, sum_mrr_10)
summery(20, sum_hr_20, sum_ndcg_20, sum_mrr_20)
summery(50, sum_hr_50, sum_ndcg_50, sum_mrr_50)
eval_topk()
self.logger.info('End test. \tTest Cost time: %.2fs' %
(time.time() - test_start))
# Start training
self.logger.info(
'Training....\tmax_epochs:%d\tepoch_size:%d' %
(self.FLAGS.max_epochs, self.FLAGS.train_batch_size))
start_time, avg_loss, self.best_auc, self.best_recall, self.best_ndcg = time.time(
), 0.0, 0.0, 0.0, 0.0
for epoch in range(self.FLAGS.max_epochs):
#if epoch > 2:
#lr = lr/1.5
random.shuffle(self.train_set)
self.logger.info('tain_set:%d' % len(self.train_set))
epoch_start_time = time.time()
learning_rate = self.FLAGS.learning_rate
for step_i, train_batch_data in DataInput(
self.train_set, self.FLAGS.train_batch_size):
# try:
#print(self.sess.run(global_step_lr))
if learning_rate > 0.001:
learning_rate = self.sess.run(
lr1, feed_dict={global_step_lr: self.global_step})
else:
learning_rate = self.sess.run(
lr2, feed_dict={global_step_lr: self.global_step})
#print(learning_rate)
add_summary = bool(self.global_step %
self.FLAGS.display_freq == 0)
step_loss, merge = self.model.train(
self.sess, train_batch_data, learning_rate,
add_summary, self.global_step, epoch)
self.sess.graph.finalize()
self.model.train_writer.add_summary(
merge, self.global_step)
avg_loss = avg_loss + step_loss
self.global_step = self.global_step + 1
self.one_epoch_step = self.one_epoch_step + 1
#evaluate for eval steps
if self.global_step % self.FLAGS.eval_freq == 0:
print(learning_rate)
self.logger.info("Epoch step is " +
str(self.one_epoch_step))
self.logger.info("Global step is " +
str(self.global_step))
self.logger.info("Train_loss is " +
str(avg_loss / self.FLAGS.eval_freq))
# train_auc = self.model.metrics(sess=self.sess, batch_data=train_batch_data,
# global_step=self.global_step,name='train auc')
# self.logger.info('Batch Train AUC: %.4f' % train_auc)
# self.test_auc = self.model.metrics(sess=self.sess, batch_data=self.test_set,
# global_step=self.global_step,name='test auc')
# self.logger.info('Test AUC: %.4f' % self.test_auc)
eval_topk()
avg_loss = 0
self.save_model()
if self.FLAGS.draw_pic == True:
self.save_fig()
# except Exception as e:
# self.logger.info("Error！！！！！！！！！！！！")
# self.logger.info(e)
# traceback.print_exc()
self.logger.info('one epoch Cost time: %.2f' %
(time.time() - epoch_start_time))
self.logger.info("Epoch step is " + str(self.one_epoch_step))
self.logger.info("Global step is " + str(self.global_step))
self.logger.info("Train_loss is " + str(step_loss))
eval_topk()
with open('best_result_hr_' + self.FLAGS.version, 'w+') as f:
f.write(str(self.best_result_hr))
with open('best_result_ndcg' + self.FLAGS.version, 'w+') as f:
f.write(str(self.best_result_ndcg))
self.logger.info('Max recall rate @ 1: %.4f ndcg @ 1: %.4f' %
(self.hr_1, self.ndcg_1))
self.logger.info('Max recall rate @ 5: %.4f ndcg @ 5: %.4f' %
(self.hr_5, self.ndcg_5))
self.logger.info(
'Max recall rate @ 10: %.4f ndcg @ 10: %.4f' %
(self.hr_10, self.ndcg_10))
self.logger.info(
'Max recall rate @ 20: %.4f ndcg @ 20: %.4f' %
(self.hr_20, self.ndcg_20))
self.logger.info(
'Max recall rate @ 50: %.4f ndcg @ 50: %.4f' %
(self.hr_50, self.ndcg_50))
if self.best_dev_hr_10 == self.last_dev_hr_10 and self.best_dev_ndcg_10 == self.last_dev_ndcg_10:
self.max_stay_count += 1
else:
self.last_dev_hr_10 = self.best_dev_hr_10
self.last_dev_ndcg_10 = self.best_dev_ndcg_10
self.max_stay_count = 0
if self.max_stay_count > 5:
break
self.one_epoch_step = 0
#if self.global_step > 1000:
#lr = lr / 2
#if lr < 0.0005:
#lr = lr * 0.99
#elif self.FLAGS.type == "tmall":
#lr = lr * 0.5
#else:
#lr = lr * 0.98
self.logger.info('Epoch %d DONE\tCost time: %.2f' %
(self.now_epoch, time.time() - start_time))
self.now_epoch = self.now_epoch + 1
self.one_epoch_step = 0
# self.model.save(self.sess,self.global_step)
self.logger.info('best test_auc: ' + str(self.best_auc))
self.logger.info('best recall: ' + str(self.best_recall))
self.logger.info('Finished')
def main():
tf.disable_eager_execution()
workspace_path = os.environ.get('AE_WORKSPACE_PATH')
if workspace_path is None:
print('Please define a workspace path:\n')
print('export AE_WORKSPACE_PATH=/path/to/workspace\n')
exit(-1)
gentle_stop = np.array((1,), dtype=np.bool)
gentle_stop[0] = False
def on_ctrl_c(signal, frame):
gentle_stop[0] = True
signal.signal(signal.SIGINT, on_ctrl_c)
parser = argparse.ArgumentParser()
parser.add_argument("experiment_name")
parser.add_argument("-d", action='store_true', default=False)
parser.add_argument("-gen", action='store_true', default=False)
arguments = parser.parse_args()
full_name = arguments.experiment_name.split('/')
experiment_name = full_name.pop()
experiment_group = full_name.pop() if len(full_name) > 0 else ''
debug_mode = arguments.d
generate_data = arguments.gen
cfg_file_path = u.get_config_file_path(workspace_path, experiment_name, experiment_group)
log_dir = u.get_log_dir(workspace_path, experiment_name, experiment_group)
checkpoint_file = u.get_checkpoint_basefilename(log_dir)
ckpt_dir = u.get_checkpoint_dir(log_dir)
train_fig_dir = u.get_train_fig_dir(log_dir)
dataset_path = u.get_dataset_path(workspace_path)
if not os.path.exists(cfg_file_path):
print('Could not find config file:\n')
print('{}\n'.format(cfg_file_path))
exit(-1)
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
if not os.path.exists(train_fig_dir):
os.makedirs(train_fig_dir)
if not os.path.exists(dataset_path):
os.makedirs(dataset_path)
args = configparser.ConfigParser()
args.read(cfg_file_path)
shutil.copy2(cfg_file_path, log_dir)
with tf.variable_scope(experiment_name):
dataset = factory.build_dataset(dataset_path, args)
queue = factory.build_queue(dataset, args)
encoder = factory.build_encoder(queue.x, args, is_training=True)
decoder = factory.build_decoder(queue.y, encoder, args, is_training=True)
ae = factory.build_ae(encoder, decoder, args)
codebook = factory.build_codebook(encoder, dataset, args)
train_op = factory.build_train_op(ae, args)
saver = tf.train.Saver(save_relative_paths=True)
num_iter = args.getint('Training', 'NUM_ITER') if not debug_mode else 100000
save_interval = args.getint('Training', 'SAVE_INTERVAL')
model_type = args.get('Dataset', 'MODEL')
if model_type=='dsprites':
dataset.get_sprite_training_images(args)
else:
dataset.get_training_images(dataset_path, args)
dataset.load_bg_images(dataset_path)
if generate_data:
print('finished generating synthetic training data for ' + experiment_name)
print('exiting...')
exit()
widgets = ['Training: ', progressbar.Percentage(),
' ', progressbar.Bar(),
' ', progressbar.Counter(), ' / %s' % num_iter,
' ', progressbar.ETA(), ' ']
bar = progressbar.ProgressBar(maxval=num_iter,widgets=widgets)
gpu_options = tf.GPUOptions(allow_growth=True, per_process_gpu_memory_fraction=0.8)
config = tf.ConfigProto(gpu_options=gpu_options)
with tf.Session(config=config) as sess:
chkpt = tf.train.get_checkpoint_state(ckpt_dir)
if chkpt and chkpt.model_checkpoint_path:
saver.restore(sess, chkpt.model_checkpoint_path)
else:
sess.run(tf.global_variables_initializer())
merged_loss_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(ckpt_dir, sess.graph)
if not debug_mode:
print('Training with %s model' % args.get('Dataset','MODEL'), os.path.basename(args.get('Paths','MODEL_PATH')))
bar.start()
queue.start(sess)
for i in range(ae.global_step.eval(), num_iter):
if not debug_mode:
sess.run(train_op)
if i % 10 == 0:
loss = sess.run(merged_loss_summary)
summary_writer.add_summary(loss, i)
bar.update(i)
if (i+1) % save_interval == 0:
saver.save(sess, checkpoint_file, global_step=ae.global_step)
this_x, this_y = sess.run([queue.x, queue.y])
reconstr_train = sess.run(decoder.x,feed_dict={queue.x:this_x})
train_imgs = np.hstack(( u.tiles(this_x, 4, 4), u.tiles(reconstr_train, 4,4),u.tiles(this_y, 4, 4)))
cv2.imwrite(os.path.join(train_fig_dir,'training_images_%s.png' % i), train_imgs*255)
else:
this_x, this_y = sess.run([queue.x, queue.y])
reconstr_train = sess.run(decoder.x,feed_dict={queue.x:this_x})
cv2.imshow('sample batch', np.hstack(( u.tiles(this_x, 3, 3), u.tiles(reconstr_train, 3,3),u.tiles(this_y, 3, 3))) )
k = cv2.waitKey(0)
if k == 27:
break
if gentle_stop[0]:
break
queue.stop(sess)
if not debug_mode:
bar.finish()
if not gentle_stop[0] and not debug_mode:
print('To create the embedding run:\n')
print('ae_embed {}\n'.format(full_name))
def __init__(self,
N,
rw_len,
walk_generator,
generator_layers=[40],
discriminator_layers=[30],
W_down_generator_size=128,
W_down_discriminator_size=128,
batch_size=128,
noise_dim=16,
noise_type="Gaussian",
learning_rate=0.0003,
disc_iters=3,
wasserstein_penalty=10,
l2_penalty_generator=1e-7,
l2_penalty_discriminator=5e-5,
temp_start=5.0,
min_temperature=0.5,
temperature_decay=1 - 5e-5,
seed=15,
gpu_id=0,
use_gumbel=True,
legacy_generator=False):
"""
Initialize NetGAN.
Parameters
----------
N: int
Number of nodes in the graph to generate.
rw_len: int
Length of random walks to generate.
walk_generator: function
Function that generates a single random walk and takes no arguments.
generator_layers: list of integers, default: [40], i.e. a single layer with 40 units.
The layer sizes of the generator LSTM layers
discriminator_layers: list of integers, default: [30], i.e. a single layer with 30 units.
The sizes of the discriminator LSTM layers
W_down_generator_size: int, default: 128
The size of the weight matrix W_down of the generator. See our paper for details.
W_down_discriminator_size: int, default: 128
The size of the weight matrix W_down of the discriminator. See our paper for details.
batch_size: int, default: 128
The batch size.
noise_dim: int, default: 16
The dimension of the random noise that is used as input to the generator.
noise_type: str in ["Gaussian", "Uniform], default: "Gaussian"
The noise type to feed into the generator.
learning_rate: float, default: 0.0003
The learning rate.
disc_iters: int, default: 3
The number of discriminator iterations per generator training iteration.
wasserstein_penalty: float, default: 10
The Wasserstein gradient penalty applied to the discriminator. See the Wasserstein GAN
paper for details.
l2_penalty_generator: float, default: 1e-7
L2 penalty on the generator weights.
l2_penalty_discriminator: float, default: 5e-5
L2 penalty on the discriminator weights.
temp_start: float, default: 5.0
The initial temperature for the Gumbel softmax.
min_temperature: float, default: 0.5
The minimal temperature for the Gumbel softmax.
temperature_decay: float, default: 1-5e-5
After each evaluation, the current temperature is updated as
current_temp := max(temperature_decay*current_temp, min_temperature)
seed: int, default: 15
Random seed.
gpu_id: int or None, default: 0
The ID of the GPU to be used for training. If None, CPU only.
use_gumbel: bool, default: True
Use the Gumbel softmax trick.
legacy_generator: bool, default: False
If True, the hidden and cell states of the generator LSTM are initialized by two separate feed-forward networks.
If False (recommended), the hidden layer is shared, which has less parameters and performs just as good.
"""
self.params = {
'noise_dim': noise_dim,
'noise_type': noise_type,
'Generator_Layers': generator_layers,
'Discriminator_Layers': discriminator_layers,
'W_Down_Generator_size': W_down_generator_size,
'W_Down_Discriminator_size': W_down_discriminator_size,
'l2_penalty_generator': l2_penalty_generator,
'l2_penalty_discriminator': l2_penalty_discriminator,
'learning_rate': learning_rate,
'batch_size': batch_size,
'Wasserstein_penalty': wasserstein_penalty,
'temp_start': temp_start,
'min_temperature': min_temperature,
'temperature_decay': temperature_decay,
'disc_iters': disc_iters,
'use_gumbel': use_gumbel,
'legacy_generator': legacy_generator
}
assert rw_len > 1, "Random walk length must be > 1."
tf.set_random_seed(seed)
#tf.random.set_seed(seed)
self.N = N
self.rw_len = rw_len
self.noise_dim = self.params['noise_dim']
self.G_layers = self.params['Generator_Layers']
self.D_layers = self.params['Discriminator_Layers']
self.tau = tf.placeholder(1.0, shape=(), name="temperature")
# W_down and W_up for generator and discriminator
self.W_down_generator = tf.get_variable(
'Generator.W_Down',
shape=[self.N, self.params['W_Down_Generator_size']],
dtype=tf.float32,
initializer=tf.glorot_uniform_initializer())
self.W_down_discriminator = tf.get_variable(
'Discriminator.W_Down',
shape=[self.N, self.params['W_Down_Discriminator_size']],
dtype=tf.float32,
initializer=tf.glorot_uniform_initializer())
self.W_up = tf.get_variable(
"Generator.W_up",
shape=[self.G_layers[-1], self.N],
dtype=tf.float32,
initializer=tf.glorot_uniform_initializer())
self.b_W_up = tf.get_variable("Generator.W_up_bias",
dtype=tf.float32,
initializer=tf.zeros_initializer,
shape=self.N)
self.generator_function = self.generator_recurrent
self.discriminator_function = self.discriminator_recurrent
self.fake_inputs = self.generator_function(self.params['batch_size'],
reuse=False,
gumbel=use_gumbel,
legacy=legacy_generator)
self.fake_inputs_discrete = self.generate_discrete(
self.params['batch_size'],
reuse=True,
gumbel=use_gumbel,
legacy=legacy_generator)
# Pre-fetch real random walks
dataset = tf.data.Dataset.from_generator(
walk_generator, tf.int32, [self.params['batch_size'], self.rw_len])
#dataset_batch = dataset.prefetch(2).batch(self.params['batch_size'])
dataset_batch = dataset.prefetch(100)
batch_iterator = dataset_batch.make_one_shot_iterator()
real_data = batch_iterator.get_next()
self.real_inputs_discrete = real_data
self.real_inputs = tf.one_hot(self.real_inputs_discrete, self.N)
self.disc_real = self.discriminator_function(self.real_inputs)
self.disc_fake = self.discriminator_function(self.fake_inputs,
reuse=True)
self.disc_cost = tf.reduce_mean(self.disc_fake) - tf.reduce_mean(
self.disc_real)
self.gen_cost = -tf.reduce_mean(self.disc_fake)
# WGAN lipschitz-penalty
alpha = tf.random_uniform(shape=[self.params['batch_size'], 1, 1],
minval=0.,
maxval=1.)
self.differences = self.fake_inputs - self.real_inputs
self.interpolates = self.real_inputs + (alpha * self.differences)
self.gradients = tf.gradients(
self.discriminator_function(self.interpolates, reuse=True),
self.interpolates)[0]
self.slopes = tf.sqrt(
tf.reduce_sum(tf.square(self.gradients), reduction_indices=[1, 2]))
self.gradient_penalty = tf.reduce_mean((self.slopes - 1.)**2)
self.disc_cost += self.params[
'Wasserstein_penalty'] * self.gradient_penalty
# weight regularization; we omit W_down from regularization
self.disc_l2_loss = tf.add_n([
tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'Disc' in v.name and not 'W_down' in v.name
]) * self.params['l2_penalty_discriminator']
self.disc_cost += self.disc_l2_loss
# weight regularization; we omit W_down from regularization
self.gen_l2_loss = tf.add_n([
tf.nn.l2_loss(v) for v in tf.trainable_variables()
if 'Gen' in v.name and not 'W_down' in v.name
]) * self.params['l2_penalty_generator']
self.gen_cost += self.gen_l2_loss
self.gen_params = [
v for v in tf.trainable_variables() if 'Generator' in v.name
]
self.disc_params = [
v for v in tf.trainable_variables() if 'Discriminator' in v.name
]
self.gen_train_op = tf.train.AdamOptimizer(
learning_rate=self.params['learning_rate'], beta1=0.5,
beta2=0.9).minimize(self.gen_cost, var_list=self.gen_params)
self.disc_train_op = tf.train.AdamOptimizer(
learning_rate=self.params['learning_rate'], beta1=0.5,
beta2=0.9).minimize(self.disc_cost, var_list=self.disc_params)
if gpu_id is None:
config = tf.ConfigProto(device_count={'GPU': 0})
else:
gpu_options = tf.GPUOptions(
visible_device_list='{}'.format(gpu_id), allow_growth=True)
config = tf.ConfigProto(gpu_options=gpu_options)
self.session = tf.InteractiveSession(config=config)
self.init_op = tf.global_variables_initializer()