def main():
args = parser.parse_args()
# We store all arguments in a json file. This has two advantages:
# 1. We can always get back and see what exactly that experiment was
# 2. We can resume an experiment as-is without needing to remember all flags.
args_file = os.path.join(args.experiment_root, 'args.json')
if args.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
for key, value in args.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
comand = input('Would you like to restore it?(yes/no)')
if comand == 'yes':
args.__dict__[key] = resumed_value
print(
'For the argument `{}` we are using the loaded value `{}`.'
.format(key, args.__dict__[key]))
else:
print(
'For the argument `{}` we are using the provided value `{}`.'
.format(key, args.__dict__[key]))
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
os.remove(args_file)
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
else:
# If the experiment directory exists already, we bail in fear.
if os.path.exists(args.experiment_root):
if os.listdir(args.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(args.experiment_root))
exit(1)
else:
os.makedirs(args.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
log_file = os.path.join(args.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.
if not args.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not args.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
######################################################################################
#prepare the training dataset
# Load the data from the TxT file. see Common.load_dataset function for details
pids_train, fids_train = common.load_dataset(args.train_set,
args.image_root)
max_fid_len = max(map(len,
fids_train)) # We'll need this later for logfiles.
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids = np.unique(pids_train)
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
dataset = dataset.take((len(unique_pids) // args.batch_p) * args.batch_p)
dataset = dataset.repeat(None) # Repeat forever. Funny way of stating it.
# For every PID, get K images.
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids_train, all_pids=pids_train, batch_k=args.batch_k
)) # now the dataset has been modified as [selected_fids
# , pid] due to the return of the function 'sample_k_fids_for_pid'
# Ungroup/flatten the batches for easy loading of the files.
dataset = dataset.apply(tf.contrib.data.unbatch())
# Convert filenames to actual image tensors.
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset = dataset.map(
lambda fid, pid: common.fid_to_image(fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size if args.
crop_augment else net_input_size),
num_parallel_calls=args.loading_threads
) # now the dataset has been modified as [selected_images
# , fid， pid] due to the return of the function 'fid_to_image'
# Augment the data if specified by the arguments.
if args.flip_augment:
dataset = dataset.map(lambda im, fid, pid:
(tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset = dataset.map(lambda im, fid, pid: (tf.random_crop(
im, net_input_size + (3, )), fid, pid))
# Group it back into PK batches.
batch_size = args.batch_p * args.batch_k
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(1)
# Since we repeat the data infinitely, we only need a one-shot iterator.
images_train, fids_train, pids_train = dataset.make_one_shot_iterator(
).get_next()
########################################################################################################################
#prepare the validation set
pids_val, fids_val = common.load_dataset(args.validation_set,
args.validation_image_root)
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids_val = np.unique(pids_val)
dataset_val = tf.data.Dataset.from_tensor_slices(unique_pids_val)
dataset_val = dataset_val.shuffle(len(unique_pids_val))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
dataset_val = dataset_val.take(
(len(unique_pids_val) // args.batch_p) * args.batch_p)
dataset_val = dataset_val.repeat(
None) # Repeat forever. Funny way of stating it.
# For every PID, get K images.
dataset_val = dataset_val.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids_val, all_pids=pids_val, batch_k=args.batch_k
)) # now the dataset has been modified as [selected_fids
# , pid] due to the return of the function 'sample_k_fids_for_pid'
# Ungroup/flatten the batches for easy loading of the files.
dataset_val = dataset_val.apply(tf.contrib.data.unbatch())
# Convert filenames to actual image tensors.
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset_val = dataset_val.map(
lambda fid, pid: common.fid_to_image(
fid,
pid,
image_root=args.validation_image_root,
image_size=pre_crop_size if args.crop_augment else net_input_size),
num_parallel_calls=args.loading_threads
) # now the dataset has been modified as [selected_images
# , fid， pid] due to the return of the function 'fid_to_image'
# Augment the data if specified by the arguments.
if args.flip_augment:
dataset_val = dataset_val.map(lambda im, fid, pid: (
tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset_val = dataset_val.map(lambda im, fid, pid: (tf.random_crop(
im, net_input_size + (3, )), fid, pid))
# Group it back into PK batches.
dataset_val = dataset_val.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset_val = dataset_val.prefetch(1)
# Since we repeat the data infinitely, we only need a one-shot iterator.
images_val, fids_val, pids_val = dataset_val.make_one_shot_iterator(
).get_next()
####################################################################################################################
# Create the model and an embedding head.
model = import_module('nets.' + args.model_name)
head = import_module('heads.' + args.head_name)
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
input_images = tf.placeholder(
dtype=tf.float32,
shape=[None, args.net_input_height, args.net_input_width, 3],
name='input')
pids = tf.placeholder(dtype=tf.string, shape=[
None,
], name='pids')
fids = tf.placeholder(dtype=tf.string, shape=[
None,
], name='fids')
endpoints, body_prefix = model.endpoints(input_images, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, args.embedding_dim, is_training=True)
# Create the loss in two steps:
# 1. Compute all pairwise distances according to the specified metric.
# 2. For each anchor along the first dimension, compute its loss.
# dists = loss.cdist(endpoints['emb'], endpoints['emb'], metric=args.metric)
# losses, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[args.loss](
# dists, pids, args.margin, batch_precision_at_k=args.batch_k-1)
# # '_' stands for the boolean matrix shows topK where the correct match of the identities occurs
# shape=(batch_size,K)
# 更改
# loss1
dists1 = loss.cdist(endpoints['feature1'],
endpoints['feature1'],
metric=args.metric)
losses1, _, _, _, _, _ = loss.LOSS_CHOICES[args.loss](
dists1, pids, args.margin, batch_precision_at_k=args.batch_k - 1)
dists2 = loss.cdist(endpoints['feature2'],
endpoints['feature2'],
metric=args.metric)
losses2, _, _, _, _, _ = loss.LOSS_CHOICES[args.loss](
dists2, pids, args.margin, batch_precision_at_k=args.batch_k - 1)
dists3 = loss.cdist(endpoints['feature3'],
endpoints['feature3'],
metric=args.metric)
losses3, _, _, _, _, _ = loss.LOSS_CHOICES[args.loss](
dists3, pids, args.margin, batch_precision_at_k=args.batch_k - 1)
dists4 = loss.cdist(endpoints['feature4'],
endpoints['feature4'],
metric=args.metric)
losses4, _, _, _, _, _ = loss.LOSS_CHOICES[args.loss](
dists4, pids, args.margin, batch_precision_at_k=args.batch_k - 1)
dists_fu = loss.cdist(endpoints['fusion_layer'],
endpoints['fusion_layer'],
metric=args.metric)
losses_fu, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[
args.loss](dists_fu,
pids,
args.margin,
batch_precision_at_k=args.batch_k - 1)
losses = losses1 + losses2 + losses3 + losses4 + losses_fu
# 更改
#loss
# losses_fu, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[args.loss](
# endpoints,pids, model_type=args.model_name, metric=args.metric, batch_precision_at_k=args.batch_k - 1
# )
# Count the number of active entries, and compute the total batch loss.
num_active = tf.reduce_sum(tf.cast(tf.greater(losses, 1e-5), tf.float32))
# 此处losses即为 pospair 比 negpair+margin 还大的部分
loss_mean = tf.reduce_mean(losses)
# Some logging for tensorboard.
tf.summary.histogram('loss_distribution', losses)
tf.summary.scalar('loss', loss_mean)
tf.summary.scalar('batch_top1', train_top1)
tf.summary.scalar('batch_prec_at_{}'.format(args.batch_k - 1), prec_at_k)
tf.summary.scalar('active_count', num_active)
#tf.summary.histogram('embedding_dists', dists)
tf.summary.histogram('embedding_pos_dists', pos_dists)
tf.summary.histogram('embedding_neg_dists', neg_dists)
tf.summary.histogram('embedding_lengths',
tf.norm(endpoints['emb_raw'], axis=1))
# Create the mem-mapped arrays in which we'll log all training detail in
# addition to tensorboard, because tensorboard is annoying for detailed
# inspection and actually discards data in histogram summaries.
if args.detailed_logs:
log_embs = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'embeddings'),
dtype=np.float32,
shape=(args.train_iterations, batch_size, args.embedding_dim))
log_loss = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'losses'),
dtype=np.float32,
shape=(args.train_iterations, batch_size))
log_fids = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'fids'),
dtype='S' + str(max_fid_len),
shape=(args.train_iterations, batch_size))
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
global_step = tf.Variable(
0, name='global_step',
trainable=False) # 'global_step' means the number of batches seen
# by graph
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.train.exponential_decay(
args.learning_rate,
tf.maximum(0, global_step - args.decay_start_iteration
), # decay every 'lr_decay_steps' after the
# 'decay_start_iteration'
# args.train_iterations - args.decay_start_iteration, args.weight_decay_factor)
args.lr_decay_steps,
args.lr_decay_factor,
staircase=True)
else:
learning_rate = args.learning_rate # the case when we set 'decay_start_iteration' as -1
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate, epsilon=1e-3)
# Feel free to try others!
# optimizer = tf.train.AdadeltaOptimizer(learning_rate)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.minimize(loss_mean, global_step=global_step)
# Define a saver for the complete model.
checkpoint_saver = tf.train.Saver(max_to_keep=0)
with tf.Session(config=config) as sess:
if args.resume:
# In case we're resuming, simply load the full checkpoint to init.
if args.checkpoint is None:
last_checkpoint = tf.train.latest_checkpoint(
args.experiment_root)
log.info(
'Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
ckpt_path = os.path.join(args.experiment_root, args.checkpoint)
log.info('Restoring from checkpoint: {}'.format(
args.checkpoint))
checkpoint_saver.restore(sess, ckpt_path)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(
sess, args.initial_checkpoint
) # restore the pre-trained parameter from online model
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.experiment_root,
sess.graph)
start_step = sess.run(global_step)
log.info('Starting training from iteration {}.'.format(start_step))
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, args.train_iterations):
# Compute gradients, update weights, store logs!
start_time = time.time()
_, summary, step, b_prec_at_k, b_embs, b_loss, b_fids = \
sess.run([train_op, merged_summary, global_step,
prec_at_k, endpoints['emb'], losses, fids], feed_dict={input_images:images_train.eval(),
pids:pids_train.eval(),
fids:fids_train.eval()})
elapsed_time = time.time() - start_time
# Compute the iteration speed and add it to the summary.
# We did observe some weird spikes that we couldn't track down.
summary2 = tf.Summary()
summary2.value.add(tag='secs_per_iter',
simple_value=elapsed_time)
summary_writer.add_summary(summary2, step)
summary_writer.add_summary(summary, step)
if args.detailed_logs:
log_embs[i], log_loss[i], log_fids[
i] = b_embs, b_loss, b_fids
# Do a huge print out of the current progress.
seconds_todo = (args.train_iterations - step) * elapsed_time
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'batch-p@{}: {:.2%}, ETA: {} ({:.2f}s/it)'.format(
step, float(np.min(b_loss)), float(np.mean(b_loss)),
float(np.max(b_loss)), args.batch_k - 1,
float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)), elapsed_time))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.
if (args.checkpoint_frequency > 0
and step % args.checkpoint_frequency == 0):
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=step)
#get validation results
if (args.validation_frequency > 0
and step % args.validation_frequency == 0):
b_prec_at_k_val, b_loss, b_fids = \
sess.run([prec_at_k, losses, fids], feed_dict={input_images : images_val.eval(),
pids:pids_val.eval(),
fids:fids_val.eval()})
log.info(
'Validation @:{:6d} iteration, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'batch-p@{}: {:.2%}'.format(step,
float(np.min(b_loss)),
float(np.mean(b_loss)),
float(np.max(b_loss)),
args.batch_k - 1,
float(b_prec_at_k_val)))
sys.stdout.flush()
sys.stderr.flush()
summary3 = tf.Summary()
summary3.value.add(tag='secs_per_iter',
simple_value=float(np.mean(b_loss)))
summary_writer.add_summary(summary3, step)
summary_writer.add_summary(summary3, step)
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root, 'checkpoint'),
global_step=step)
def main():
args = parser.parse_args()
# Data augmentation
global seq_geo
global seq_img
seq_geo = iaa.SomeOf(
(0, 5),
[
iaa.Fliplr(0.5), # horizontally flip 50% of the images
iaa.PerspectiveTransform(scale=(0, 0.075)),
iaa.Affine(
scale={
"x": (0.8, 1.0),
"y": (0.8, 1.0)
},
rotate=(-5, 5),
translate_percent={
"x": (-0.1, 0.1),
"y": (-0.1, 0.1)
},
), # rotate by -45 to +45 degrees),
iaa.Crop(pc=(
0, 0.125
)), # crop images from each side by 0 to 12.5% (randomly chosen)
iaa.CoarsePepper(p=0.01, size_percent=0.1)
],
random_order=False)
# Content transformation
seq_img = iaa.SomeOf(
(0, 3),
[
iaa.GaussianBlur(
sigma=(0, 1.0)), # blur images with a sigma of 0 to 2.0
iaa.ContrastNormalization(alpha=(0.9, 1.1)),
iaa.Grayscale(alpha=(0, 0.2)),
iaa.Multiply((0.9, 1.1))
])
# We store all arguments in a json file. This has two advantages:
# 1. We can always get back and see what exactly that experiment was
# 2. We can resume an experiment as-is without needing to remember all flags.
args_file = os.path.join(args.experiment_root, 'args.json')
if args.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
for key, value in args.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
args.__dict__[key] = resumed_value
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
else:
# If the experiment directory exists already, we bail in fear.
if os.path.exists(args.experiment_root):
if os.listdir(args.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(args.experiment_root))
exit(1)
else:
os.makedirs(args.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
log_file = os.path.join(args.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.
if not args.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not args.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
# Load the data from the CSV file.
pids, fids = common.load_dataset(args.train_set, args.image_root)
max_fid_len = max(map(len, fids)) # We'll need this later for logfiles.
# Load feature embeddings
if args.hard_pool_size > 0:
with h5py.File(args.train_embeddings, 'r') as f_train:
train_embs = np.array(f_train['emb'])
f_dists = scipy.spatial.distance.cdist(train_embs, train_embs)
hard_ids = get_hard_id_pool(pids, f_dists, args.hard_pool_size)
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids = np.unique(pids)
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
if args.hard_pool_size == 0:
dataset = dataset.take(
(len(unique_pids) // args.batch_p) * args.batch_p)
dataset = dataset.repeat(
None) # Repeat forever. Funny way of stating it.
else:
dataset = dataset.repeat(
None) # Repeat forever. Funny way of stating it.
dataset = dataset.map(lambda pid: sample_batch_ids_for_pid(
pid, all_pids=pids, batch_p=args.batch_p, all_hard_pids=hard_ids))
# Unbatch the P PIDs
dataset = dataset.apply(tf.contrib.data.unbatch())
# For every PID, get K images.
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids, all_pids=pids, batch_k=args.batch_k))
# Ungroup/flatten the batches for easy loading of the files.
dataset = dataset.apply(tf.contrib.data.unbatch())
# Convert filenames to actual image tensors.
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset = dataset.map(lambda im, fid, pid: common.fid_to_image(
fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size if args.crop_augment else net_input_size),
num_parallel_calls=args.loading_threads)
# Augment the data if specified by the arguments.
if args.augment == False:
dataset = dataset.map(
lambda fid, pid: common.fid_to_image(fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size
if args.crop_augment else
net_input_size), #Ergys
num_parallel_calls=args.loading_threads)
if args.flip_augment:
dataset = dataset.map(lambda im, fid, pid: (
tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset = dataset.map(lambda im, fid, pid: (tf.random_crop(
im, net_input_size + (3, )), fid, pid))
else:
dataset = dataset.map(lambda im, fid, pid: common.fid_to_image(
fid, pid, image_root=args.image_root, image_size=net_input_size),
num_parallel_calls=args.loading_threads)
dataset = dataset.map(lambda im, fid, pid: (tf.py_func(
augment_images, [im], [tf.float32]), fid, pid))
dataset = dataset.map(lambda im, fid, pid: (tf.reshape(
im[0],
(args.net_input_height, args.net_input_width, 3)), fid, pid))
# Group it back into PK batches.
batch_size = args.batch_p * args.batch_k
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(batch_size * 2)
# Since we repeat the data infinitely, we only need a one-shot iterator.
images, fids, pids = dataset.make_one_shot_iterator().get_next()
# Create the model and an embedding head.
model = import_module('nets.' + args.model_name)
head = import_module('heads.' + args.head_name)
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, args.embedding_dim, is_training=True)
# Create the loss in two steps:
# 1. Compute all pairwise distances according to the specified metric.
# 2. For each anchor along the first dimension, compute its loss.
dists = loss.cdist(endpoints['emb'], endpoints['emb'], metric=args.metric)
losses, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[
args.loss](dists,
pids,
args.margin,
batch_precision_at_k=args.batch_k - 1)
# Count the number of active entries, and compute the total batch loss.
num_active = tf.reduce_sum(tf.cast(tf.greater(losses, 1e-5), tf.float32))
loss_mean = tf.reduce_mean(losses)
# Some logging for tensorboard.
tf.summary.histogram('loss_distribution', losses)
tf.summary.scalar('loss', loss_mean)
tf.summary.scalar('batch_top1', train_top1)
tf.summary.scalar('batch_prec_at_{}'.format(args.batch_k - 1), prec_at_k)
tf.summary.scalar('active_count', num_active)
tf.summary.histogram('embedding_dists', dists)
tf.summary.histogram('embedding_pos_dists', pos_dists)
tf.summary.histogram('embedding_neg_dists', neg_dists)
tf.summary.histogram('embedding_lengths',
tf.norm(endpoints['emb_raw'], axis=1))
# Create the mem-mapped arrays in which we'll log all training detail in
# addition to tensorboard, because tensorboard is annoying for detailed
# inspection and actually discards data in histogram summaries.
if args.detailed_logs:
log_embs = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'embeddings'),
dtype=np.float32,
shape=(args.train_iterations, batch_size, args.embedding_dim))
log_loss = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'losses'),
dtype=np.float32,
shape=(args.train_iterations, batch_size))
log_fids = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'fids'),
dtype='S' + str(max_fid_len),
shape=(args.train_iterations, batch_size))
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
global_step = tf.Variable(0, name='global_step', trainable=False)
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.train.exponential_decay(
args.learning_rate,
tf.maximum(0, global_step - args.decay_start_iteration),
args.train_iterations - args.decay_start_iteration, 0.001)
else:
learning_rate = args.learning_rate
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
# Feel free to try others!
# optimizer = tf.train.AdadeltaOptimizer(learning_rate)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.minimize(loss_mean, global_step=global_step)
# Define a saver for the complete model.
checkpoint_saver = tf.train.Saver(max_to_keep=0)
with tf.Session() as sess:
if args.resume:
# In case we're resuming, simply load the full checkpoint to init.
last_checkpoint = tf.train.latest_checkpoint(args.experiment_root)
log.info('Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(sess, args.initial_checkpoint)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.experiment_root,
sess.graph)
start_step = sess.run(global_step)
log.info('Starting training from iteration {}.'.format(start_step))
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, args.train_iterations):
# Compute gradients, update weights, store logs!
start_time = time.time()
_, summary, step, b_prec_at_k, b_embs, b_loss, b_fids = \
sess.run([train_op, merged_summary, global_step,
prec_at_k, endpoints['emb'], losses, fids])
elapsed_time = time.time() - start_time
# Compute the iteration speed and add it to the summary.
# We did observe some weird spikes that we couldn't track down.
summary2 = tf.Summary()
summary2.value.add(tag='secs_per_iter',
simple_value=elapsed_time)
summary_writer.add_summary(summary2, step)
summary_writer.add_summary(summary, step)
if args.detailed_logs:
log_embs[i], log_loss[i], log_fids[
i] = b_embs, b_loss, b_fids
# Do a huge print out of the current progress.
seconds_todo = (args.train_iterations - step) * elapsed_time
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'batch-p@{}: {:.2%}, ETA: {} ({:.2f}s/it)'.format(
step, float(np.min(b_loss)), float(np.mean(b_loss)),
float(np.max(b_loss)), args.batch_k - 1,
float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)), elapsed_time))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.
if (args.checkpoint_frequency > 0
and step % args.checkpoint_frequency == 0):
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=step)
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root, 'checkpoint'),
global_step=step)
def main():
args = parser.parse_args()
# We store all arguments in a json file. This has two advantages:
# 1. We can always get back and see what exactly that experiment was
# 2. We can resume an experiment as-is without needing to remember all flags.
args_file = os.path.join(args.experiment_root, 'args.json')
if args.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
for key, value in args.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
args.__dict__[key] = resumed_value
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
else:
# If the experiment directory exists already, we bail in fear.
if os.path.exists(args.experiment_root):
if os.listdir(args.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(args.experiment_root))
exit(1)
else:
os.makedirs(args.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
log_file = os.path.join(args.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.
if not args.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not args.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
# Load the data from the CSV file.
pids, fids = common.load_dataset(args.train_set, args.image_root)
max_fid_len = max(map(len, fids)) # We'll need this later for logfiles.
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids = np.unique(pids)
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
dataset = dataset.take((len(unique_pids) // args.batch_p) * args.batch_p)
dataset = dataset.repeat(None) # Repeat forever. Funny way of stating it.
# For every PID, get K images.
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids, all_pids=pids, batch_k=args.batch_k))
# Ungroup/flatten the batches for easy loading of the files.
dataset = dataset.apply(tf.contrib.data.unbatch())
# Convert filenames to actual image tensors.
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset = dataset.map(lambda fid, pid: common.fid_to_image(
fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size if args.crop_augment else net_input_size),
num_parallel_calls=args.loading_threads)
# Augment the data if specified by the arguments.
if args.flip_augment:
dataset = dataset.map(lambda im, fid, pid:
(tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset = dataset.map(lambda im, fid, pid: (tf.random_crop(
im, net_input_size + (3, )), fid, pid))
# Group it back into PK batches.
batch_size = args.batch_p * args.batch_k
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(1)
# Since we repeat the data infinitely, we only need a one-shot iterator.
images, fids, pids = dataset.make_one_shot_iterator().get_next()
# Create the model and an embedding head.
model = import_module('nets.' + args.model_name)
head = import_module('heads.' + args.head_name)
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, args.embedding_dim, is_training=True)
# Create the loss in two steps:
# 1. Compute all pairwise distances according to the specified metric.
# 2. For each anchor along the first dimension, compute its loss.
dists = loss.cdist(endpoints['emb'], endpoints['emb'], metric=args.metric)
losses, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[
args.loss](dists,
pids,
args.margin,
batch_precision_at_k=args.batch_k - 1)
decDense = tf.layers.dense(
inputs=endpoints['emb'], units=5120,
name='decDense') # ,activation = tf.nn.relu ################
unflat = tf.reshape(decDense, shape=[tf.shape(decDense)[0], 32, 16, 10])
unp3shape = tf.TensorShape(
[2 * di for di in unflat.get_shape().as_list()[1:-1]])
unPool3 = tf.image.resize_nearest_neighbor(unflat,
unp3shape,
name='unpool3')
deConv3 = tf.layers.conv2d(inputs=unPool3,
filters=64,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv3')
unp2shape = tf.TensorShape(
[2 * di for di in deConv3.get_shape().as_list()[1:-1]])
unPool2 = tf.image.resize_nearest_neighbor(deConv3,
unp2shape,
name='unpool2')
deConv2 = tf.layers.conv2d(inputs=unPool2,
filters=32,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv2')
unp1shape = tf.TensorShape(
[2 * di for di in deConv2.get_shape().as_list()[1:-1]])
unPool1 = tf.image.resize_nearest_neighbor(deConv2,
unp1shape,
name='unpool1')
deConv1 = tf.layers.conv2d(inputs=unPool1,
filters=3,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=None,
name='deConv1')
imClip = deConv1 #tf.clip_by_value(t = deConv1,clip_value_min = -1.0,clip_value_max = 1.0,name='clipRelu')
print('RconstructeddImage : ', imClip.name)
recLoss = tf.multiply(
0.01, tf.losses.mean_squared_error(
labels=images,
predictions=imClip,
))
print('recLoss : ', recLoss.name)
decDense1 = tf.layers.dense(
inputs=endpoints['emb'], units=5120,
name='decDense1') # ,activation = tf.nn.relu ################
unflat1 = tf.reshape(decDense1, shape=[tf.shape(decDense1)[0], 32, 16, 10])
unp3shape1 = tf.TensorShape(
[2 * di for di in unflat1.get_shape().as_list()[1:-1]])
unPool3_new = tf.image.resize_nearest_neighbor(unflat1,
unp3shape1,
name='unpool3_new')
deConv3_new = tf.layers.conv2d(inputs=unPool3_new,
filters=64,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv3_new')
unp2shape_new = tf.TensorShape(
[2 * di for di in deConv3_new.get_shape().as_list()[1:-1]])
unPool2_new = tf.image.resize_nearest_neighbor(deConv3_new,
unp2shape_new,
name='unpool2_new')
deConv2_new = tf.layers.conv2d(inputs=unPool2_new,
filters=3,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv2_new')
unp1shape_new = tf.TensorShape(
[2 * di for di in deConv2_new.get_shape().as_list()[1:-1]])
unPool1_new = tf.image.resize_nearest_neighbor(deConv2_new,
unp1shape_new,
name='unpool1_new')
deConv1_new = tf.layers.conv2d(inputs=unPool1_new,
filters=3,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=None,
name='deConv1_new')
imClip1 = deConv2_new
print('RconstructeddImage : ', imClip1.name)
print(imClip1.shape)
images2 = tf.image.resize_images(images, [128, 64])
print(images2.shape)
recLoss1 = tf.multiply(
0.01,
tf.losses.mean_squared_error(
labels=images2,
predictions=imClip1,
))
print('recLoss_new : ', recLoss1.name)
decDense2 = tf.layers.dense(
inputs=endpoints['emb'], units=5120,
name='decDense2') # ,activation = tf.nn.relu ################
unflat12 = tf.reshape(decDense2,
shape=[tf.shape(decDense2)[0], 32, 16, 10])
unp3shape12 = tf.TensorShape(
[2 * di for di in unflat12.get_shape().as_list()[1:-1]])
unPool3_new2 = tf.image.resize_nearest_neighbor(unflat12,
unp3shape12,
name='unpool3_new2')
deConv3_new2 = tf.layers.conv2d(inputs=unPool3_new2,
filters=3,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv3_new2')
unp2shape_new2 = tf.TensorShape(
[2 * di for di in deConv3_new2.get_shape().as_list()[1:-1]])
unPool2_new2 = tf.image.resize_nearest_neighbor(deConv3_new2,
unp2shape_new2,
name='unpool2_new2')
imClip11 = deConv3_new2
images21 = tf.image.resize_images(images, [64, 32])
recLoss2 = tf.multiply(
0.01,
tf.losses.mean_squared_error(
labels=images21,
predictions=imClip11,
))
print('recLoss_new : ', recLoss2.name)
decDensel = tf.layers.dense(
inputs=endpoints['emb'], units=5120,
name='decDensel') # ,activation = tf.nn.relu ################
unflatl = tf.reshape(decDensel, shape=[tf.shape(decDensel)[0], 32, 16, 10])
unp3shapel = tf.TensorShape(
[2 * di for di in unflatl.get_shape().as_list()[1:-1]])
unPool3l = tf.image.resize_nearest_neighbor(unflatl,
unp3shapel,
name='unpool3l')
deConv3l = tf.layers.conv2d(inputs=unPool3l,
filters=64,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv3l')
unp2shapel = tf.TensorShape(
[2 * di for di in deConv3l.get_shape().as_list()[1:-1]])
unPool2l = tf.image.resize_nearest_neighbor(deConv3l,
unp2shapel,
name='unpool2l')
deConv2l = tf.layers.conv2d(inputs=unPool2l,
filters=32,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=tf.nn.relu,
name='deConv2l')
unp1shapel = tf.TensorShape(
[2 * di for di in deConv2l.get_shape().as_list()[1:-1]])
unPool1l = tf.image.resize_nearest_neighbor(deConv2l,
unp1shapel,
name='unpool1l')
deConv1l = tf.layers.conv2d(inputs=unPool1l,
filters=3,
kernel_size=[5, 5],
strides=(1, 1),
padding='same',
activation=None,
name='deConv1l')
imClipl = deConv1l #tf.clip_by_value(t = deConv1,clip_value_min = -1.0,clip_value_max = 1.0,name='clipRelu')
print('RconstructeddImage : ', imClipl.name)
recLossl = tf.multiply(
0.01, tf.losses.mean_squared_error(
labels=images,
predictions=imClipl,
))
print('recLoss : ', recLossl.name)
# Count the number of active entries, and compute the total batch loss.
num_active = tf.reduce_sum(tf.cast(tf.greater(losses, 1e-5), tf.float32))
loss_mean = tf.reduce_mean(losses)
# Some logging for tensorboard.
tf.summary.histogram('loss_distribution', losses)
tf.summary.scalar('loss', loss_mean)
tf.summary.scalar('batch_top1', train_top1)
tf.summary.scalar('batch_prec_at_{}'.format(args.batch_k - 1), prec_at_k)
tf.summary.scalar('active_count', num_active)
tf.summary.histogram('embedding_dists', dists)
tf.summary.histogram('embedding_pos_dists', pos_dists)
tf.summary.histogram('embedding_neg_dists', neg_dists)
tf.summary.histogram('embedding_lengths',
tf.norm(endpoints['emb_raw'], axis=1))
# Create the mem-mapped arrays in which we'll log all training detail in
# addition to tensorboard, because tensorboard is annoying for detailed
# inspection and actually discards data in histogram summaries.
if args.detailed_logs:
log_embs = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'embeddings'),
dtype=np.float32,
shape=(args.train_iterations, batch_size, args.embedding_dim))
log_loss = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'losses'),
dtype=np.float32,
shape=(args.train_iterations, batch_size))
log_fids = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'fids'),
dtype='S' + str(max_fid_len),
shape=(args.train_iterations, batch_size))
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
global_step = tf.Variable(0, name='global_step', trainable=False)
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.train.exponential_decay(
args.learning_rate,
tf.maximum(0, global_step - args.decay_start_iteration),
args.train_iterations - args.decay_start_iteration, 0.001)
else:
learning_rate = args.learning_rate
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
# Feel free to try others!
# optimizer = tf.train.AdadeltaOptimizer(learning_rate)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.minimize(tf.add(
loss_mean,
tf.add(recLoss, tf.add(recLoss1, tf.add(recLoss2, recLossl)))),
global_step=global_step)
# Define a saver for the complete model.
checkpoint_saver = tf.train.Saver(max_to_keep=0)
with tf.Session() as sess:
if args.resume:
# In case we're resuming, simply load the full checkpoint to init.
last_checkpoint = tf.train.latest_checkpoint(args.experiment_root)
log.info('Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(sess, args.initial_checkpoint)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.experiment_root,
sess.graph)
start_step = sess.run(global_step)
log.info('Starting training from iteration {}.'.format(start_step))
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, args.train_iterations):
# Compute gradients, update weights, store logs!
start_time = time.time()
_, summary, step, b_prec_at_k, b_embs, b_loss, b_fids ,b_rec, b_rec1= \
sess.run([train_op, merged_summary, global_step,
prec_at_k, endpoints['emb'], losses, fids,recLoss, recLoss1])
elapsed_time = time.time() - start_time
# Compute the iteration speed and add it to the summary.
# We did observe some weird spikes that we couldn't track down.
summary2 = tf.Summary()
summary2.value.add(tag='secs_per_iter',
simple_value=elapsed_time)
summary_writer.add_summary(summary2, step)
summary_writer.add_summary(summary, step)
if args.detailed_logs:
log_embs[i], log_loss[i], log_fids[
i] = b_embs, b_loss, b_fids
# Do a huge print out of the current progress.
seconds_todo = (args.train_iterations - step) * elapsed_time
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'recLoss: {:.3f} batch-p@{}: {:.2%}, ETA: {} ({:.2f}s/it)'.
format(step, float(np.min(b_loss)), float(np.mean(b_loss)),
float(np.max(b_loss)), b_rec, args.batch_k - 1,
float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)), elapsed_time))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.
if (args.checkpoint_frequency > 0
and step % args.checkpoint_frequency == 0):
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=step)
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root, 'checkpoint'),
global_step=step)
def main(argv):
args = parser.parse_args(argv)
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
# We store all arguments in a json file. This has two advantages:
# 1. We can always get back and see what exactly that experiment was
# 2. We can resume an experiment as-is without needing to remember all flags.
args_file = os.path.join(args.experiment_root, 'args.json')
if args.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
for key, value in args.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
args.__dict__[key] = resumed_value
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
else:
# If the experiment directory exists already, we bail in fear.
if os.path.exists(args.experiment_root):
if os.listdir(args.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(args.experiment_root))
exit(1)
else:
os.makedirs(args.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
log_file = os.path.join(args.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.
if not args.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not args.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
# Load the data from the CSV file.
pids, fids = common.load_dataset(args.train_set, args.image_root)
max_fid_len = max(map(len, fids)) # We'll need this later for logfiles.
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids = np.unique(pids)
if len(unique_pids) < args.batch_p:
unique_pids = np.tile(unique_pids,
int(np.ceil(args.batch_p / len(unique_pids))))
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
dataset = dataset.take((len(unique_pids) // args.batch_p) * args.batch_p)
dataset = dataset.repeat(None) # Repeat forever. Funny way of stating it.
# For every PID, get K images.
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids, all_pids=pids, batch_k=args.batch_k))
# Ungroup/flatten the batches for easy loading of the files.
dataset = dataset.apply(tf.contrib.data.unbatch())
# Convert filenames to actual image tensors.
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset = dataset.map(lambda fid, pid: common.fid_to_image(
fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size if args.crop_augment else net_input_size),
num_parallel_calls=args.loading_threads)
# Augment the data if specified by the arguments.
dataset = dataset.map(
lambda im, fid, pid: common.fid_to_image(fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size
if args.crop_augment else
net_input_size), # Ergys
num_parallel_calls=args.loading_threads)
if args.flip_augment:
dataset = dataset.map(lambda im, fid, pid:
(tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset = dataset.map(lambda im, fid, pid: (tf.random_crop(
im, net_input_size + (3, )), fid, pid))
# Group it back into PK batches.
batch_size = args.batch_p * args.batch_k
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(1)
# Since we repeat the data infinitely, we only need a one-shot iterator.
images, fids, pids = dataset.make_one_shot_iterator().get_next()
# Create the model and an embedding head.
model = import_module('nets.' + args.model_name)
head = import_module('heads.' + args.head_name)
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
weight_decay = 10e-4
weights_regularizer = tf.contrib.layers.l2_regularizer(scale=weight_decay)
endpoints, body_prefix = model.endpoints(images, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints,
args.embedding_dim,
is_training=True,
weights_regularizer=weights_regularizer)
# Create the loss in two steps:
# 1. Compute all pairwise distances according to the specified metric.
# 2. For each anchor along the first dimension, compute its loss.
# batch_embedding = endpoints['emb']
batch_embedding = endpoints['emb']
if args.loss == 'semi_hard_triplet':
triplet_loss = triplet_semihard_loss(batch_embedding, pids,
args.margin)
elif args.loss == 'hard_triplet':
triplet_loss = batch_hard(batch_embedding, pids, args.margin,
args.metric)
elif args.loss == 'lifted_loss':
triplet_loss = lifted_loss(pids, batch_embedding, margin=args.margin)
elif args.loss == 'contrastive_loss':
assert batch_size % 2 == 0
assert args.batch_k == 4 ## Can work with other number but will need tuning
contrastive_idx = np.tile([0, 1, 4, 3, 2, 5, 6, 7], args.batch_p // 2)
for i in range(args.batch_p // 2):
contrastive_idx[i * 8:i * 8 + 8] += i * 8
contrastive_idx = np.expand_dims(contrastive_idx, 1)
batch_embedding_ordered = tf.gather_nd(batch_embedding,
contrastive_idx)
pids_ordered = tf.gather_nd(pids, contrastive_idx)
# batch_embedding_ordered = tf.Print(batch_embedding_ordered,[pids_ordered],'pids_ordered :: ',summarize=1000)
embeddings_anchor, embeddings_positive = tf.unstack(
tf.reshape(batch_embedding_ordered, [-1, 2, args.embedding_dim]),
2, 1)
# embeddings_anchor = tf.Print(embeddings_anchor,[pids_ordered,embeddings_anchor,embeddings_positive,batch_embedding,batch_embedding_ordered],"Tensors ", summarize=1000)
fixed_labels = np.tile([1, 0, 0, 1], args.batch_p // 2)
# fixed_labels = np.reshape(fixed_labels,(len(fixed_labels),1))
# print(fixed_labels)
labels = tf.constant(fixed_labels)
# labels = tf.Print(labels,[labels],'labels ',summarize=1000)
triplet_loss = contrastive_loss(labels,
embeddings_anchor,
embeddings_positive,
margin=args.margin)
elif args.loss == 'angular_loss':
embeddings_anchor, embeddings_positive = tf.unstack(
tf.reshape(batch_embedding, [-1, 2, args.embedding_dim]), 2, 1)
# pids = tf.Print(pids, [pids], 'pids:: ', summarize=100)
pids, _ = tf.unstack(tf.reshape(pids, [-1, 2, 1]), 2, 1)
# pids = tf.Print(pids,[pids],'pids:: ',summarize=100)
triplet_loss = angular_loss(pids,
embeddings_anchor,
embeddings_positive,
batch_size=args.batch_p,
with_l2reg=True)
elif args.loss == 'npairs_loss':
assert args.batch_k == 2 ## Single positive pair per class
embeddings_anchor, embeddings_positive = tf.unstack(
tf.reshape(batch_embedding, [-1, 2, args.embedding_dim]), 2, 1)
pids, _ = tf.unstack(tf.reshape(pids, [-1, 2, 1]), 2, 1)
pids = tf.reshape(pids, [-1])
triplet_loss = npairs_loss(pids, embeddings_anchor,
embeddings_positive)
else:
raise NotImplementedError('loss function {} NotImplemented'.format(
args.loss))
loss_mean = tf.reduce_mean(triplet_loss)
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
global_step = tf.Variable(0, name='global_step', trainable=False)
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.train.polynomial_decay(args.learning_rate,
global_step,
args.train_iterations,
end_learning_rate=1e-7,
power=1)
else:
learning_rate = args.learning_rate
if args.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(learning_rate)
elif args.optimizer == 'momentum':
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9)
else:
raise NotImplementedError('Invalid optimizer {}'.format(
args.optimizer))
#
# learning_rate = tf.train.polynomial_decay(args.learning_rate, global_step,
# args.train_iterations, end_learning_rate= 1e-7,
# power=1)
#
# Feel free to try others!
# optimizer = tf.train.AdadeltaOptimizer(learning_rate)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.minimize(loss_mean, global_step=global_step)
# Define a saver for the complete model.
checkpoint_saver = tf.train.Saver(max_to_keep=2)
gpu_options = tf.GPUOptions(allow_growth=True)
gpu_config = tf.ConfigProto(gpu_options=gpu_options)
with tf.Session(config=gpu_config) as sess:
if args.resume:
# In case we're resuming, simply load the full checkpoint to init.
last_checkpoint = tf.train.latest_checkpoint(args.experiment_root)
if last_checkpoint == None:
print('Resume with No previous checkpoint')
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(sess, args.initial_checkpoint)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
else:
log.info(
'Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(sess, args.initial_checkpoint)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
start_step = sess.run(global_step)
log.info('Starting training from iteration {}.'.format(start_step))
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, args.train_iterations):
# Compute gradients, update weights, store logs!
start_time = time.time()
_, step, b_embs, b_loss, b_fids = \
sess.run([train_op, global_step, endpoints['emb'], triplet_loss, fids])
elapsed_time = time.time() - start_time
# Do a huge print out of the current progress.
seconds_todo = (args.train_iterations - step) * elapsed_time
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, ETA: {} ({:.2f}s/it)'
.format(
step,
float(np.min(b_loss)),
float(np.mean(b_loss)),
float(np.max(b_loss)),
# args.batch_k - 1, float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)),
elapsed_time))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.
if (args.checkpoint_frequency > 0
and step % args.checkpoint_frequency == 0):
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=step)
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root, 'checkpoint'),
global_step=step)
tf.global_variables_initializer().run()
if resume:
last_checkpoint = tf.train.latest_checkpoint(tr.save_path)
saver.restore(sess, last_checkpoint)
start_step = sess.run(global_step)
logger.debug('Resume training ... Start from step %d / %d .' %
(start_step, train_nums))
resume = False
else:
start_step = 0
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, train_nums):
_, loss_value, step, ttt = sess.run(
[train_op, loss, global_step, tt])
if i % print_steps == 0:
top1, top5, top10 = sess.run([
accuracy_top1_batch, accuracy_top5_batch,
accuracy_top10_batch
])
logger.debug(
"After %d training step(s),loss on training batch is %g.The batch test accuracy = %g , %g ，%g."
% (i, loss_value, top1, top5, top10))
'''
losslist.append([step,loss_value])
accuracy.append([step,top1])
def main(args):
best_acc = -1
logger = bit_common.setup_logger(args)
cp, cn = smooth_BCE(eps=0.1)
# Lets cuDNN benchmark conv implementations and choose the fastest.
# Only good if sizes stay the same within the main loop!
torch.backends.cudnn.benchmark = True
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.info(f"Going to train on {device}")
classes = 5
train_set, valid_set, train_loader, valid_loader = mktrainval(args, logger)
logger.info(f"Loading model from {args.model}.npz")
#model = models.KNOWN_MODELS[args.model](head_size=classes, zero_head=True)
#model.load_from(np.load(f"{args.model}.npz"))
model = EfficientNet.from_pretrained(args.model, num_classes=classes)
logger.info("Moving model onto all GPUs")
model = torch.nn.DataParallel(model)
# Optionally resume from a checkpoint.
# Load it to CPU first as we'll move the model to GPU later.
# This way, we save a little bit of GPU memory when loading.
start_epoch = 0
# Note: no weight-decay!
optim = torch.optim.SGD(model.parameters(), lr=0.003, momentum=0.9)
# Resume fine-tuning if we find a saved model.
savename = pjoin(args.logdir, args.name, "bit.pth.tar")
try:
logger.info(f"Model will be saved in '{savename}'")
checkpoint = torch.load(savename, map_location="cpu")
logger.info(f"Found saved model to resume from at '{savename}'")
start_epoch = checkpoint["epoch"]
model.load_state_dict(checkpoint["model"])
optim.load_state_dict(checkpoint["optim"])
logger.info(f"Resumed at epoch {start_epoch}")
except FileNotFoundError:
logger.info("Fine-tuning from BiT")
model = model.to(device)
optim.zero_grad()
model.train()
mixup = bit_hyperrule.get_mixup(len(train_set))
#mixup = -1
cri = torch.nn.CrossEntropyLoss().to(device)
#cri = FocalLoss(cri)
logger.info("Starting training!")
chrono = lb.Chrono()
accum_steps = 0
mixup_l = np.random.beta(mixup, mixup) if mixup > 0 else 1
end = time.time()
epoches = 10
scheduler = torch.optim.lr_scheduler.OneCycleLR(optim,
max_lr=0.01,
steps_per_epoch=1,
epochs=epoches)
with lb.Uninterrupt() as u:
for epoch in range(start_epoch, epoches):
pbar = enumerate(train_loader)
pbar = tqdm.tqdm(pbar, total=len(train_loader))
scheduler.step()
all_top1, all_top5 = [], []
for param_group in optim.param_groups:
lr = param_group["lr"]
#for x, y in recycle(train_loader):
for batch_id, (x, y) in pbar:
#for batch_id, (x, y) in enumerate(train_loader):
# measure data loading time, which is spent in the `for` statement.
chrono._done("load", time.time() - end)
if u.interrupted:
break
# Schedule sending to GPU(s)
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
# Update learning-rate, including stop training if over.
#lr = bit_hyperrule.get_lr(step, len(train_set), args.base_lr)
#if lr is None:
# break
if mixup > 0.0:
x, y_a, y_b = mixup_data(x, y, mixup_l)
# compute output
with chrono.measure("fprop"):
logits = model(x)
top1, top5 = topk(logits, y, ks=(1, 5))
all_top1.extend(top1.cpu())
all_top5.extend(top5.cpu())
if mixup > 0.0:
c = mixup_criterion(cri, logits, y_a, y_b, mixup_l)
else:
c = cri(logits, y)
train_loss = c.item()
train_acc = np.mean(all_top1) * 100.0
# Accumulate grads
with chrono.measure("grads"):
(c / args.batch_split).backward()
accum_steps += 1
accstep = f"({accum_steps}/{args.batch_split})" if args.batch_split > 1 else ""
s = f"epoch={epoch} batch {batch_id}{accstep}: loss={train_loss:.5f} train_acc={train_acc:.2f} lr={lr:.1e}"
#s = f"epoch={epoch} batch {batch_id}{accstep}: loss={c.item():.5f} lr={lr:.1e}"
pbar.set_description(s)
#logger.info(f"[batch {batch_id}{accstep}]: loss={c_num:.5f} (lr={lr:.1e})") # pylint: disable=logging-format-interpolation
logger.flush()
# Update params
with chrono.measure("update"):
optim.step()
optim.zero_grad()
# Sample new mixup ratio for next batch
mixup_l = np.random.beta(mixup, mixup) if mixup > 0 else 1
# Run evaluation and save the model.
val_loss, val_acc = run_eval(model, valid_loader, device, chrono,
logger, epoch)
best = val_acc > best_acc
if best:
best_acc = val_acc
torch.save(
{
"epoch": epoch,
"val_loss": val_loss,
"val_acc": val_acc,
"train_acc": train_acc,
"model": model.state_dict(),
"optim": optim.state_dict(),
}, savename)
end = time.time()
logger.info(f"Timings:\n{chrono}")
def main():
args = parser.parse_args()
# We store all arguments in a json file. This has two advantages: 我们将所有参数存储在json文件中。 这有两个好处：
# 1. We can always get back and see what exactly that experiment was 1.我们总是可以回头看看实验是什么
# 2. We can resume an experiment as-is without needing to remember all flags.2.我们可以按原样恢复实验，无需记住所有标志。
args_file = os.path.join(args.experiment_root, 'args.json')
if args.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
#恢复时，我们不仅需要使用文件中的值填充args对象，但我们也想检查加载参数和给定参数之间的一些可能的冲突。
for key, value in args.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
args.__dict__[key] = resumed_value
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
else:
# If the experiment directory exists already, we bail in fear.如果实验目录已经存在，我们就会担心。
if os.path.exists(args.experiment_root):
if os.listdir(args.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(args.experiment_root))
exit(1)
else:
os.makedirs(args.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.将传递的参数存储起来，以便稍后以一种很好且可读的格式恢复和刷新。
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
log_file = os.path.join(args.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
#同时在开始时显示所有参数值，便于读取日志。
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.在这里检查他们，
if not args.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not args.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
# Load the data from the CSV file. 加载CSV文件中的数据。
pids, fids = common.load_dataset(args.train_set, args.image_root)
max_fid_len = max(map(
len, fids)) # We'll need this later for logfiles.我们稍后需要这个日志文件。
# Setup a tf.Dataset where one "epoch" loops over all PIDS. 设置一个tf.Dataset，其中一个“epoch”在所有PIDS上循环。
# PIDS are shuffled after every epoch and continue indefinitely.PIDS在每个时代之后都会被洗牌并无限期地继续下去。
unique_pids = np.unique(pids)
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.将数据集大小限制为批量的倍数，以便在每个时期结束时我们不会重叠。
dataset = dataset.take((len(unique_pids) // args.batch_p) * args.batch_p)
dataset = dataset.repeat(
None) # Repeat forever. Funny way of stating it. 永远重复。 说明它的有趣方式。
# For every PID, get K images.对于每个PID，获得K个图像。
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids, all_pids=pids, batch_k=args.batch_k))
# Ungroup/flatten the batches for easy loading of the files.取消组合/拼合批次以便轻松加载文件。
dataset = dataset.apply(tf.contrib.data.unbatch())
# Convert filenames to actual image tensors. 将文件名转换为实际图像张量。
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset = dataset.map(lambda fid, pid: common.fid_to_image(
fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size if args.crop_augment else net_input_size),
num_parallel_calls=args.loading_threads)
# Augment the data if specified by the arguments. 如果由参数指定，则增加数据。
if args.flip_augment:
dataset = dataset.map(lambda im, fid, pid:
(tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset = dataset.map(lambda im, fid, pid: (tf.random_crop(
im, net_input_size + (3, )), fid, pid))
# Group it back into PK batches. 将其重新分组为PK批次。
batch_size = args.batch_p * args.batch_k
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(1)
# Since we repeat the data infinitely, we only need a one-shot iterator.由于我们无限地重复数据，因此我们只需要一个one-shot 迭代器。
images, fids, pids = dataset.make_one_shot_iterator().get_next()
# Create the model and an embedding head.
model = import_module('nets.' + args.model_name)
head = import_module('heads.' + args.head_name)
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
#通过模型提供图像。返回的`body_prefix`将进一步用于加载具有此前缀的所有变量的预先训练权重。
endpoints, body_prefix = model.endpoints(images, is_training=True)
with tf.name_scope('head'):
endpoints = head.head(endpoints, args.embedding_dim, is_training=True)
# Create the loss in two steps: 分两步创建损失：
# 1. Compute all pairwise distances according to the specified metric.1.根据指定的度量计算所有成对距离。
# 2. For each anchor along the first dimension, compute its loss. 2.对于第一维中的每个锚，计算其损失。
dists = loss.cdist(endpoints['emb'], endpoints['emb'], metric=args.metric)
losses, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[
args.loss](dists,
pids,
args.margin,
batch_precision_at_k=args.batch_k - 1)
# Count the number of active entries, and compute the total batch loss. 计算活动条目的数量，并计算总批量损失。
num_active = tf.reduce_sum(tf.cast(tf.greater(losses, 1e-5), tf.float32))
loss_mean = tf.reduce_mean(losses)
# Some logging for tensorboard. 一些日志记录在tensorboard。
tf.summary.histogram('loss_distribution', losses)
tf.summary.scalar('loss', loss_mean)
tf.summary.scalar('batch_top1', train_top1)
tf.summary.scalar('batch_prec_at_{}'.format(args.batch_k - 1), prec_at_k)
tf.summary.scalar('active_count', num_active)
tf.summary.histogram('embedding_dists', dists)
tf.summary.histogram('embedding_pos_dists', pos_dists)
tf.summary.histogram('embedding_neg_dists', neg_dists)
tf.summary.histogram('embedding_lengths',
tf.norm(endpoints['emb_raw'], axis=1))
# Create the mem-mapped arrays in which we'll log all training detail in
# addition to tensorboard, because tensorboard is annoying for detailed
# inspection and actually discards data in histogram summaries.
#创建mem - mapped数组，我们将记录除tensorboard之外的所有训练细节，因为tensorboard对于详细检查很烦人，实际上在直方图总结中丢弃了数据。
if args.detailed_logs:
log_embs = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'embeddings'),
dtype=np.float32,
shape=(args.train_iterations, batch_size, args.embedding_dim))
log_loss = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'losses'),
dtype=np.float32,
shape=(args.train_iterations, batch_size))
log_fids = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'fids'),
dtype='S' + str(max_fid_len),
shape=(args.train_iterations, batch_size))
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
#在我们添加优化器之前收集这些信息，因为根据优化器的不同，它可能会添加额外的插槽，这些插槽也是全局变量，具有完全相同的前缀。
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
# Define the optimizer and the learning-rate schedule. 定义优化器和学习率计划。
# Unfortunately, we get NaNs if we don't handle no-decay separately. 不幸的是，如果我们不单独处理无衰减，我们会得到NaNs。
global_step = tf.Variable(0, name='global_step', trainable=False)
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.train.exponential_decay(
args.learning_rate,
tf.maximum(0, global_step - args.decay_start_iteration),
args.train_iterations - args.decay_start_iteration, 0.001)
else:
learning_rate = args.learning_rate
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
# Feel free to try others! 随意尝试别的的优化器！
# optimizer = tf.train.AdadeltaOptimizer(learning_rate)
# Update_ops are used to update batchnorm stats. Update_ops用于更新batchnorm统计信息。
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.minimize(loss_mean, global_step=global_step)
# Define a saver for the complete model. 为整个模型定义一个保存器。
checkpoint_saver = tf.train.Saver(max_to_keep=0)
with tf.Session() as sess:
if args.resume:
# In case we're resuming, simply load the full checkpoint to init.如果我们正在恢复，只需将完整的检查点加载到init。
last_checkpoint = tf.train.latest_checkpoint(args.experiment_root)
log.info('Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
#但是如果我们从头开始，我们可能需要从预先训练的权重中加载一些变量，并随机初始化其他的变量。
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(sess, args.initial_checkpoint)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
#无论如何，我们也将这个初始化作为检查点存储，以便我们可以运行完全可再生的实验。
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.experiment_root,
sess.graph)
start_step = sess.run(global_step)
log.info('Starting training from iteration {}.'.format(start_step))
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
#最后，这里是主循环。这个`Uninterrupt`是一个非常方便的工具，可以在Ctrl + C之后完成迭代才停止，我们可以干净地停止训练。
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, args.train_iterations):
# Compute gradients, update weights, store logs!计算梯度，更新权重，存储日志！
start_time = time.time()
_, summary, step, b_prec_at_k, b_embs, b_loss, b_fids = \
sess.run([train_op, merged_summary, global_step,
prec_at_k, endpoints['emb'], losses, fids])
elapsed_time = time.time() - start_time
# Compute the iteration speed and add it to the summary.计算迭代速度并将其添加到摘要中。
# We did observe some weird spikes that we couldn't track down.我们确实观察到一些我们无法追查的奇怪尖峰。
summary2 = tf.Summary()
summary2.value.add(tag='secs_per_iter',
simple_value=elapsed_time)
summary_writer.add_summary(summary2, step)
summary_writer.add_summary(summary, step)
if args.detailed_logs:
log_embs[i], log_loss[i], log_fids[
i] = b_embs, b_loss, b_fids
# Do a huge print out of the current progress. 在当前的进展中做大量的印刷。
seconds_todo = (args.train_iterations - step) * elapsed_time
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'batch-p@{}: {:.2%}, ETA: {} ({:.2f}s/it)'.format(
step, float(np.min(b_loss)), float(np.mean(b_loss)),
float(np.max(b_loss)), args.batch_k - 1,
float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)), elapsed_time))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.每隔一段时间保存一次训练的检查点checkpoint。
if (args.checkpoint_frequency > 0
and step % args.checkpoint_frequency == 0):
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=step)
# Stop the main-loop at the end of the step, if requested. 如果需要，在步骤结束时停止主循环。
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
#存储一个最终检查点。 这可能是多余的，但是在中间存储被禁用的情况下它是至关重要的，并且在进程被中断时保存检查点。
checkpoint_saver.save(sess,
os.path.join(args.experiment_root, 'checkpoint'),
global_step=step)
def main(args):
logger = common.setup_logger(args)
# Lets cuDNN benchmark conv implementations and choose the fastest.
# Only good if sizes stay the same within the main loop!
torch.backends.cudnn.benchmark = True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
logger.info(f"Going to train on {device}")
train_set, valid_set, train_loader, valid_loader = mktrainval(args, logger)
logger.info(f"Loading model from {args.model}.npz")
model = models.KNOWN_MODELS[args.model](head_size=len(valid_set.classes),
zero_head=True)
model.load_from(
np.load(os.path.join(args.pretrained_dir, f"{args.model}.npz")))
logger.info("Moving model onto all GPUs")
model = torch.nn.DataParallel(model)
# Optionally resume from a checkpoint.
# Load it to CPU first as we'll move the model to GPU later.
# This way, we save a little bit of GPU memory when loading.
step = 0
# Note: no weight-decay!
optim = torch.optim.SGD(model.parameters(), lr=args.base_lr, momentum=0.9)
writer = SummaryWriter(os.path.join(args.logdir, args.name))
# Resume fine-tuning if we find a saved model.
savename = pjoin(args.logdir, args.name, "model.tar")
try:
logger.info(f"Model will be saved in '{savename}'")
checkpoint = torch.load(savename, map_location="cpu")
logger.info(f"Found saved model to resume from at '{savename}'")
step = checkpoint["step"]
model.load_state_dict(checkpoint["model"])
optim.load_state_dict(checkpoint["optim"])
logger.info(f"Resumed at step {step}")
except FileNotFoundError:
logger.info("Fine-tuning from BiT")
model = model.to(device)
optim.zero_grad()
model.train()
mixup = hyperrule.get_mixup(len(train_set))
cri = torch.nn.CrossEntropyLoss().to(device)
logger.info("Starting training!")
chrono = lb.Chrono()
accum_steps = 0
mixup_l = np.random.beta(mixup, mixup) if mixup > 0 else 1
end = time.time()
with lb.Uninterrupt() as u:
for x, y in recycle(train_loader):
# measure data loading time, which is spent in the `for` statement.
chrono._done("load", time.time() - end)
if u.interrupted:
break
# Schedule sending to GPU(s)
x = x.to(device, non_blocking=True)
y = y.to(device, non_blocking=True)
# Update learning-rate, including stop training if over.
lr = hyperrule.get_lr(step, len(train_set), args.base_lr)
if lr is None:
break
for param_group in optim.param_groups:
param_group["lr"] = lr
if mixup > 0.0:
x, y_a, y_b = mixup_data(x, y, mixup_l)
# compute output
with chrono.measure("fprop"):
logits = model(x)
if mixup > 0.0:
c = mixup_criterion(cri, logits, y_a, y_b, mixup_l)
else:
c = cri(logits, y)
c_num = float(
c.data.cpu().numpy()) # Also ensures a sync point.
# Accumulate grads
with chrono.measure("grads"):
(c / args.batch_split).backward()
accum_steps += 1
accstep = f" ({accum_steps}/{args.batch_split})" if args.batch_split > 1 else ""
logger.info(
f"[step {step}{accstep}]: loss={c_num:.5f} (lr={lr:.1e})") # pylint: disable=logging-format-interpolation
logger.flush()
writer.add_scalar('Train/loss', c_num, step)
writer.add_scalar('Train/lr', lr, step)
# Update params
if accum_steps == args.batch_split:
with chrono.measure("update"):
optim.step()
optim.zero_grad()
step += 1
accum_steps = 0
# Sample new mixup ratio for next batch
mixup_l = np.random.beta(mixup, mixup) if mixup > 0 else 1
# Run evaluation and save the model.
if args.eval_every and step % args.eval_every == 0:
run_eval(model, valid_loader, device, chrono, logger,
writer, step)
if args.save and step % args.save_every == 0:
step_savename = pjoin(args.logdir, args.name,
"model_" + str(step) + ".tar")
torch.save(
{
"step": step,
"model": model.state_dict(),
"optim": optim.state_dict()
}, step_savename)
end = time.time()
# Final eval at end of training.
run_eval(model, valid_loader, device, chrono, logger, writer, step)
logger.info(f"Timings:\n{chrono}")
def train(args, images, fids, pids, max_fid_len, log):
'''
Creation model and training neural network
:param args: all stored arguments
:param images: prepared images for training
:param fids: figure id (relative paths from image_root to images)
:param pids: person id (or car id) for all images
:param log: log file, where logs from training are stored
:return: saved files (checkpoints, train log file)
'''
###################################################################################################################
# CREATE MODEL
###################################################################################################################
# Create the model and an embedding head.
model = import_module('nets.resnet_v1_50')
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
drops = {}
if args.dropout is not None:
drops = getDropoutProbs(args.dropout)
b4_layers = None
try:
b4_layers = int(args.b4_layers)
if b4_layers not in [1, 2, 3]: raise ValueError()
except:
ValueError("Argument exception: b4_layers has to be in [1, 2, 3]")
endpoints, body_prefix = model.endpoints(images,
b4_layers,
drops,
is_training=True,
resnet_stride=int(
args.resnet_stride))
endpoints['emb'] = endpoints['emb_raw'] = slim.fully_connected(
endpoints['model_output'],
args.embedding_dim,
activation_fn=None,
weights_initializer=tf.orthogonal_initializer(),
scope='emb')
step_pl = tf.placeholder(dtype=tf.float32)
features = endpoints['emb']
# Create the loss in two steps:
# 1. Compute all pairwise distances according to the specified metric.
# 2. For each anchor along the first dimension, compute its loss.
dists = loss.cdist(features, features, metric=args.metric)
losses, train_top1, prec_at_k, _, probe_neg_dists, pos_dists, neg_dists = loss.loss_function(
dists,
pids, [args.alpha1, args.alpha2, args.alpha3],
batch_precision_at_k=args.batch_k - 1)
# Count the number of active entries, and compute the total batch loss.
num_active = tf.reduce_sum(tf.cast(tf.greater(losses, 1e-5), tf.float32))
loss_mean = tf.reduce_mean(losses)
# Some logging for tensorboard.
tf.summary.histogram('loss_distribution', losses)
tf.summary.scalar('loss', loss_mean)
tf.summary.scalar('batch_top1', train_top1)
tf.summary.scalar('batch_prec_at_{}'.format(args.batch_k - 1), prec_at_k)
tf.summary.scalar('active_count', num_active)
tf.summary.scalar('embedding_pos_dists', tf.reduce_mean(pos_dists))
tf.summary.scalar('embedding_probe_neg_dists',
tf.reduce_mean(probe_neg_dists))
tf.summary.scalar('embedding_neg_dists', tf.reduce_mean(neg_dists))
tf.summary.histogram('embedding_dists', dists)
tf.summary.histogram('embedding_pos_dists', pos_dists)
tf.summary.histogram('embedding_probe_neg_dists', probe_neg_dists)
tf.summary.histogram('embedding_neg_dists', neg_dists)
tf.summary.histogram('embedding_lengths',
tf.norm(endpoints['emb_raw'], axis=1))
# Create the mem-mapped arrays in which we'll log all training detail in
# addition to tensorboard, because tensorboard is annoying for detailed
# inspection and actually discards data in histogram summaries.
batch_size = args.batch_p * args.batch_k
if args.detailed_logs:
log_embs = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'embeddings'),
dtype=np.float32,
shape=(args.train_iterations, batch_size, args.embedding_dim))
log_loss = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'losses'),
dtype=np.float32,
shape=(args.train_iterations, batch_size))
log_fids = lb.create_or_resize_dat(
os.path.join(args.experiment_root, 'fids'),
dtype='S' + str(max_fid_len),
shape=(args.train_iterations, batch_size))
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
body_prefix)
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
global_step = tf.Variable(0, name='global_step', trainable=False)
if args.sgdr:
learning_rate = tf.train.cosine_decay_restarts(
learning_rate=args.learning_rate,
global_step=global_step,
first_decay_steps=4000,
t_mul=1.5)
else:
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.train.exponential_decay(
args.learning_rate,
tf.maximum(0, global_step - args.decay_start_iteration),
args.train_iterations - args.decay_start_iteration,
float(args.lr_decay))
else:
learning_rate = args.learning_rate
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(tf.convert_to_tensor(learning_rate))
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
train_op = optimizer.minimize(loss_mean, global_step=global_step)
# Define a saver for the complete model.
checkpoint_saver = tf.train.Saver(max_to_keep=0)
with tf.Session() as sess:
if args.resume:
# In case we're resuming, simply load the full checkpoint to init.
last_checkpoint = tf.train.latest_checkpoint(args.experiment_root)
log.info('Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if args.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables)
saver.restore(sess, args.initial_checkpoint)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=0)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(args.experiment_root,
sess.graph)
start_step = sess.run(global_step)
step = start_step
log.info('Starting training from iteration {}.'.format(start_step))
###################################################################################################################
# TRAINING
###################################################################################################################
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, args.train_iterations):
# Compute gradients, update weights, store logs!
start_time = time.time()
_, summary, step, b_prec_at_k, b_embs, b_loss, b_fids = \
sess.run([train_op, merged_summary, global_step,
prec_at_k, features, losses, fids], feed_dict={step_pl: step})
elapsed_time = time.time() - start_time
# Compute the iteration speed and add it to the summary.
# We did observe some weird spikes that we couldn't track down.
summary2 = tf.Summary()
summary2.value.add(tag='secs_per_iter',
simple_value=elapsed_time)
summary_writer.add_summary(summary2, step)
summary_writer.add_summary(summary, step)
if args.detailed_logs:
log_embs[i], log_loss[i], log_fids[
i] = b_embs, b_loss, b_fids
# Do a huge print out of the current progress. Maybe steal from here.
seconds_todo = (args.train_iterations - step) * elapsed_time
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'batch-p@{}: {:.2%}, ETA: {} ({:.2f}s/it), lr={:.4g}'.
format(step, float(np.min(b_loss)), float(np.mean(b_loss)),
float(np.max(b_loss)), args.batch_k - 1,
float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)), elapsed_time,
sess.run(optimizer._lr)))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.
if (args.checkpoint_frequency > 0
and step % args.checkpoint_frequency == 0):
checkpoint_saver.save(sess,
os.path.join(args.experiment_root,
'checkpoint'),
global_step=step)
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
checkpoint_saver.save(sess,
os.path.join(args.experiment_root, 'checkpoint'),
global_step=step)
def main(argv):
args = parser.parse_args(argv)
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
# tf.compat.v1.disable_eager_execution()
# physical_devices = tf.config.experimental.list_physical_devices('GPU')
# tf.config.experimental.set_memory_growth(physical_devices[0], True)
# We store all arguments in a json file. This has two advantages:
# 1. We can always get back and see what exactly that experiment was
# 2. We can resume an experiment as-is without needing to remember all flags.
args_file = os.path.join(args.experiment_root, 'args.json')
if args.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
for key, value in args.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
args.__dict__[key] = resumed_value
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
else:
# If the experiment directory exists already, we bail in fear.
if os.path.exists(args.experiment_root):
if os.listdir(args.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(args.experiment_root))
exit(1)
else:
os.makedirs(args.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.
with open(args_file, 'w') as f:
json.dump(vars(args),
f,
ensure_ascii=False,
indent=2,
sort_keys=True)
log_file = os.path.join(args.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.
if not args.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not args.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
# Load the data from the CSV file.
pids, fids = common.load_dataset(args.train_set, args.image_root)
max_fid_len = max(map(len, fids)) # We'll need this later for logfiles.
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids = np.unique(pids)
if len(unique_pids) < args.batch_p:
unique_pids = np.tile(unique_pids,
int(np.ceil(args.batch_p / len(unique_pids))))
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
dataset = dataset.take((len(unique_pids) // args.batch_p) * args.batch_p)
dataset = dataset.repeat(None) # Repeat forever. Funny way of stating it.
# For every PID, get K images.
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids, all_pids=pids, batch_k=args.batch_k))
# Ungroup/flatten the batches for easy loading of the files.
dataset = dataset.unbatch()
# Convert filenames to actual image tensors.
net_input_size = (args.net_input_height, args.net_input_width)
pre_crop_size = (args.pre_crop_height, args.pre_crop_width)
dataset = dataset.map(lambda fid, pid: common.fid_to_image(
fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size if args.crop_augment else net_input_size),
num_parallel_calls=args.loading_threads)
# Augment the data if specified by the arguments.
dataset = dataset.map(
lambda im, fid, pid: common.fid_to_image(fid,
pid,
image_root=args.image_root,
image_size=pre_crop_size
if args.crop_augment else
net_input_size), # Ergys
num_parallel_calls=args.loading_threads)
if args.flip_augment:
dataset = dataset.map(lambda im, fid, pid:
(tf.image.random_flip_left_right(im), fid, pid))
if args.crop_augment:
dataset = dataset.map(lambda im, fid, pid: (tf.image.random_crop(
im, net_input_size + (3, )), fid, pid))
# Create the model and an embedding head.
tf.keras.backend.set_learning_phase(1)
emb_model = EmbeddingModel(args)
# Group it back into PK batches.
batch_size = args.batch_p * args.batch_k
dataset = dataset.map(lambda im, fid, pid:
(emb_model.preprocess_input(im), fid, pid))
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(1)
# Since we repeat the data infinitely, we only need a one-shot iterator.
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
# all_trainable_variables = embedding_head.trainable_variables+base_model.trainable_variables
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
if 0 <= args.decay_start_iteration < args.train_iterations:
learning_rate = tf.optimizers.schedules.PolynomialDecay(
args.learning_rate, args.train_iterations, end_learning_rate=1e-7)
else:
learning_rate = args.learning_rate
if args.optimizer == 'adam':
optimizer = tf.keras.optimizers.Adam(learning_rate)
elif args.optimizer == 'momentum':
optimizer = tf.keras.optimizers.SGD(learning_rate, momentum=0.9)
else:
raise NotImplementedError('Invalid optimizer {}'.format(
args.optimizer))
@tf.function
def train_step(images, pids):
with tf.GradientTape() as tape:
batch_embedding = emb_model(images)
if args.loss == 'semi_hard_triplet':
embedding_loss = triplet_semihard_loss(batch_embedding, pids,
args.margin)
elif args.loss == 'hard_triplet':
embedding_loss = batch_hard(batch_embedding, pids, args.margin,
args.metric)
elif args.loss == 'lifted_loss':
embedding_loss = lifted_loss(pids,
batch_embedding,
margin=args.margin)
elif args.loss == 'contrastive_loss':
assert batch_size % 2 == 0
assert args.batch_k == 4 ## Can work with other number but will need tuning
contrastive_idx = np.tile([0, 1, 4, 3, 2, 5, 6, 7],
args.batch_p // 2)
for i in range(args.batch_p // 2):
contrastive_idx[i * 8:i * 8 + 8] += i * 8
contrastive_idx = np.expand_dims(contrastive_idx, 1)
batch_embedding_ordered = tf.gather_nd(batch_embedding,
contrastive_idx)
pids_ordered = tf.gather_nd(pids, contrastive_idx)
# batch_embedding_ordered = tf.Print(batch_embedding_ordered,[pids_ordered],'pids_ordered :: ',summarize=1000)
embeddings_anchor, embeddings_positive = tf.unstack(
tf.reshape(batch_embedding_ordered,
[-1, 2, args.embedding_dim]), 2, 1)
# embeddings_anchor = tf.Print(embeddings_anchor,[pids_ordered,embeddings_anchor,embeddings_positive,batch_embedding,batch_embedding_ordered],"Tensors ", summarize=1000)
fixed_labels = np.tile([1, 0, 0, 1], args.batch_p // 2)
# fixed_labels = np.reshape(fixed_labels,(len(fixed_labels),1))
# print(fixed_labels)
labels = tf.constant(fixed_labels)
# labels = tf.Print(labels,[labels],'labels ',summarize=1000)
embedding_loss = contrastive_loss(labels,
embeddings_anchor,
embeddings_positive,
margin=args.margin)
elif args.loss == 'angular_loss':
embeddings_anchor, embeddings_positive = tf.unstack(
tf.reshape(batch_embedding, [-1, 2, args.embedding_dim]),
2, 1)
# pids = tf.Print(pids, [pids], 'pids:: ', summarize=100)
pids, _ = tf.unstack(tf.reshape(pids, [-1, 2, 1]), 2, 1)
# pids = tf.Print(pids,[pids],'pids:: ',summarize=100)
embedding_loss = angular_loss(pids,
embeddings_anchor,
embeddings_positive,
batch_size=args.batch_p,
with_l2reg=True)
elif args.loss == 'npairs_loss':
assert args.batch_k == 2 ## Single positive pair per class
embeddings_anchor, embeddings_positive = tf.unstack(
tf.reshape(batch_embedding, [-1, 2, args.embedding_dim]),
2, 1)
pids, _ = tf.unstack(tf.reshape(pids, [-1, 2, 1]), 2, 1)
pids = tf.reshape(pids, [-1])
embedding_loss = npairs_loss(pids, embeddings_anchor,
embeddings_positive)
else:
raise NotImplementedError('Invalid Loss {}'.format(args.loss))
loss_mean = tf.reduce_mean(embedding_loss)
gradients = tape.gradient(loss_mean, emb_model.trainable_variables)
optimizer.apply_gradients(zip(gradients,
emb_model.trainable_variables))
return embedding_loss
# sess = tf.compat.v1.Session()
# start_step = sess.run(global_step)
# checkpoint_saver = tf.train.Saver(max_to_keep=2)
start_step = 0
log.info('Starting training from iteration {}.'.format(start_step))
dataset_iter = iter(dataset)
ckpt = tf.train.Checkpoint(step=tf.Variable(1),
optimizer=optimizer,
net=emb_model)
manager = tf.train.CheckpointManager(ckpt,
osp.join(args.experiment_root,
'tf_ckpts'),
max_to_keep=3)
ckpt.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
print("Restored from {}".format(manager.latest_checkpoint))
else:
print("Initializing from scratch.")
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(ckpt.step.numpy(), args.train_iterations):
# for batch_idx, batch in enumerate():
start_time = time.time()
images, fids, pids = next(dataset_iter)
batch_loss = train_step(images, pids)
elapsed_time = time.time() - start_time
seconds_todo = (args.train_iterations - i) * elapsed_time
# print(tf.reduce_min(batch_loss).numpy(),tf.reduce_mean(batch_loss).numpy(),tf.reduce_max(batch_loss).numpy())
log.info(
'iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, ETA: {} ({:.2f}s/it)'
.format(
i,
tf.reduce_min(batch_loss).numpy(),
tf.reduce_mean(batch_loss).numpy(),
tf.reduce_max(batch_loss).numpy(),
# args.batch_k - 1, float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)),
elapsed_time))
ckpt.step.assign_add(1)
if (args.checkpoint_frequency > 0
and i % args.checkpoint_frequency == 0):
# uncomment if you want to save the model weight separately
# emb_model.save_weights(os.path.join(args.experiment_root, 'model_weights_{0:04d}.w'.format(i)))
manager.save()
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
def main():
# args = parser.parse_args()
# We store all arguments in a json file. This has two advantages:
# 1. We can always get back and see what exactly that experiment was
# 2. We can resume an experiment as-is without needing to remember all flags.
train_config = cfg.TrainConfig()
args_file = os.path.join(train_config.experiment_root, 'args.json')
if train_config.resume:
if not os.path.isfile(args_file):
raise IOError('`args.json` not found in {}'.format(args_file))
print('Loading args from {}.'.format(args_file))
with open(args_file, 'r') as f:
args_resumed = json.load(f)
args_resumed['resume'] = True # This would be overwritten.
# When resuming, we not only want to populate the args object with the
# values from the file, but we also want to check for some possible
# conflicts between loaded and given arguments.
for key, value in train_config.__dict__.items():
if key in args_resumed:
resumed_value = args_resumed[key]
if resumed_value != value:
print('Warning: For the argument `{}` we are using the'
' loaded value `{}`. The provided value was `{}`'
'.'.format(key, resumed_value, value))
train_config.__dict__[key] = resumed_value
else:
print('Warning: A new argument was added since the last run:'
' `{}`. Using the new value: `{}`.'.format(key, value))
else:
# If the experiment directory exists already, we bail in fear.
if os.path.exists(train_config.experiment_root):
if os.listdir(train_config.experiment_root):
print('The directory {} already exists and is not empty.'
' If you want to resume training, append --resume to'
' your call.'.format(train_config.experiment_root))
exit(1)
else:
os.makedirs(train_config.experiment_root)
# Store the passed arguments for later resuming and grepping in a nice
# and readable format.
with open(args_file, 'w') as f:
json.dump(vars(args), f, ensure_ascii=False, indent=2, sort_keys=True)
log_file = os.path.join(train_config.experiment_root, "train")
logging.config.dictConfig(common.get_logging_dict(log_file))
log = logging.getLogger('train')
# Also show all parameter values at the start, for ease of reading logs.
log.info('Training using the following parameters:')
for key, value in sorted(vars(args).items()):
log.info('{}: {}'.format(key, value))
# Check them here, so they are not required when --resume-ing.
if not train_config.train_set:
parser.print_help()
log.error("You did not specify the `train_set` argument!")
sys.exit(1)
if not train_config.image_root:
parser.print_help()
log.error("You did not specify the required `image_root` argument!")
sys.exit(1)
# Load the data from the CSV file.
pids, fids = common.load_dataset(train_config.train_set, train_config.image_root, is_train=True)
max_fid_len = max(map(len, fids)) # We'll need this later for logfiles
# Setup a tf.Dataset where one "epoch" loops over all PIDS.
# PIDS are shuffled after every epoch and continue indefinitely.
unique_pids = np.unique(pids)
dataset = tf.data.Dataset.from_tensor_slices(unique_pids)
dataset = dataset.shuffle(len(unique_pids))
# Constrain the dataset size to a multiple of the batch-size, so that
# we don't get overlap at the end of each epoch.
dataset = dataset.take((len(unique_pids) // train_config.batch_p) * train_config.batch_p)
# take(count) Creates a Dataset with at most count elements from this dataset.
dataset = dataset.repeat(None) # Repeat forever. Funny way of stating it.
# Repeats this dataset count times.
# For every PID, get K images.
dataset = dataset.map(lambda pid: sample_k_fids_for_pid(
pid, all_fids=fids, all_pids=pids, batch_k=train_config.batch_k))
# Ungroup/flatten the batches for easy loading of the files.
dataset = dataset.apply(tf.contrib.data.unbatch())
# apply(transformation_func) Apply a transformation function to this dataset.
# apply enables chaining of custom Dataset transformations, which are represented as functions that take one Dataset argument and return a transformed Dataset.
# Convert filenames to actual image tensors.
net_input_size = (train_config.net_input_height, train_config.net_input_width)
# 256，128
pre_crop_size = (train_config.pre_crop_height, train_config.pre_crop_width)
# 288，144
dataset = dataset.map(
lambda fid, pid: common.fid_to_image_label(
fid, pid, image_root=train_config.image_root,
image_size=pre_crop_size if train_config.crop_augment else net_input_size),
num_parallel_calls=train_config.loading_threads)
###########################################################################################
dataset = dataset.map(
lambda im, keypt, mask, fid, pid: (tf.concat([im, keypt, mask], 2), fid, pid))
###########################################################################################
# Augment the data if specified by the arguments.
if train_config.flip_augment:
dataset = dataset.map(
lambda im, fid, pid: (tf.image.random_flip_left_right(im), fid, pid))
# net_input_size_aug = net_input_size + (4,)
if train_config.crop_augment:
dataset = dataset.map(
lambda im, fid, pid: (tf.random_crop(im, net_input_size + (21,)), fid, pid))
# net_input_size + (21,) = (256, 128, 21)
# split
#############################################################################################
dataset = dataset.map(
lambda im, fid, pid: (common.split(im, fid, pid)))
#############################################################################################
# Group it back into PK batches.
batch_size = train_config.batch_p * train_config.batch_k
dataset = dataset.batch(batch_size)
# Overlap producing and consuming for parallelism.
dataset = dataset.prefetch(1)
# prefetch(buffer_size) Creates a Dataset that prefetches elements from this dataset.
# Since we repeat the data infinitely, we only need a one-shot iterator.
images, keypts, masks, fids, pids = dataset.make_one_shot_iterator().get_next()
# tf.summary.image('image',images,10)
# Create the model and an embedding head.
model = import_module('nets.' + train_config.model_name)
head = import_module('heads.' + train_config.head_name)
# Feed the image through the model. The returned `body_prefix` will be used
# further down to load the pre-trained weights for all variables with this
# prefix.
endpoints, body_prefix = model.endpoints(images, is_training=True)
heatmap_in = endpoints[train_config.model_name + '/block4']
# resnet_block_4_out = heatmap.resnet_block_4(heatmap_in)
# resnet_block_3_4_out = heatmap.resnet_block_3_4(heatmap_in)
# resnet_block_2_3_4_out = heatmap.resnet_block_2_3_4(heatmap_in)
# head for heatmap
with tf.name_scope('heatmap'):
# heatmap_in = endpoints['model_output']
# heatmap_out_layer_0 = heatmap.hmnet_layer_0(resnet_block_4_out, 1)
# heatmap_out_layer_0 = heatmap.hmnet_layer_0(resnet_block_3_4_out, 1)
# heatmap_out_layer_0 = heatmap.hmnet_layer_0(resnet_block_2_3_4_out, 1)
heatmap_out_layer_0 = VAC.hmnet_layer_0(heatmap_in[:, :, :, 1020:2048], 1)
heatmap_out_layer_1 = VAC.hmnet_layer_1(heatmap_out_layer_0, 1)
heatmap_out_layer_2 = VAC.hmnet_layer_2(heatmap_out_layer_1, 1)
heatmap_out_layer_3 = VAC.hmnet_layer_3(heatmap_out_layer_2, 1)
heatmap_out_layer_4 = VAC.hmnet_layer_4(heatmap_out_layer_3, 1)
heatmap_out = heatmap_out_layer_4
heatmap_loss = VAC.loss_mutilayer(heatmap_out_layer_0, heatmap_out_layer_1, heatmap_out_layer_2,
heatmap_out_layer_3, heatmap_out_layer_4, masks, net_input_size)
# heatmap_loss = heatmap.loss(heatmap_out, labels, net_input_size)
# heatmap_loss_mean = heatmap_loss
with tf.name_scope('head'):
# heatmap_sum = tf.reduce_sum(heatmap_out, axis=3)
# heatmap_resize = tf.image.resize_images(tf.expand_dims(heatmap_sum, axis=3), [8, 4])
# featuremap_tmp = tf.multiply(heatmap_resize, endpoints[args.model_name + '/block4'])
# endpoints[args.model_name + '/block4'] = featuremap_tmp
endpoints = head.head(endpoints, train_config.embedding_dim, is_training=True)
tf.summary.image('feature_map', tf.expand_dims(endpoints[train_config.model_name + '/block4'][:, :, :, 0], axis=3), 4)
with tf.name_scope('keypoints_pre'):
keypoints_pre_in = endpoints[train_config.model_name + '/block4']
# keypoints_pre_in_0 = keypoints_pre_in[:, :, :, 0:256]
# keypoints_pre_in_1 = keypoints_pre_in[:, :, :, 256:512]
# keypoints_pre_in_2 = keypoints_pre_in[:, :, :, 512:768]
# keypoints_pre_in_3 = keypoints_pre_in[:, :, :, 768:1024]
keypoints_pre_in_0 = keypoints_pre_in[:, :, :, 0:170]
keypoints_pre_in_1 = keypoints_pre_in[:, :, :, 170:340]
keypoints_pre_in_2 = keypoints_pre_in[:, :, :, 340:510]
keypoints_pre_in_3 = keypoints_pre_in[:, :, :, 510:680]
keypoints_pre_in_4 = keypoints_pre_in[:, :, :, 680:850]
keypoints_pre_in_5 = keypoints_pre_in[:, :, :, 850:1020]
labels = tf.image.resize_images(keypts, [128, 64])
# keypoints_gt_0 = tf.concat([labels[:, :, :, 0:5], labels[:, :, :, 14:15], labels[:, :, :, 15:16], labels[:, :, :, 16:17], labels[:, :, :, 17:18]], 3)
# keypoints_gt_1 = tf.concat([labels[:, :, :, 1:2], labels[:, :, :, 2:3], labels[:, :, :, 3:4], labels[:, :, :, 5:6]], 3)
# keypoints_gt_2 = tf.concat([labels[:, :, :, 4:5], labels[:, :, :, 7:8], labels[:, :, :, 8:9], labels[:, :, :, 11:12]], 3)
# keypoints_gt_3 = tf.concat([labels[:, :, :, 9:10], labels[:, :, :, 10:11], labels[:, :, :, 12:13], labels[:, :, :, 13:14]], 3)
keypoints_gt_0 = labels[:, :, :, 0:5]
keypoints_gt_1 = labels[:, :, :, 5:7]
keypoints_gt_2 = labels[:, :, :, 7:9]
keypoints_gt_3 = labels[:, :, :, 9:13]
keypoints_gt_4 = labels[:, :, :, 13:15]
keypoints_gt_5 = labels[:, :, :, 15:17]
keypoints_pre_0 = PAC.tran_conv_0(keypoints_pre_in, kp_num=5)
keypoints_pre_1 = PAC.tran_conv_1(keypoints_pre_in, kp_num=2)
keypoints_pre_2 = PAC.tran_conv_2(keypoints_pre_in, kp_num=2)
keypoints_pre_3 = PAC.tran_conv_3(keypoints_pre_in, kp_num=4)
keypoints_pre_4 = PAC.tran_conv_4(keypoints_pre_in, kp_num=2)
keypoints_pre_5 = PAC.tran_conv_5(keypoints_pre_in, kp_num=2)
keypoints_loss_0 = PAC.keypoints_loss(keypoints_pre_0, keypoints_gt_0)
keypoints_loss_1 = PAC.keypoints_loss(keypoints_pre_1, keypoints_gt_1)
keypoints_loss_2 = PAC.keypoints_loss(keypoints_pre_2, keypoints_gt_2)
keypoints_loss_3 = PAC.keypoints_loss(keypoints_pre_3, keypoints_gt_3)
keypoints_loss_4 = PAC.keypoints_loss(keypoints_pre_4, keypoints_gt_4)
keypoints_loss_5 = PAC.keypoints_loss(keypoints_pre_5, keypoints_gt_5)
keypoints_loss = 5/17*keypoints_loss_0 + 2/17*keypoints_loss_1 + 2/17*keypoints_loss_2 + 4/17*keypoints_loss_3 + 2/17*keypoints_loss_4 + 2/17*keypoints_loss_5
# Create the loss in two steps:
# 1. Compute all pairwise distances according to the specified metric.
# 2. For each anchor along the first dimension, compute its loss.
dists = loss.cdist(endpoints['emb'], endpoints['emb'], metric=train_config.metric)
losses, train_top1, prec_at_k, _, neg_dists, pos_dists = loss.LOSS_CHOICES[train_config.loss](
dists, pids, train_config.margin, batch_precision_at_k=train_config.batch_k-1)
# Count the number of active entries, and compute the total batch loss.
num_active = tf.reduce_sum(tf.cast(tf.greater(losses, 1e-5), tf.float32))
loss_mean = tf.reduce_mean(losses)
scale_rate_0 = 1E-7
scale_rate_1 = 6E-8
total_loss = loss_mean + keypoints_loss*scale_rate_0 + heatmap_loss*scale_rate_1
# total_loss = loss_mean + keypoints_loss * scale_rate_0
# total_loss = loss_mean
# Some logging for tensorboard.
tf.summary.histogram('loss_distribution', losses)
tf.summary.scalar('loss', loss_mean)
############################################################################################
# tf.summary.histogram('hm_loss_distribution', heatmap_loss)
tf.summary.scalar('keypt_loss_0', keypoints_loss_0)
tf.summary.scalar('keypt_loss_1', keypoints_loss_1)
tf.summary.scalar('keypt_loss_2', keypoints_loss_2)
tf.summary.scalar('keypt_loss_3', keypoints_loss_3)
tf.summary.scalar('keypt_loss_all', keypoints_loss)
############################################################################################
tf.summary.scalar('total_loss', total_loss)
tf.summary.scalar('batch_top1', train_top1)
tf.summary.scalar('batch_prec_at_{}'.format(args.batch_k-1), prec_at_k)
tf.summary.scalar('active_count', num_active)
tf.summary.histogram('embedding_dists', dists)
tf.summary.histogram('embedding_pos_dists', pos_dists)
tf.summary.histogram('embedding_neg_dists', neg_dists)
tf.summary.histogram('embedding_lengths',
tf.norm(endpoints['emb_raw'], axis=1))
# Create the mem-mapped arrays in which we'll log all training detail in
# addition to tensorboard, because tensorboard is annoying for detailed
# inspection and actually discards data in histogram summaries.
if args.detailed_logs:
log_embs = lb.create_or_resize_dat(
os.path.join(train_config.experiment_root, 'embeddings'),
dtype=np.float32, shape=(train_config.train_iterations, batch_size, args.embedding_dim))
log_loss = lb.create_or_resize_dat(
os.path.join(train_config.experiment_root, 'losses'),
dtype=np.float32, shape=(train_config.train_iterations, batch_size))
log_fids = lb.create_or_resize_dat(
os.path.join(train_config.experiment_root, 'fids'),
dtype='S' + str(max_fid_len), shape=(train_config.train_iterations, batch_size))
# These are collected here before we add the optimizer, because depending
# on the optimizer, it might add extra slots, which are also global
# variables, with the exact same prefix.
model_variables = tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, body_prefix)
# Define the optimizer and the learning-rate schedule.
# Unfortunately, we get NaNs if we don't handle no-decay separately.
global_step = tf.Variable(0, name='global_step', trainable=False)
if 0 <= train_config.decay_start_iteration < train_config.train_iterations:
learning_rate = tf.train.exponential_decay(
train_config.learning_rate,
tf.maximum(0, global_step - train_config.decay_start_iteration),
train_config.train_iterations - train_config.decay_start_iteration, 0.001)
else:
learning_rate = train_config.learning_rate
tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
# Feel free to try others!
# optimizer = tf.train.AdadeltaOptimizer(learning_rate)
# Update_ops are used to update batchnorm stats.
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
# train_op = optimizer.minimize(loss_mean, global_step=global_step)
train_op = optimizer.minimize(total_loss, global_step=global_step)
#
# Define a saver for the complete model.
checkpoint_saver = tf.train.Saver(max_to_keep=0)
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
if train_config.resume:
# In case we're resuming, simply load the full checkpoint to init.
last_checkpoint = tf.train.latest_checkpoint(args.experiment_root)
log.info('Restoring from checkpoint: {}'.format(last_checkpoint))
checkpoint_saver.restore(sess, last_checkpoint)
else:
# But if we're starting from scratch, we may need to load some
# variables from the pre-trained weights, and random init others.
sess.run(tf.global_variables_initializer())
if train_config.initial_checkpoint is not None:
saver = tf.train.Saver(model_variables, write_version=tf.train.SaverDef.V1)
saver.restore(sess, train_config.initial_checkpoint)
# name_11 = 'resnet_v1_50/block4'
# name_12 = 'resnet_v1_50/block3'
# name_13 = 'resnet_v1_50/block2'
# name_21 = 'Resnet_block_2_3_4/block4'
# name_22 = 'Resnet_block_2_3_4/block3'
# name_23 = 'Resnet_block_2_3_4/block2'
# for var in tf.trainable_variables():
# var_name = var.name
# if re.match(name_11, var_name):
# dst_name = var_name.replace(name_11, name_21)
# tensor = tf.get_default_graph().get_tensor_by_name(var_name)
# dst_tensor = tf.get_default_graph().get_tensor_by_name(dst_name)
# tf.assign(dst_tensor, tensor)
# if re.match(name_12, var_name):
# dst_name = var_name.replace(name_12, name_22)
# tensor = tf.get_default_graph().get_tensor_by_name(var_name)
# dst_tensor = tf.get_default_graph().get_tensor_by_name(dst_name)
# tf.assign(dst_tensor, tensor)
# if re.match(name_13, var_name):
# dst_name = var_name.replace(name_13, name_23)
# tensor = tf.get_default_graph().get_tensor_by_name(var_name)
# dst_tensor = tf.get_default_graph().get_tensor_by_name(dst_name)
# tf.assign(dst_tensor, tensor)
# In any case, we also store this initialization as a checkpoint,
# such that we could run exactly reproduceable experiments.
checkpoint_saver.save(sess, os.path.join(
train_config.experiment_root, 'checkpoint'), global_step=0)
merged_summary = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(train_config.experiment_root, sess.graph)
start_step = sess.run(global_step)
log.info('Starting training from iteration {}.'.format(start_step))
# Finally, here comes the main-loop. This `Uninterrupt` is a handy
# utility such that an iteration still finishes on Ctrl+C and we can
# stop the training cleanly.
with lb.Uninterrupt(sigs=[SIGINT, SIGTERM], verbose=True) as u:
for i in range(start_step, train_config.train_iterations):
# Compute gradients, update weights, store logs!
start_time = time.time()
_, summary, step, b_prec_at_k, b_embs, b_loss, b_fids = \
sess.run([train_op, merged_summary, global_step,
prec_at_k, endpoints['emb'], losses, fids])
elapsed_time = time.time() - start_time
# Compute the iteration speed and add it to the summary.
# We did observe some weird spikes that we couldn't track down.
summary2 = tf.Summary()
summary2.value.add(tag='secs_per_iter', simple_value=elapsed_time)
summary_writer.add_summary(summary2, step)
summary_writer.add_summary(summary, step)
if train_config.detailed_logs:
log_embs[i], log_loss[i], log_fids[i] = b_embs, b_loss, b_fids
# Do a huge print out of the current progress.
seconds_todo = (train_config.train_iterations - step) * elapsed_time
log.info('iter:{:6d}, loss min|avg|max: {:.3f}|{:.3f}|{:6.3f}, '
'batch-p@{}: {:.2%}, ETA: {} ({:.2f}s/it)'.format(
step,
float(np.min(b_loss)),
float(np.mean(b_loss)),
float(np.max(b_loss)),
train_config.batch_k-1, float(b_prec_at_k),
timedelta(seconds=int(seconds_todo)),
elapsed_time))
sys.stdout.flush()
sys.stderr.flush()
# Save a checkpoint of training every so often.
if (train_config.checkpoint_frequency > 0 and
step % train_config.checkpoint_frequency == 0):
checkpoint_saver.save(sess, os.path.join(
train_config.experiment_root, 'checkpoint'), global_step=step)
# Stop the main-loop at the end of the step, if requested.
if u.interrupted:
log.info("Interrupted on request!")
break
# Store one final checkpoint. This might be redundant, but it is crucial
# in case intermediate storing was disabled and it saves a checkpoint
# when the process was interrupted.
checkpoint_saver.save(sess, os.path.join(
train_config.experiment_root, 'checkpoint'), global_step=step)
