def __init__(self, imdb):
self.model_path = "trained_models/music/RefineNet-Res101/semseg"
self.model_name = "RefineNet-Res101"
self.saved_net = 'backbone'
# has to be adjusted according to the training scheme used
self.energy_loss = "softmax"
self.class_loss = "softmax"
self.bbox_loss = "reg"
self.tf_session = None
self.root_dir = cfg.ROOT_DIR
self.sess = tf.Session()
print('Loading model')
self.input = tf.placeholder(tf.float32, shape=[None, None, None, 1])
self.network_heads, self.init_fn = build_dwd_net(
self.input,
model=self.model_name,
num_classes=imdb.num_classes,
pretrained_dir="",
substract_mean=False)
self.saver = tf.train.Saver(max_to_keep=1000)
self.sess.run(tf.global_variables_initializer())
print("Loading weights")
self.saver.restore(
self.sess,
self.root_dir + "/" + self.model_path + "/" + self.saved_net)
self.tf_session = self.sess
def __init__(self, parsed, path, imdb):
self.model_path = path
self.config = parsed
self.imdb = imdb
self.saved_net = "backbone"
self.tf_session = None
self.sess = tf.Session()
print('Loading model')
if "DeepScores" not in self.model_path:
self.input = tf.placeholder(tf.float32,
shape=[None, None, None, 3])
else:
self.input = tf.placeholder(tf.float32,
shape=[None, None, None, 1])
print("Initializing Model:" + self.config.model)
used_heads = set()
self.used_heads_loss = []
for _, assign in enumerate(self.config.training_assignements):
used_heads.add(assign["stamp_func"])
self.used_heads_loss.append(
[assign["stamp_func"], assign["stamp_args"]["loss"]])
self.used_heads = list(used_heads)
self.used_heads.sort()
# for now:
#self.used_heads = ['stamp_class', 'stamp_bbox', 'stamp_energy']
#self.used_heads = ['stamp_class', 'stamp_energy', 'stamp_bbox']
#self.used_heads = ['stamp_bbox', 'stamp_class', 'stamp_energy']
#self.used_heads = ['stamp_bbox', 'stamp_energy', 'stamp_class']
#self.used_heads = ['stamp_energy', 'stamp_class', 'stamp_bbox']
#self.used_heads = ['stamp_energy', 'stamp_bbox', 'stamp_class']
self.network_heads, init_fn = build_dwd_net(
self.input,
model=parsed.model,
num_classes=len(self.imdb._classes),
pretrained_dir="",
max_energy=parsed.max_energy,
substract_mean=False,
individual_upsamp=self.config.individual_upsamp,
paired_mode=self.config.paired_data,
used_heads=self.used_heads,
sparse_heads=parsed.sparse_heads)
self.saver = tf.train.Saver(max_to_keep=1000)
# self.sess.run(tf.global_variables_initializer())
# self.sess.run(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='deep_watershed/energy_logits_pair0_3'))
print("Loading weights")
self.saver.restore(self.sess, self.model_path + "/" + self.saved_net)
self.tf_session = self.sess
self.counter = 0
def __init__(self):
np.random.seed(self.seed)
tf.set_random_seed(self.seed)
print('Loading mappings ...')
mapping = pa.read_csv(self.root_dir + "/" + self.mapping_name)
mapping = mapping[mapping["Init_name"] != "-1"]
mapping = mapping.drop(mapping.columns[[0, 2]], axis=1)
# strip values
mapping['Init_name'] = mapping['Init_name'].str.strip()
# reset index
self.mapping = mapping.reset_index(drop=True)
# has to be adjusted according to the training scheme used
self.energy_loss = "softmax"
self.class_loss = "softmax"
self.bbox_loss = "reg"
sess = tf.Session()
print('Loading model')
self.input = tf.placeholder(tf.float32,
shape=[None, None, None, 1],
name='image_input')
dws_heads, init_fn = build_dwd_net(self.input,
model=self.model_name,
num_classes=len(self.mapping),
pretrained_dir="",
substract_mean=False)
[self.dws_energy, self.class_logits, self.bbox_size] = [
dws_heads["stamp_energy"][self.energy_loss][-1],
dws_heads["stamp_class"][self.class_loss][-1],
dws_heads["stamp_bbox"][self.bbox_loss][-1]
]
# Name the tensors
saver = tf.train.Saver(max_to_keep=1000)
sess.run(tf.global_variables_initializer())
print("Loading weights")
saver.restore(
sess,
self.root_dir + "/" + self.model_path + "/" + self.model_name)
self.tf_session = sess
def __init__(self, imdb, path, pa, individual_upsamp=False):
self.model_path = path
self.model_name = pa.net_type
self.saved_net = pa.saved_net
# has to be adjusted according to the training scheme used
self.energy_loss = pa.energy_loss
self.class_loss = pa.class_loss
self.bbox_loss = pa.bbox_loss
self.tf_session = None
self.root_dir = cfg.ROOT_DIR
self.sess = tf.Session()
print('Loading model')
if "realistic" in self.model_path:
self.input = tf.placeholder(tf.float32,
shape=[None, None, None, 3])
else:
self.input = tf.placeholder(tf.float32,
shape=[None, None, None, 1])
self.network_heads, self.init_fn = build_dwd_net(
self.input,
model=self.model_name,
num_classes=imdb.num_classes,
pretrained_dir="",
substract_mean=False,
individual_upsamp=individual_upsamp)
self.saver = tf.train.Saver(max_to_keep=1000)
self.sess.run(tf.global_variables_initializer())
print("Loading weights")
self.saver.restore(
self.sess,
self.root_dir + "/" + self.model_path + "/" + self.saved_net)
self.tf_session = self.sess
self.counter = 0
def main(parsed):
args = parsed[0]
print(args)
iteration = 1
np.random.seed(cfg.RNG_SEED)
# load database
imdb, roidb, imdb_val, roidb_val, data_layer, data_layer_val = load_database(
args)
global nr_classes
nr_classes = len(imdb._classes)
args.nr_classes.append(nr_classes)
# replaces keywords with function handles in training assignements
save_objectness_function_handles(args, imdb)
# Debug stuffs
#
# try_all_assign(data_layer,args,100)
# dws_list = perform_dws(data["dws_energy"], data["class_map"], data["bbox_fcn"])
#
# def show_image(data, gt_boxes=None):
# from PIL import Image, ImageDraw
# if gt_boxes is None:
# im = Image.fromarray(data[0].astype("uint8"))
# im.show()
# else:
# im = Image.fromarray(data[0].astype("uint8"))
# draw = ImageDraw.Draw(im)
# # overlay GT boxes
# for row in gt_boxes[0]:
# draw.rectangle(((row[0], row[1]), (row[2], row[3])), fill="red")
# im.show()
# return
#
# fetched_dat = data_layer.forward(args, [args.training_assignements[0]], None)
# show_image(fetched_dat["data"],fetched_dat["gt_boxes"])
# show_image(fetched_dat["data"], None)
# tensorflow session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# input and output tensors
if "DeepScores" in args.dataset:
input = tf.placeholder(tf.float32, shape=[None, None, None, 1])
resnet_dir = cfg.PRETRAINED_DIR + "/DeepScores/"
refinenet_dir = cfg.PRETRAINED_DIR + "/DeepScores_semseg/"
elif "MUSICMA" in args.dataset:
input = tf.placeholder(tf.float32, shape=[None, None, None, 1])
resnet_dir = cfg.PRETRAINED_DIR + "/DeepScores/"
refinenet_dir = cfg.PRETRAINED_DIR + "/DeepScores_semseg/"
else:
input = tf.placeholder(tf.float32, shape=[None, None, None, 3])
resnet_dir = cfg.PRETRAINED_DIR + "/ImageNet/"
refinenet_dir = cfg.PRETRAINED_DIR + "/VOC2012/"
if not (len(args.training_help) == 1 and args.training_help[0] is None):
# initialize helper_input
helper_input = tf.placeholder(
tf.float32, shape=[None, None, None, input.shape[-1] + 1])
feed_head = slim.conv2d(helper_input,
input.shape[-1], [3, 3],
scope='gt_feed_head')
input = feed_head
print("Initializing Model:" + args.model)
# model has all possible output heads (even if unused) to ensure saving and loading goes smoothly
network_heads, init_fn = build_dwd_net(input,
model=args.model,
num_classes=nr_classes,
pretrained_dir=resnet_dir,
substract_mean=False)
# initialize tasks
preped_assign = []
for assign in args.training_assignements:
[
loss, optim, gt_placeholders, scalar_summary_op, images_summary_op,
images_placeholders, mask_placholders
] = initialize_assignement(assign, imdb, network_heads, sess,
data_layer, input, args)
preped_assign.append([
loss, optim, gt_placeholders, scalar_summary_op, images_summary_op,
images_placeholders, mask_placholders
])
# init tensorflow session
saver = tf.train.Saver(max_to_keep=1000)
sess.run(tf.global_variables_initializer())
# load model weights
checkpoint_dir = get_checkpoint_dir(args)
checkpoint_name = "backbone"
if args.continue_training == "True":
print("Loading checkpoint")
saver.restore(sess, checkpoint_dir + "/" + checkpoint_name)
elif args.pretrain_lvl == "deepscores_to_musicma":
pretrained_vars = []
for var in slim.get_model_variables():
if not ("class_pred" in var.name):
pretrained_vars.append(var)
print("Loading network pretrained on Deepscores for Muscima")
loading_checkpoint_name = cfg.PRETRAINED_DIR + "/DeepScores_to_Muscima/" + "backbone"
init_fn = slim.assign_from_checkpoint_fn(loading_checkpoint_name,
pretrained_vars)
init_fn(sess)
else:
if args.pretrain_lvl == "semseg":
#load all variables except the ones in scope "deep_watershed"
pretrained_vars = []
for var in slim.get_model_variables():
if not ("deep_watershed" in var.name
or "gt_feed_head" in var.name):
pretrained_vars.append(var)
print("Loading network pretrained on semantic segmentation")
loading_checkpoint_name = refinenet_dir + args.model + ".ckpt"
init_fn = slim.assign_from_checkpoint_fn(loading_checkpoint_name,
pretrained_vars)
init_fn(sess)
elif args.pretrain_lvl == "class":
print("Loading pretrained weights for level: " + args.pretrain_lvl)
init_fn(sess)
else:
print("Not loading a pretrained network")
# set up tensorboard
writer = tf.summary.FileWriter(checkpoint_dir, sess.graph)
# execute tasks
for do_a in args.do_assign:
assign_nr = do_a["assign"]
do_itr = do_a["Itrs"]
training_help = args.training_help[do_a["help"]]
iteration = execute_assign(args, input, saver, sess, checkpoint_dir,
checkpoint_name, data_layer, writer,
network_heads, do_itr,
args.training_assignements[assign_nr],
preped_assign[assign_nr], iteration,
training_help)
# execute combined tasks
for do_comb_a in args.combined_assignements:
#iteration = execute_combined_assign(do_comb_a)
do_comb_itr = do_comb_a["Itrs"]
rm_length = do_comb_a["Running_Mean_Length"]
loss_factors = do_comb_a["loss_factors"]
orig_assign = [
args.training_assignements[i] for i in do_comb_a["assigns"]
]
preped_assigns = [preped_assign[i] for i in do_comb_a["assigns"]]
training_help = None # unused atm
execute_combined_assign(args, data_layer, training_help, orig_assign,
preped_assigns, loss_factors, do_comb_itr,
iteration, input, rm_length, network_heads,
sess, checkpoint_dir, checkpoint_name, saver,
writer)
print("done :)")
def main(parsed):
args = parsed[0]
print(args)
iteration = 1
np.random.seed(cfg.RNG_SEED)
# load database
imdb, roidb, imdb_val, roidb_val, data_layer, data_layer_val = load_database(
args)
global nr_classes
nr_classes = len(imdb._classes)
args.nr_classes.append(nr_classes)
# replaces keywords with function handles in training assignements
save_objectness_function_handles(args, imdb)
# tensorflow session
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# input and output tensors
if "DeepScores_300dpi" in args.dataset:
input = tf.placeholder(tf.float32, shape=[None, None, None, 1])
resnet_dir = cfg.PRETRAINED_DIR + "/DeepScores/"
refinenet_dir = cfg.PRETRAINED_DIR + "/DeepScores_semseg/"
elif "DeepScores" in args.dataset:
input = tf.placeholder(tf.float32, shape=[None, None, None, 1])
resnet_dir = cfg.PRETRAINED_DIR + "/DeepScores/"
refinenet_dir = cfg.PRETRAINED_DIR + "/DeepScores_semseg/"
elif "MUSICMA" in args.dataset:
input = tf.placeholder(tf.float32, shape=[None, None, None, 1])
resnet_dir = cfg.PRETRAINED_DIR + "/DeepScores/"
refinenet_dir = cfg.PRETRAINED_DIR + "/DeepScores_semseg/"
else:
input = tf.placeholder(tf.float32, shape=[None, None, None, 3])
resnet_dir = cfg.PRETRAINED_DIR + "/ImageNet/"
refinenet_dir = cfg.PRETRAINED_DIR + "/VOC2012/"
if not (len(args.training_help) == 1 and args.training_help[0] is None):
# initialize helper_input
helper_input = tf.placeholder(
tf.float32, shape=[None, None, None, input.shape[-1] + 1])
feed_head = slim.conv2d(helper_input,
input.shape[-1], [3, 3],
scope='gt_feed_head')
input = feed_head
print("Initializing Model:" + args.model)
# model has all possible output heads (even if unused) to ensure saving and loading goes smoothly
network_heads, init_fn = build_dwd_net(
input,
model=args.model,
num_classes=nr_classes,
pretrained_dir=resnet_dir,
substract_mean=False,
individual_upsamp=args.individual_upsamp)
# use just one image summary OP for all tasks
final_pred_placeholder = tf.placeholder(tf.uint8, shape=[1, None, None, 3])
images_sums = []
images_placeholders = []
images_placeholders.append(final_pred_placeholder)
images_sums.append(
tf.summary.image('DWD_debug_img', final_pred_placeholder))
images_summary_op = tf.summary.merge(images_sums)
# initialize tasks
preped_assign = []
for assign in args.training_assignements:
[loss, optim, gt_placeholders, scalar_summary_op,
mask_placholders] = initialize_assignement(assign, imdb,
network_heads, sess,
data_layer, input, args)
preped_assign.append([
loss, optim, gt_placeholders, scalar_summary_op, images_summary_op,
images_placeholders, mask_placholders
])
# init tensorflow session
saver = tf.train.Saver(max_to_keep=1000)
sess.run(tf.global_variables_initializer())
# load model weights
checkpoint_dir = get_checkpoint_dir(args)
checkpoint_name = "backbone"
if args.continue_training == "True":
print("Loading checkpoint")
saver.restore(sess, checkpoint_dir + "/" + checkpoint_name)
elif args.pretrain_lvl == "deepscores_to_musicma":
pretrained_vars = []
for var in slim.get_model_variables():
if not ("class_pred" in var.name):
pretrained_vars.append(var)
print("Loading network pretrained on Deepscores for Muscima")
loading_checkpoint_name = cfg.PRETRAINED_DIR + "/DeepScores_to_Muscima/" + "backbone"
init_fn = slim.assign_from_checkpoint_fn(loading_checkpoint_name,
pretrained_vars)
init_fn(sess)
elif args.pretrain_lvl == "DeepScores_to_300dpi":
pretrained_vars = []
for var in slim.get_model_variables():
if not ("class_pred" in var.name):
pretrained_vars.append(var)
print("Loading network pretrained on Deepscores for Muscima")
loading_checkpoint_name = cfg.PRETRAINED_DIR + "/DeepScores_to_300dpi/" + "backbone"
init_fn = slim.assign_from_checkpoint_fn(loading_checkpoint_name,
pretrained_vars)
init_fn(sess)
else:
if args.pretrain_lvl == "semseg":
# load all variables except the ones in scope "deep_watershed"
pretrained_vars = []
for var in slim.get_model_variables():
if not ("deep_watershed" in var.name
or "gt_feed_head" in var.name):
pretrained_vars.append(var)
print("Loading network pretrained on semantic segmentation")
loading_checkpoint_name = refinenet_dir + args.model + ".ckpt"
init_fn = slim.assign_from_checkpoint_fn(loading_checkpoint_name,
pretrained_vars)
init_fn(sess)
elif args.pretrain_lvl == "class":
print("Loading pretrained weights for level: " + args.pretrain_lvl)
init_fn(sess)
else:
print("Not loading a pretrained network")
# set up tensorboard
writer = tf.summary.FileWriter(checkpoint_dir, sess.graph)
# execute tasks
for do_a in args.do_assign:
assign_nr = do_a["assign"]
do_itr = do_a["Itrs"]
training_help = args.training_help[do_a["help"]]
iteration = execute_assign(args, input, saver, sess, checkpoint_dir,
checkpoint_name, data_layer, writer,
network_heads, do_itr,
args.training_assignements[assign_nr],
preped_assign[assign_nr], iteration,
training_help)
# execute combined tasks
for do_comb_a in args.combined_assignements:
do_comb_itr = do_comb_a["Itrs"]
rm_length = do_comb_a["Running_Mean_Length"]
loss_factors = do_comb_a["loss_factors"]
orig_assign = [
args.training_assignements[i] for i in do_comb_a["assigns"]
]
preped_assigns = [preped_assign[i] for i in do_comb_a["assigns"]]
training_help = None # unused atm
execute_combined_assign(args, data_layer, training_help, orig_assign,
preped_assigns, loss_factors, do_comb_itr,
iteration, input, rm_length, network_heads,
sess, checkpoint_dir, checkpoint_name, saver,
writer)
print("done :)")