def train_net(net,
train_path,
num_classes,
batch_size,
data_shape,
mean_pixels,
resume,
finetune,
pretrained,
epoch,
prefix,
ctx,
begin_epoch,
end_epoch,
frequent,
learning_rate,
momentum,
weight_decay,
lr_refactor_step,
lr_refactor_ratio,
freeze_layer_pattern='',
shape_range=(320, 512),
random_shape_step=0,
random_shape_epoch=10,
num_example=10000,
label_pad_width=350,
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
nms_topk=400,
force_suppress=False,
train_list="",
val_path="",
val_list="",
iter_monitor=0,
monitor_pattern=".*",
log_file=None):
"""
Wrapper for training phase.
Parameters:
----------
net : str
symbol name for the network structure
train_path : str
record file path for training
num_classes : int
number of object classes, not including background
batch_size : int
training batch-size
data_shape : int or tuple
width/height as integer or (3, height, width) tuple
mean_pixels : tuple of floats
mean pixel values for red, green and blue
resume : int
resume from previous checkpoint if > 0
finetune : int
fine-tune from previous checkpoint if > 0
pretrained : str
prefix of pretrained model, including path
epoch : int
load epoch of either resume/finetune/pretrained model
prefix : str
prefix for saving checkpoints
ctx : [mx.cpu()] or [mx.gpu(x)]
list of mxnet contexts
begin_epoch : int
starting epoch for training, should be 0 if not otherwise specified
end_epoch : int
end epoch of training
frequent : int
frequency to print out training status
learning_rate : float
training learning rate
momentum : float
trainig momentum
weight_decay : float
training weight decay param
lr_refactor_ratio : float
multiplier for reducing learning rate
lr_refactor_step : comma separated integers
at which epoch to rescale learning rate, e.g. '30, 60, 90'
freeze_layer_pattern : str
regex pattern for layers need to be fixed
shape_range : tuple of (min, max)
random data shape range
random_shape_step : int
step size for random data shape, defined by network, 0 to disable
random_step_epoch : int
number of epoch before next random shape
num_example : int
number of training images
label_pad_width : int
force padding training and validation labels to sync their label widths
nms_thresh : float
non-maximum suppression threshold for validation
force_nms : boolean
suppress overlaped objects from different classes
train_list : str
list file path for training, this will replace the embeded labels in record
val_path : str
record file path for validation
val_list : str
list file path for validation, this will replace the embeded labels in record
iter_monitor : int
monitor internal stats in networks if > 0, specified by monitor_pattern
monitor_pattern : str
regex pattern for monitoring network stats
log_file : str
log to file if enabled
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if log_file:
fh = logging.FileHandler(log_file)
logger.addHandler(fh)
# check args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
prefix += '_' + net.strip('_yolo') + '_' + str(data_shape[1])
if isinstance(mean_pixels, (int, float)):
mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
assert len(mean_pixels) == 3, "must provide all RGB mean values"
# load symbol
sys.path.append(os.path.join(cfg.ROOT_DIR, 'symbol'))
symbol_module = importlib.import_module("symbol_" + net)
net = symbol_module.get_symbol(num_classes,
nms_thresh=nms_thresh,
force_suppress=force_suppress,
nms_topk=nms_topk)
# define layers with fixed weight/bias
if freeze_layer_pattern.strip():
re_prog = re.compile(freeze_layer_pattern)
fixed_param_names = [
name for name in net.list_arguments() if re_prog.match(name)
]
else:
fixed_param_names = None
# load pretrained or resume from previous state
ctx_str = '(' + ','.join([str(c) for c in ctx]) + ')'
allow_missing = True
if resume > 0:
logger.info("Resume training with {} from epoch {}".format(
ctx_str, resume))
_, args, auxs = mx.model.load_checkpoint(prefix, resume)
begin_epoch = resume
allow_missing = False
elif finetune > 0:
logger.info("Start finetuning with {} from epoch {}".format(
ctx_str, finetune))
_, args, auxs = mx.model.load_checkpoint(prefix, finetune)
begin_epoch = finetune
# the prediction convolution layers name starts with relu, so it's fine
fixed_param_names = [name for name in net.list_arguments() \
if name.startswith('conv')]
elif pretrained:
logger.info("Start training with {} from pretrained model {}".format(
ctx_str, pretrained))
_, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
args = convert_pretrained(pretrained, args)
else:
logger.info("Experimental: start training from scratch with {}".format(
ctx_str))
args = None
auxs = None
fixed_param_names = None
# helper information
if fixed_param_names:
logger.info("Freezed parameters: [" + ','.join(fixed_param_names) +
']')
# fit parameters
batch_end_callback = mx.callback.Speedometer(batch_size, frequent=frequent)
epoch_end_callback = mx.callback.do_checkpoint(prefix)
monitor = mx.mon.Monitor(
iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
# run fit net, every n epochs we run evaluation network to get mAP
if voc07_metric:
valid_metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=0)
else:
valid_metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=0)
# init training module
mod = mx.mod.Module(net,
label_names=('yolo_output_label', ),
logger=logger,
context=ctx,
fixed_param_names=fixed_param_names)
random_shape_step = int(random_shape_step)
if random_shape_step > 0:
fit_begins = list(range(begin_epoch, end_epoch, random_shape_epoch))
fit_ends = fit_begins[1:] + [end_epoch]
assert (len(shape_range) == 2)
data_shapes = [(3, x * random_shape_step, x * random_shape_step) \
for x in range(shape_range[0] // random_shape_step,
shape_range[1] // random_shape_step + 1)]
logger.info("Candidate random shapes:" + str(data_shapes))
else:
fit_begins = [begin_epoch]
fit_ends = [end_epoch]
data_shapes = [data_shape]
for begin, end in zip(fit_begins, fit_ends):
if len(data_shapes) == 1:
data_shape = data_shapes[0]
else:
data_shape = data_shapes[random.randint(0, len(data_shapes) - 1)]
logger.info("Setting random data shape: " + str(data_shape))
train_iter = DetRecordIter(train_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=train_list,
**cfg.train)
if val_path:
val_iter = DetRecordIter(val_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=val_list,
**cfg.valid)
else:
val_iter = None
learning_rate, lr_scheduler = get_lr_scheduler(learning_rate,
lr_refactor_step,
lr_refactor_ratio,
num_example, batch_size,
begin_epoch)
optimizer_params = {
'learning_rate': learning_rate,
'momentum': momentum,
'wd': weight_decay,
'lr_scheduler': lr_scheduler,
'clip_gradient': 10,
'rescale_grad': 1.0
}
mod.fit(train_iter,
val_iter,
eval_metric=MultiBoxMetric(),
validation_metric=valid_metric,
batch_end_callback=batch_end_callback,
epoch_end_callback=epoch_end_callback,
optimizer='sgd',
optimizer_params=optimizer_params,
begin_epoch=begin,
num_epoch=end,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=allow_missing,
monitor=monitor,
force_rebind=True,
force_init=True)
args, auxs = mod.get_params()
allow_missing = False
def train_net(net, train_path, num_classes, batch_size,
data_shape, mean_pixels, resume, finetune, pretrained, epoch,
prefix, ctx, begin_epoch, end_epoch, frequent, learning_rate,
momentum, weight_decay, lr_refactor_step, lr_refactor_ratio,
freeze_layer_pattern='',
num_example=10000, label_pad_width=350,
nms_thresh=0.45, force_nms=False, ovp_thresh=0.5,
use_difficult=False, class_names=None,
voc07_metric=False, nms_topk=400, force_suppress=False,
train_list="", val_path="", val_list="", iter_monitor=0,
monitor_pattern=".*", log_file=None):
"""
Wrapper for training phase.
Parameters:
----------
net : str
symbol name for the network structure
train_path : str
record file path for training
num_classes : int
number of object classes, not including background
batch_size : int
training batch-size
data_shape : int or tuple
width/height as integer or (3, height, width) tuple
mean_pixels : tuple of floats
mean pixel values for red, green and blue
resume : int
resume from previous checkpoint if > 0
finetune : int
fine-tune from previous checkpoint if > 0
pretrained : str
prefix of pretrained model, including path
epoch : int
load epoch of either resume/finetune/pretrained model
prefix : str
prefix for saving checkpoints
ctx : [mx.cpu()] or [mx.gpu(x)]
list of mxnet contexts
begin_epoch : int
starting epoch for training, should be 0 if not otherwise specified
end_epoch : int
end epoch of training
frequent : int
frequency to print out training status
learning_rate : float
training learning rate
momentum : float
trainig momentum
weight_decay : float
training weight decay param
lr_refactor_ratio : float
multiplier for reducing learning rate
lr_refactor_step : comma separated integers
at which epoch to rescale learning rate, e.g. '30, 60, 90'
freeze_layer_pattern : str
regex pattern for layers need to be fixed
num_example : int
number of training images
label_pad_width : int
force padding training and validation labels to sync their label widths
nms_thresh : float
non-maximum suppression threshold for validation
force_nms : boolean
suppress overlaped objects from different classes
train_list : str
list file path for training, this will replace the embeded labels in record
val_path : str
record file path for validation
val_list : str
list file path for validation, this will replace the embeded labels in record
iter_monitor : int
monitor internal stats in networks if > 0, specified by monitor_pattern
monitor_pattern : str
regex pattern for monitoring network stats
log_file : str
log to file if enabled
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if log_file:
fh = logging.FileHandler(log_file)
logger.addHandler(fh)
# check args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
if prefix.endswith('_'):
prefix += '_' + str(data_shape[1])
if isinstance(mean_pixels, (int, float)):
mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
assert len(mean_pixels) == 3, "must provide all RGB mean values"
train_iter = DetRecordIter(train_path, batch_size, data_shape, mean_pixels=mean_pixels,
label_pad_width=label_pad_width, path_imglist=train_list, **cfg.train)
if val_path:
val_iter = DetRecordIter(val_path, batch_size, data_shape, mean_pixels=mean_pixels,
label_pad_width=label_pad_width, path_imglist=val_list, **cfg.valid)
else:
val_iter = None
# load symbol
net = get_symbol_train(net, data_shape[1], num_classes=num_classes,
nms_thresh=nms_thresh, force_suppress=force_suppress, nms_topk=nms_topk)
# define layers with fixed weight/bias
if freeze_layer_pattern.strip():
re_prog = re.compile(freeze_layer_pattern)
fixed_param_names = [name for name in net.list_arguments() if re_prog.match(name)]
else:
fixed_param_names = None
# load pretrained or resume from previous state
ctx_str = '('+ ','.join([str(c) for c in ctx]) + ')'
if resume > 0:
logger.info("Resume training with {} from epoch {}"
.format(ctx_str, resume))
_, args, auxs = mx.model.load_checkpoint(prefix, resume)
begin_epoch = resume
elif finetune > 0:
logger.info("Start finetuning with {} from epoch {}"
.format(ctx_str, finetune))
_, args, auxs = mx.model.load_checkpoint(prefix, finetune)
begin_epoch = finetune
# check what layers mismatch with the loaded parameters
exe = net.simple_bind(mx.cpu(), data=(1, 3, 300, 300), label=(1, 1, 5), grad_req='null')
arg_dict = exe.arg_dict
fixed_param_names = []
for k, v in arg_dict.items():
if k in args:
if v.shape != args[k].shape:
del args[k]
logging.info("Removed %s" % k)
else:
if not 'pred' in k:
fixed_param_names.append(k)
elif pretrained:
logger.info("Start training with {} from pretrained model {}"
.format(ctx_str, pretrained))
_, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
args = convert_pretrained(pretrained, args)
else:
logger.info("Experimental: start training from scratch with {}"
.format(ctx_str))
args = None
auxs = None
fixed_param_names = None
# helper information
if fixed_param_names:
logger.info("Freezed parameters: [" + ','.join(fixed_param_names) + ']')
# init training module
mod = mx.mod.Module(net, label_names=('label',), logger=logger, context=ctx,
fixed_param_names=fixed_param_names)
# fit parameters
batch_end_callback = mx.callback.Speedometer(train_iter.batch_size, frequent=frequent)
epoch_end_callback = mx.callback.do_checkpoint(prefix)
learning_rate, lr_scheduler = get_lr_scheduler(learning_rate, lr_refactor_step,
lr_refactor_ratio, num_example, batch_size, begin_epoch)
optimizer_params={'learning_rate':learning_rate,
'momentum':momentum,
'wd':weight_decay,
'lr_scheduler':lr_scheduler,
'clip_gradient':None,
'rescale_grad': 1.0 / len(ctx) if len(ctx) > 0 else 1.0 }
monitor = mx.mon.Monitor(iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
# run fit net, every n epochs we run evaluation network to get mAP
if voc07_metric:
valid_metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=3)
else:
valid_metric = MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=3)
mod.fit(train_iter,
val_iter,
eval_metric=MultiBoxMetric(),
validation_metric=valid_metric,
batch_end_callback=batch_end_callback,
epoch_end_callback=epoch_end_callback,
optimizer='sgd',
optimizer_params=optimizer_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=True,
monitor=monitor)
def train_net(network,
train_path,
num_classes,
batch_size,
data_shape,
mean_pixels,
resume,
finetune,
pretrained,
epoch,
prefix,
ctx,
begin_epoch,
end_epoch,
frequent,
learning_rate,
momentum,
weight_decay,
lr_refactor_step,
lr_refactor_ratio,
freeze_layer_pattern='',
num_example=10000,
label_pad_width=350,
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
nms_topk=400,
force_suppress=False,
train_list="",
val_path="",
val_list="",
iter_monitor=0,
monitor_pattern=".*",
log_file=None,
optimizer='sgd',
tensorboard=False,
checkpoint_period=5,
min_neg_samples=0):
"""
Wrapper for training phase.
Parameters:
----------
net : str
symbol name for the network structure
train_path : str
record file path for training
num_classes : int
number of object classes, not including background
batch_size : int
training batch-size
data_shape : int or tuple
width/height as integer or (3, height, width) tuple
mean_pixels : tuple of floats
mean pixel values for red, green and blue
resume : int
resume from previous checkpoint if > 0
finetune : int
fine-tune from previous checkpoint if > 0
pretrained : str
prefix of pretrained model, including path
epoch : int
load epoch of either resume/finetune/pretrained model
prefix : str
prefix for saving checkpoints
ctx : [mx.cpu()] or [mx.gpu(x)]
list of mxnet contexts
begin_epoch : int
starting epoch for training, should be 0 if not otherwise specified
end_epoch : int
end epoch of training
frequent : int
frequency to print out training status
optimizer : str
usage of different optimizers, other then default sgd
learning_rate : float
training learning rate
momentum : float
trainig momentum
weight_decay : float
training weight decay param
lr_refactor_ratio : float
multiplier for reducing learning rate
lr_refactor_step : comma separated integers
at which epoch to rescale learning rate, e.g. '30, 60, 90'
freeze_layer_pattern : str
regex pattern for layers need to be fixed
num_example : int
number of training images
label_pad_width : int
force padding training and validation labels to sync their label widths
nms_thresh : float
non-maximum suppression threshold for validation
force_nms : boolean
suppress overlaped objects from different classes
train_list : str
list file path for training, this will replace the embeded labels in record
val_path : str
record file path for validation
val_list : str
list file path for validation, this will replace the embeded labels in record
iter_monitor : int
monitor internal stats in networks if > 0, specified by monitor_pattern
monitor_pattern : str
regex pattern for monitoring network stats
log_file : str
log to file if enabled
tensorboard : bool
record logs into tensorboard
min_neg_samples : int
always have some negative examples, no matter how many positive there are.
this is useful when training on images with no ground-truth.
checkpoint_period : int
a checkpoint will be saved every "checkpoint_period" epochs
"""
# check actual number of train_images
if os.path.exists(train_path.replace('rec', 'idx')):
with open(train_path.replace('rec', 'idx'), 'r') as f:
txt = f.readlines()
num_example = len(txt)
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if log_file:
log_file_path = os.path.join(os.path.dirname(prefix), log_file)
if not os.path.exists(os.path.dirname(log_file_path)):
os.makedirs(os.path.dirname(log_file_path))
fh = logging.FileHandler(log_file_path)
logger.addHandler(fh)
# check args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
if prefix.endswith('_'):
prefix += '_' + str(data_shape[1])
if isinstance(mean_pixels, (int, float)):
mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
assert len(mean_pixels) == 3, "must provide all RGB mean values"
train_iter = DetRecordIter(train_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=train_list,
**cfg.train)
if val_path:
val_iter = DetRecordIter(val_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=val_list,
**cfg.valid)
else:
val_iter = None
# load symbol
net = get_symbol_train(network,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_suppress,
nms_topk=nms_topk,
minimum_negative_samples=min_neg_samples)
# define layers with fixed weight/bias
if freeze_layer_pattern.strip():
re_prog = re.compile(freeze_layer_pattern)
fixed_param_names = [
name for name in net.list_arguments() if re_prog.match(name)
]
else:
fixed_param_names = None
# load pretrained or resume from previous state
ctx_str = '(' + ','.join([str(c) for c in ctx]) + ')'
if resume > 0:
logger.info("Resume training with {} from epoch {}".format(
ctx_str, resume))
_, args, auxs = mx.model.load_checkpoint(prefix, resume)
begin_epoch = resume
elif finetune > 0:
logger.info("Start finetuning with {} from epoch {}".format(
ctx_str, finetune))
# _, args, auxs = mx.model.load_checkpoint(prefix, finetune)
# Gang Chen changed
_, args, auxs = mx.model.load_checkpoint(pretrained, finetune)
begin_epoch = finetune
# check what layers mismatch with the loaded parameters
exe = net.simple_bind(mx.cpu(),
data=(1, 3, 300, 300),
label=(1, 1, 5),
grad_req='null')
arg_dict = exe.arg_dict
fixed_param_names = []
for k, v in arg_dict.items():
if k in args:
if v.shape != args[k].shape:
del args[k]
logging.info("Removed %s" % k)
else:
if not 'pred' in k:
#Gang Chen
#fixed_param_names.append(k)
pass
elif pretrained and not finetune:
logger.info("Start training with {} from pretrained model {}".format(
ctx_str, pretrained))
_, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
args = convert_pretrained(pretrained, args)
else:
logger.info("Experimental: start training from scratch with {}".format(
ctx_str))
args = None
auxs = None
fixed_param_names = None
# helper information
if fixed_param_names:
logger.info("Freezed parameters: [" + ','.join(fixed_param_names) +
']')
# visualize net - both train and test
net_visualization(net=net,
network=network,
data_shape=data_shape[2],
output_dir=os.path.dirname(prefix),
train=True)
net_visualization(net=None,
network=network,
data_shape=data_shape[2],
output_dir=os.path.dirname(prefix),
train=False,
num_classes=num_classes)
# init training module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=fixed_param_names)
batch_end_callback = []
eval_end_callback = []
epoch_end_callback = [
mx.callback.do_checkpoint(prefix, period=checkpoint_period)
]
# add logging to tensorboard
if tensorboard:
tensorboard_dir = os.path.join(os.path.dirname(prefix), 'logs')
if not os.path.exists(tensorboard_dir):
os.makedirs(os.path.join(tensorboard_dir, 'train', 'scalar'))
os.makedirs(os.path.join(tensorboard_dir, 'train', 'dist'))
os.makedirs(os.path.join(tensorboard_dir, 'val', 'roc'))
os.makedirs(os.path.join(tensorboard_dir, 'val', 'scalar'))
os.makedirs(os.path.join(tensorboard_dir, 'val', 'images'))
batch_end_callback.append(
ParseLogCallback(
dist_logging_dir=os.path.join(tensorboard_dir, 'train',
'dist'),
scalar_logging_dir=os.path.join(tensorboard_dir, 'train',
'scalar'),
logfile_path=log_file_path,
batch_size=batch_size,
iter_monitor=iter_monitor,
frequent=frequent))
eval_end_callback.append(
mx.contrib.tensorboard.LogMetricsCallback(
os.path.join(tensorboard_dir, 'val/scalar'), 'ssd'))
eval_end_callback.append(
LogROCCallback(logging_dir=os.path.join(tensorboard_dir,
'val/roc'),
roc_path=os.path.join(os.path.dirname(prefix),
'roc'),
class_names=class_names))
eval_end_callback.append(
LogDetectionsCallback(
logging_dir=os.path.join(tensorboard_dir, 'val/images'),
images_path=os.path.join(os.path.dirname(prefix), 'images'),
class_names=class_names,
batch_size=batch_size,
mean_pixels=mean_pixels))
# this callback should be the last in a serie of batch_callbacks
# since it is resetting the metric evaluation every $frequent batches
batch_end_callback.append(
mx.callback.Speedometer(train_iter.batch_size, frequent=frequent))
learning_rate, lr_scheduler = get_lr_scheduler(learning_rate,
lr_refactor_step,
lr_refactor_ratio,
num_example, batch_size,
begin_epoch)
# add possibility for different optimizer
opt, opt_params = get_optimizer_params(optimizer=optimizer,
learning_rate=learning_rate,
momentum=momentum,
weight_decay=weight_decay,
lr_scheduler=lr_scheduler,
ctx=ctx,
logger=logger)
# TODO monitor the gradient flow as in 'https://github.com/dmlc/tensorboard/blob/master/docs/tutorial/understanding-vanish-gradient.ipynb'
monitor = mx.mon.Monitor(
iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
# run fit net, every n epochs we run evaluation network to get mAP
if voc07_metric:
valid_metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=3,
roc_output_path=os.path.join(
os.path.dirname(prefix), 'roc'))
else:
valid_metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=3,
roc_output_path=os.path.join(
os.path.dirname(prefix), 'roc'))
mod.fit(train_iter,
val_iter,
eval_metric=MultiBoxMetric(),
validation_metric=valid_metric,
batch_end_callback=batch_end_callback,
eval_end_callback=eval_end_callback,
epoch_end_callback=epoch_end_callback,
optimizer=opt,
optimizer_params=opt_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=True,
monitor=monitor)
arg_params=arg_params,
aux_params=aux_params,
ctx=ctx,
excluded_sym_names=excluded_sym_names,
calib_mode=calib_mode,
quantized_dtype=args.quantized_dtype,
logger=logger)
sym_name = '%s-symbol.json' % ('./model/qssd_vgg16_reduced_300')
param_name = '%s-%04d.params' % ('./model/qssd_vgg16_reduced_300',
epoch)
save_symbol(sym_name, qsym, logger)
else:
logger.info('Creating ImageRecordIter for reading calibration dataset')
eval_iter = DetRecordIter(os.path.join(os.getcwd(), 'data', 'val.rec'),
batch_size,
data_shape,
mean_pixels=(123, 117, 104),
path_imglist="",
**cfg.valid)
qsym, qarg_params, aux_params = quantize_model(
sym=sym,
arg_params=arg_params,
aux_params=aux_params,
ctx=ctx,
excluded_sym_names=excluded_sym_names,
calib_mode=calib_mode,
calib_data=eval_iter,
num_calib_examples=num_calib_batches * batch_size,
calib_layer=calib_layer,
quantized_dtype=args.quantized_dtype,
label_names=(label_name, ),
def evaluate_net(net,
path_imgrec,
num_classes,
mean_pixels,
data_shape,
model_prefix,
epoch,
ctx=mx.cpu(),
batch_size=1,
path_imglist="",
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False):
"""
evalute network given validation record file
Parameters:
----------
net : str or None
Network name or use None to load from json without modifying
path_imgrec : str
path to the record validation file
path_imglist : str
path to the list file to replace labels in record file, optional
num_classes : int
number of classes, not including background
mean_pixels : tuple
(mean_r, mean_g, mean_b)
data_shape : tuple or int
(3, height, width) or height/width
model_prefix : str
model prefix of saved checkpoint
epoch : int
load model epoch
ctx : mx.ctx
mx.gpu() or mx.cpu()
batch_size : int
validation batch size
nms_thresh : float
non-maximum suppression threshold
force_nms : boolean
whether suppress different class objects
ovp_thresh : float
AP overlap threshold for true/false postives
use_difficult : boolean
whether to use difficult objects in evaluation if applicable
class_names : comma separated str
class names in string, must correspond to num_classes if set
voc07_metric : boolean
whether to use 11-point evluation as in VOC07 competition
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
model_prefix += '_' + str(data_shape[1])
# iterator
eval_iter = DetRecordIter(path_imgrec,
batch_size,
data_shape,
mean_pixels=mean_pixels,
path_imglist=path_imglist,
**cfg.valid)
# model params
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
# network
if net is None:
net = load_net
else:
net = get_symbol(net,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_nms)
if not 'label' in net.list_arguments():
label = mx.sym.Variable(name='label')
net = mx.sym.Group([net, label])
# init module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=net.list_arguments())
mod.bind(data_shapes=eval_iter.provide_data,
label_shapes=eval_iter.provide_label)
mod.set_params(args, auxs, allow_missing=False, force_init=True)
# run evaluation
if voc07_metric:
metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names)
else:
metric = MApMetric(ovp_thresh, use_difficult, class_names)
results = mod.score(eval_iter, metric, num_batch=None)
for k, v in results:
print("{}: {}".format(k, v))
def evaluate_net(net,
path_imgrec,
num_classes,
num_batch,
mean_pixels,
data_shape,
model_prefix,
epoch,
ctx=mx.cpu(),
batch_size=32,
path_imglist="",
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
lite=False):
"""
evalute network given validation record file
Parameters:
----------
net : str or None
Network name or use None to load from json without modifying
path_imgrec : str
path to the record validation file
path_imglist : str
path to the list file to replace labels in record file, optional
num_classes : int
number of classes, not including background
mean_pixels : tuple
(mean_r, mean_g, mean_b)
data_shape : tuple or int
(3, height, width) or height/width
model_prefix : str
model prefix of saved checkpoint
epoch : int
load model epoch
ctx : mx.ctx
mx.gpu() or mx.cpu()
batch_size : int
validation batch size
nms_thresh : float
non-maximum suppression threshold
force_nms : boolean
whether suppress different class objects
ovp_thresh : float
AP overlap threshold for true/false postives
use_difficult : boolean
whether to use difficult objects in evaluation if applicable
class_names : comma separated str
class names in string, must correspond to num_classes if set
voc07_metric : boolean
whether to use 11-point evluation as in VOC07 competition
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
model_prefix += '_' + str(data_shape[1]) + '_' + str(data_shape[2])
# iterator
eval_iter = DetRecordIter(path_imgrec,
batch_size,
data_shape,
mean_pixels=mean_pixels,
path_imglist=path_imglist,
**cfg.valid)
# model params
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
# network
if net is None:
net = load_net
else:
net = get_symbol(net,
data_shape,
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_nms,
lite=lite)
if not 'label' in net.list_arguments():
label = mx.sym.Variable(name='label')
net = mx.sym.Group([net, label])
# init module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=net.list_arguments())
mod.bind(data_shapes=eval_iter.provide_data,
label_shapes=eval_iter.provide_label)
mod.set_params(args, auxs, allow_missing=False, force_init=True)
# run evaluation
if voc07_metric:
metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names)
else:
metric = MApMetric(ovp_thresh, use_difficult, class_names)
num = num_batch * batch_size
data = [
mx.random.uniform(-1.0, 1.0, shape=shape, ctx=ctx)
for _, shape in mod.data_shapes
]
batch = mx.io.DataBatch(data, []) # empty label
dry_run = 5 # use 5 iterations to warm up
for i in range(dry_run):
mod.forward(batch, is_train=False)
for output in mod.get_outputs():
output.wait_to_read()
tic = time.time()
results = mod.score(eval_iter,
metric,
num_batch=None,
batch_end_callback=mx.callback.Speedometer(
batch_size, frequent=10, auto_reset=False))
speed = num / (time.time() - tic)
if logger is not None:
logger.info('Finished inference with %d images' % num)
logger.info('Finished with %f images per second', speed)
for k, v in results:
print("{}: {}".format(k, v))
def evaluate_net(net,
path_imgrec,
num_classes,
mean_pixels,
data_shape,
model_prefix,
epoch,
ctx=mx.cpu(),
batch_size=1,
path_imglist="",
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
frequent=20):
"""
evalute network given validation record file
Parameters:
----------
net : str or None
Network name or use None to load from json without modifying
path_imgrec : str
path to the record validation file
path_imglist : str
path to the list file to replace labels in record file, optional
num_classes : int
number of classes, not including background
mean_pixels : tuple
(mean_r, mean_g, mean_b)
data_shape : tuple or int
(3, height, width) or height/width
model_prefix : str
model prefix of saved checkpoint
epoch : int
load model epoch
ctx : mx.ctx
mx.gpu() or mx.cpu()
batch_size : int
validation batch size
nms_thresh : float
non-maximum suppression threshold
force_nms : boolean
whether suppress different class objects
ovp_thresh : float
AP overlap threshold for true/false postives
use_difficult : boolean
whether to use difficult objects in evaluation if applicable
class_names : comma separated str
class names in string, must correspond to num_classes if set
voc07_metric : boolean
whether to use 11-point evluation as in VOC07 competition
frequent : int
frequency to print out validation status
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
#model_prefix += '_' + str(data_shape[1])
# iterator
eval_iter = DetRecordIter(path_imgrec,
batch_size,
data_shape,
path_imglist=path_imglist,
**cfg.valid)
# model params
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
# network
if net is None:
net = load_net
else:
net = get_symbol(net,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_nms)
if not 'label' in net.list_arguments():
label = mx.sym.Variable(name='label')
net = mx.sym.Group([net, label])
data_shape = (1, 3, 300, 300)
mx.viz.plot_network_detail(net,
shape={
"data": data_shape,
"label": (1, 1, 5)
},
node_attrs={
"hide_weights": "true",
"fixedsize": 'false',
"shape": 'oval'
}).view()
# Gang Chen add
exe = net.simple_bind(mx.cpu(),
data=(1, 3, 300, 300),
label=(1, 1, 5),
grad_req='null')
arg_dict = exe.arg_dict
for k, v in args.items():
if k not in arg_dict:
del args[k]
# END Gang Chen add
# init module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=net.list_arguments())
mod.bind(data_shapes=eval_iter.provide_data,
label_shapes=eval_iter.provide_label)
mod.set_params(args, auxs, allow_missing=True, force_init=True)
# run evaluation
if voc07_metric:
metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
roc_output_path=os.path.join(
os.path.dirname(model_prefix), 'roc'))
else:
metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
roc_output_path=os.path.join(
os.path.dirname(model_prefix), 'roc'))
results = mod.score(eval_iter,
metric,
num_batch=None,
batch_end_callback=mx.callback.Speedometer(
batch_size, frequent=frequent, auto_reset=False))
for k, v in results:
print("{}: {}".format(k, v))
def evaluate_net(net, path_imgrec, num_classes, mean_img, data_shape,
model_prefix, epoch, path_img, ctx=mx.cpu(), batch_size=1,
path_imglist="", nms_thresh=0.45, force_nms=False,
ovp_thresh=0.5, use_difficult=False, class_names=None,
voc07_metric=False):
"""
evalute network given validation record file
Parameters:
----------
net : str or None
Network name or use None to load from json without modifying
path_imgrec : str
path to the record validation file
path_imglist : str
path to the list file to replace labels in record file, optional
num_classes : int
number of classes, not including background
mean_pixels : tuple
(mean_r, mean_g, mean_b)
data_shape : tuple or int
(3, height, width) or height/width
model_prefix : str
model prefix of saved checkpoint
epoch : int
load model epoch
ctx : mx.ctx
mx.gpu() or mx.cpu()
batch_size : int
validation batch size
nms_thresh : float
non-maximum suppression threshold
force_nms : boolean
whether suppress different class objects
ovp_thresh : float
AP overlap threshold for true/false postives
use_difficult : boolean
whether to use difficult objects in evaluation if applicable
class_names : comma separated str
class names in string, must correspond to num_classes if set
voc07_metric : boolean
whether to use 11-point evluation as in VOC07 competition
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
model_prefix += '_' + str(data_shape[1])
netname = net
# iterator
eval_iter = DetRecordIter(path_imgrec, batch_size, data_shape, mean_img=mean_img,
path_imglist=path_imglist, **cfg.valid)
# model params
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
# network
if net is None:
net = load_net
else:
net = get_symbol(net, data_shape[1], num_classes=num_classes,
nms_thresh=nms_thresh, force_suppress=force_nms)
if not 'label' in net.list_arguments():
label = mx.sym.Variable(name='label')
net = mx.sym.Group([net, label])
# init module
mod = mx.mod.Module(net, label_names=('label',), logger=logger, context=ctx,
fixed_param_names=net.list_arguments())
mod.bind(data_shapes=eval_iter.provide_data, label_shapes=eval_iter.provide_label)
mod.set_params(args, auxs, allow_missing=False, force_init=True)
# # run evaluation
# if voc07_metric:
# metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names)
# else:
# metric = MApMetric(ovp_thresh, use_difficult, class_names)
# results = mod.score(eval_iter, metric, num_batch=None)
# for k, v in results:
# print("{}: {}".format(k, v))
predict_results = mod.predict(eval_iter, merge_batches = True)
preds = predict_results[0]
labels = predict_results[1]
(flags, ious) = find_wrong_detection.find_wrong_detection(labels, preds, path_imglist,
path_img, ovp_thresh = ovp_thresh)
flags_dict = {0:'correct', 1:'lower iou', 2:'wrong class'}
flag_count = Counter(flags)
for flag in set(flags):
print ("%s image number is : %d"%(flags_dict[flag], flag_count[flag]))
print ("recall is %f"%((len(flags)-flag_count[1])/float(len(flags))))
if not os.path.exists('./model/iou_distribution'):
os.mkdir('./model/iou_distribution')
xmin = min(ious) - 0.1 if min(ious) > 0.1 else 0
xmax = max(ious) + 0.1 if max(ious) < 0.9 else 1
draw_hist(ious, "iou distribution", "iou", "image number", xmin, xmax, 0, len(ious)/20, netname)
def train_net(net,
train_path,
num_classes,
batch_size,
data_shape,
mean_pixels,
resume,
finetune,
pretrained,
epoch,
prefix,
ctx,
begin_epoch,
end_epoch,
frequent,
learning_rate,
momentum,
weight_decay,
use_plateau,
lr_refactor_step,
lr_refactor_ratio,
use_global_stats=0,
freeze_layer_pattern='',
num_example=10000,
label_pad_width=350,
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
ignore_names=None,
optimizer_name='sgd',
voc07_metric=False,
nms_topk=400,
force_suppress=False,
train_list="",
val_path="",
val_list="",
iter_monitor=0,
monitor_pattern=".*",
log_file=None):
"""
Wrapper for training phase.
Parameters:
----------
net : str
symbol name for the network structure
train_path : str
record file path for training
num_classes : int
number of object classes, not including background
batch_size : int
training batch-size
data_shape : int or tuple
width/height as integer or (3, height, width) tuple
mean_pixels : tuple of floats
mean pixel values for red, green and blue
resume : int
resume from previous checkpoint if > 0
finetune : int
fine-tune from previous checkpoint if > 0
pretrained : str
prefix of pretrained model, including path
epoch : int
load epoch of either resume/finetune/pretrained model
prefix : str
prefix for saving checkpoints
ctx : [mx.cpu()] or [mx.gpu(x)]
list of mxnet contexts
begin_epoch : int
starting epoch for training, should be 0 if not otherwise specified
end_epoch : int
end epoch of training
frequent : int
frequency to print out training status
learning_rate : float
training learning rate
momentum : float
trainig momentum
weight_decay : float
training weight decay param
lr_refactor_ratio : float
multiplier for reducing learning rate
lr_refactor_step : comma separated integers
at which epoch to rescale learning rate, e.g. '30, 60, 90'
freeze_layer_pattern : str
regex pattern for layers need to be fixed
num_example : int
number of training images
label_pad_width : int
force padding training and validation labels to sync their label widths
nms_thresh : float
non-maximum suppression threshold for validation
force_nms : boolean
suppress overlaped objects from different classes
train_list : str
list file path for training, this will replace the embeded labels in record
val_path : str
record file path for validation
val_list : str
list file path for validation, this will replace the embeded labels in record
iter_monitor : int
monitor internal stats in networks if > 0, specified by monitor_pattern
monitor_pattern : str
regex pattern for monitoring network stats
log_file : str
log to file if enabled
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if log_file:
fh = logging.FileHandler(log_file)
logger.addHandler(fh)
# check args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
prefix += '_' + net + '_' + str(data_shape[1])
if isinstance(mean_pixels, (int, float)):
mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
assert len(mean_pixels) == 3, "must provide all RGB mean values"
train_iter = DetRecordIter(train_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=train_list,
**cfg.train)
if val_path:
val_iter = DetRecordIter(val_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=val_list,
**cfg.valid)
else:
val_iter = None
# load symbol
net_str = net
net = get_symbol_train(net, data_shape[1], \
use_global_stats=use_global_stats, \
num_classes=num_classes, ignore_names=ignore_names, \
nms_thresh=nms_thresh, force_suppress=force_suppress, nms_topk=nms_topk)
# define layers with fixed weight/bias
if freeze_layer_pattern.strip():
re_prog = re.compile(freeze_layer_pattern)
fixed_param_names = [
name for name in net.list_arguments() if re_prog.match(name)
]
else:
fixed_param_names = None
# load pretrained or resume from previous state
ctx_str = '(' + ','.join([str(c) for c in ctx]) + ')'
if resume > 0:
logger.info("Resume training with {} from epoch {}".format(
ctx_str, resume))
_, args, auxs = mx.model.load_checkpoint(prefix, resume)
begin_epoch = resume
elif finetune > 0:
logger.info("Start finetuning with {} from epoch {}".format(
ctx_str, finetune))
_, args, auxs = mx.model.load_checkpoint(prefix, finetune)
begin_epoch = finetune
# the prediction convolution layers name starts with relu, so it's fine
fixed_param_names = [name for name in net.list_arguments() \
if name.startswith('conv')]
elif pretrained:
try:
logger.info(
"Start training with {} from pretrained model {}".format(
ctx_str, pretrained))
_, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
args = convert_pretrained(pretrained, args)
if net_str == 'ssd_pva':
args, auxs = convert_pvanet(args, auxs)
except:
logger.info(
"Failed to load the pretrained model. Start from scratch.")
args = None
auxs = None
fixed_param_names = None
else:
logger.info("Experimental: start training from scratch with {}".format(
ctx_str))
args = None
auxs = None
fixed_param_names = None
# helper information
if fixed_param_names:
logger.info("Freezed parameters: [" + ','.join(fixed_param_names) +
']')
# init training module
if not use_plateau: # focal loss does not go well with plateau
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=fixed_param_names)
else:
mod = PlateauModule(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=fixed_param_names)
# robust parameter setting
mod.bind(data_shapes=train_iter.provide_data,
label_shapes=train_iter.provide_label)
mod = set_mod_params(mod, args, auxs, logger)
# fit parameters
batch_end_callback = mx.callback.Speedometer(train_iter.batch_size,
frequent=frequent,
auto_reset=True)
epoch_end_callback = mx.callback.do_checkpoint(prefix)
monitor = mx.mon.Monitor(
iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
optimizer_params = {
'learning_rate': learning_rate,
'wd': weight_decay,
'clip_gradient': 4.0,
'rescale_grad': 1.0 / len(ctx) if len(ctx) > 0 else 1.0
}
if optimizer_name == 'sgd':
optimizer_params['momentum'] = momentum
# #7847
mod.init_optimizer(optimizer=optimizer_name,
optimizer_params=optimizer_params,
force_init=True)
if not use_plateau:
learning_rate, lr_scheduler = get_lr_scheduler(learning_rate,
lr_refactor_step,
lr_refactor_ratio,
num_example, batch_size,
begin_epoch)
else:
w_l1 = cfg.train['smoothl1_weight']
eval_weights = {
'CrossEntropy': 1.0,
'SmoothL1': w_l1,
'ObjectRecall': 0.0
}
plateau_lr = PlateauScheduler( \
patient_epochs=lr_refactor_step, factor=float(lr_refactor_ratio), eval_weights=eval_weights)
plateau_metric = MultiBoxMetric(
fn_stat='/home/hyunjoon/github/additions_mxnet/ssd/stat.txt')
mod.init_optimizer(optimizer=optimizer_name,
optimizer_params=optimizer_params)
eval_metric = MultiBoxMetric()
# run fit net, every n epochs we run evaluation network to get mAP
if voc07_metric:
map_metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=4)
recall_metric = RecallMetric(ovp_thresh, use_difficult, pred_idx=4)
valid_metric = mx.metric.create([map_metric, recall_metric])
else:
valid_metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=4)
if not use_plateau:
mod.fit(train_iter,
eval_data=val_iter,
eval_metric=eval_metric,
validation_metric=valid_metric,
batch_end_callback=batch_end_callback,
epoch_end_callback=epoch_end_callback,
optimizer=optimizer_name,
optimizer_params=optimizer_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=True,
monitor=monitor)
else:
mod.fit(train_iter,
plateau_lr,
plateau_metric=plateau_metric,
fn_curr_model=prefix + '-1000.params',
plateau_backtrace=False,
eval_data=val_iter,
eval_metric=eval_metric,
validation_metric=valid_metric,
validation_period=5,
kvstore='local',
batch_end_callback=batch_end_callback,
epoch_end_callback=epoch_end_callback,
optimizer=optimizer_name,
optimizer_params=optimizer_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=True,
monitor=monitor)
def evaluate_net(net,
path_imgrec,
num_classes,
mean_pixels,
data_shape,
model_prefix,
epoch,
ctx=mx.cpu(),
batch_size=1,
path_imglist="",
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
use_second_network=False,
net1=None,
path_imgrec1=None,
epoch1=None,
model_prefix1=None,
data_shape1=None):
"""
evalute network given validation record file
Parameters:
----------
net : str or None
Network name or use None to load from json without modifying
path_imgrec : str
path to the record validation file
path_imglist : str
path to the list file to replace labels in record file, optional
num_classes : int
number of classes, not including background
mean_pixels : tuple
(mean_r, mean_g, mean_b)
data_shape : tuple or int
(3, height, width) or height/width
model_prefix : str
model prefix of saved checkpoint
epoch : int
load model epoch
ctx : mx.ctx
mx.gpu() or mx.cpu()
batch_size : int
validation batch size
nms_thresh : float
non-maximum suppression threshold
force_nms : boolean
whether suppress different class objects
ovp_thresh : float
AP overlap threshold for true/false postives
use_difficult : boolean
whether to use difficult objects in evaluation if applicable
class_names : comma separated str
class names in string, must correspond to num_classes if set
voc07_metric : boolean
whether to use 11-point evluation as in VOC07 competition
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
elif isinstance(data_shape, list):
data_shape = (3, data_shape[0], data_shape[1])
assert len(data_shape) == 3 and data_shape[0] == 3
# model_prefix += '_' + str(data_shape[1])
# iterator
#eval_iter = DetRecordIter(path_imgrec, batch_size, data_shape,
# path_imglist=path_imglist, **cfg.valid)
curr_path = os.path.abspath(os.path.dirname(__file__))
imdb_val = load_caltech(image_set='val',
caltech_path=os.path.join(
curr_path, '..', 'data',
'caltech-pedestrian-dataset-converter'),
shuffle=False)
eval_iter = DetIter(imdb_val, batch_size, (data_shape[1], data_shape[2]), \
mean_pixels=[128, 128, 128], rand_samplers=[], \
rand_mirror=False, shuffle=False, rand_seed=None, \
is_train=True, max_crop_trial=50)
# model params
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
# network
if net is None:
net = load_net
else:
#net = get_symbol(net, data_shape[1], num_classes=num_classes,
net = get_symbol_concat(net,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_nms)
if not 'label' in net.list_arguments():
label = mx.sym.Variable(name='label')
label2 = mx.sym.Variable(name='label2')
net = mx.sym.Group([net, label, label2])
# init module
#mod = mx.mod.Module(net, label_names=('label',), logger=logger, context=ctx,
mod = mx.mod.Module(net,
label_names=('label', 'label2'),
logger=logger,
context=ctx,
fixed_param_names=net.list_arguments())
mod.bind(data_shapes=eval_iter.provide_data,
label_shapes=eval_iter.provide_label)
mod.set_params(args, auxs, allow_missing=False, force_init=True)
if voc07_metric:
#metric = VOC07MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=1)
metric = VOC07MApMetric(
ovp_thresh,
use_difficult,
class_names,
pred_idx=[0, 1],
output_names=['detection_output', 'detection2_output'],
label_names=['label', 'label2'])
else:
#metric = MApMetric(ovp_thresh, use_difficult, class_names, pred_idx=1)
metric = MApMetric(
ovp_thresh,
use_difficult,
class_names,
pred_idx=[0, 1],
output_names=['detection_output', 'detection2_output'],
label_names=['label', 'label2'])
# run evaluation
if not use_second_network:
results = mod.score(eval_iter, metric, num_batch=None)
for k, v in results:
print("{}: {}".format(k, v))
else:
logging.basicConfig()
logger1 = logging.getLogger()
logger1.setLevel(logging.INFO)
# load sub network
if isinstance(data_shape1, int):
data_shape1 = (3, data_shape1, data_shape1)
elif isinstance(data_shape1, list):
data_shape1 = (3, data_shape1[0], data_shape1[1])
assert len(data_shape1) == 3 and data_shape1[0] == 3
# iterator
eval_iter1 = DetRecordIter(path_imgrec1,
batch_size,
data_shape1,
path_imglist=path_imglist,
**cfg.valid)
# model params
load_net1, args1, auxs1 = mx.model.load_checkpoint(
model_prefix1, epoch1)
# network
if net1 is None:
net1 = load_net1
else:
net1 = net
if 'label' not in net1.list_arguments():
label1 = mx.sym.Variable(name='label')
net1 = mx.sym.Group([net1, label1])
# init module
mod1 = mx.mod.Module(net1,
label_names=('label', ),
logger=logger1,
context=ctx,
fixed_param_names=net1.list_arguments())
mod1.bind(data_shapes=eval_iter1.provide_data,
label_shapes=eval_iter1.provide_label)
mod1.set_params(args1, auxs1, allow_missing=False, force_init=True)
if voc07_metric:
metric1 = VOC07MApMetric(ovp_thresh, use_difficult, class_names)
else:
metric1 = MApMetric(ovp_thresh, use_difficult, class_names)
# filepath = '/home/binghao/workspace/MXNet-SSD/matlab/kitti/outputs/ssd/'
filepath1 = '/home/binghao/workspace/MXNet-SSD/matlab/kitti/outputs/ssd_small/'
# mod.score_m(filepath, eval_iter, metric, num_batch=None)
mod1.score_m(filepath1, eval_iter1, metric1, num_batch=None)
def evaluate_net(net,
path_imgrec,
num_classes,
mean_pixels,
data_shape,
model_prefix,
epoch,
ctx=mx.cpu(),
batch_size=1,
path_imglist="",
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
frequent=20):
"""
evalute network given validation record file
Parameters:
----------
net : str or None
Network name or use None to load from json without modifying
path_imgrec : str
path to the record validation file
path_imglist : str
path to the list file to replace labels in record file, optional
num_classes : int
number of classes, not including background
mean_pixels : tuple
(mean_r, mean_g, mean_b)
data_shape : tuple or int
(3, height, width) or height/width
model_prefix : str
model prefix of saved checkpoint
epoch : int
load model epoch
ctx : mx.ctx
mx.gpu() or mx.cpu()
batch_size : int
validation batch size
nms_thresh : float
non-maximum suppression threshold
force_nms : boolean
whether suppress different class objects
ovp_thresh : float
AP overlap threshold for true/false postives
use_difficult : boolean
whether to use difficult objects in evaluation if applicable
class_names : comma separated str
class names in string, must correspond to num_classes if set
voc07_metric : boolean
whether to use 11-point evluation as in VOC07 competition
frequent : int
frequency to print out validation status
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
#model_prefix += '_' + str(data_shape[1])
# iterator
eval_iter = DetRecordIter(path_imgrec,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=350,
path_imglist=path_imglist,
**cfg.valid)
# model params
load_net, args, auxs = mx.model.load_checkpoint(model_prefix, epoch)
# network
if net is None:
net = load_net
else:
net = get_symbol(net,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_nms)
if not 'label' in net.list_arguments():
label = mx.sym.Variable(name='label')
net = mx.sym.Group([net, label])
# init module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=net.list_arguments())
mod.bind(data_shapes=eval_iter.provide_data,
label_shapes=eval_iter.provide_label)
mod.set_params(args, auxs, allow_missing=False, force_init=True)
# run evaluation
if voc07_metric:
metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
roc_output_path=os.path.join(
os.path.dirname(model_prefix), 'roc'))
else:
metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
roc_output_path=os.path.join(
os.path.dirname(model_prefix), 'roc'))
posemetric = PoseMetric(
LINEMOD_path='/data/ZHANGXIN/DATASETS/SIXD_CHALLENGE/LINEMOD/',
classes=class_names)
# visualize bb8 results
# for nbatch, eval_batch in tqdm(enumerate(eval_iter)):
# mod.forward(eval_batch)
# preds = mod.get_outputs(merge_multi_context=True)
#
# labels = eval_batch.label[0].asnumpy()
# # get generated multi label from network
# cls_prob = preds[0]
# loc_pred = preds[4]
# bb8_pred = preds[5]
# anchors = preds[6]
#
# bb8dets = BB8MultiBoxDetection(cls_prob, loc_pred, bb8_pred, anchors, nms_threshold=0.5, force_suppress=False,
# variances=(0.1, 0.1, 0.2, 0.2), nms_topk=400)
# bb8dets = bb8dets.asnumpy()
#
# for nsample, sampleDet in enumerate(bb8dets):
# image = eval_batch.data[0][nsample].asnumpy()
# image += np.array(mean_pixels).reshape((3, 1, 1))
# image = np.transpose(image, axes=(1, 2, 0))
# draw_dets = []
# draw_cids = []
#
# for instanceDet in sampleDet:
# if instanceDet[0] == -1:
# continue
# else:
# cid = instanceDet[0].astype(np.int16)
# indices = np.where(sampleDet[:, 0] == cid)[0]
#
# if indices.size > 0:
# draw_dets.append(sampleDet[indices[0], 6:])
# draw_cids.append(cid)
# sampleDet = np.delete(sampleDet, indices, axis=0)
# show_BB8(image / 255., np.transpose(draw_dets[-1].reshape((-1, 8, 2)), axes=(0,2,1)), [cid],
# plot_path='./output/bb8results/{:04d}_{}'.format(nbatch * batch_size + nsample, class_names[cid]))
#
# # draw_dets = np.array(draw_dets)
# # draw_cids = np.array(draw_cids)
#
# # show_BB8(image / 255., np.transpose(draw_dets.reshape((-1, 8, 2)), axes=(0,2,1)), draw_cids,
# # plot_path='./output/bb8results/{:04d}'.format(nbatch * batch_size + nsample))
# quantitive results
results = mod.score(eval_iter, [metric, posemetric],
num_batch=None,
batch_end_callback=mx.callback.Speedometer(
batch_size, frequent=frequent, auto_reset=False))
results_save_path = os.path.join(os.path.dirname(model_prefix),
'evaluate_results')
with open(results_save_path, 'w') as f:
for k, v in results:
print("{}: {}".format(k, v))
f.write("{}: {}\n".format(k, v))
f.close()
reproj_save_path = os.path.join(os.path.dirname(model_prefix),
'reprojection_error')
with open(reproj_save_path, 'wb') as f:
# for k, v in metric.Reproj.items():
# f.write("{}: {}\n".format(k, v))
pickle.dump(posemetric.Reproj, f, protocol=2)
f.close()
count_save_path = os.path.join(os.path.dirname(model_prefix), 'gt_count')
with open(count_save_path, 'wb') as f:
# for k, v in metric.counts.items():
# f.write("{}: {}\n".format(k, v))
pickle.dump(posemetric.counts, f, protocol=2)
f.close()
# net.hybridize()
# out = net(data)
# net.export(path='./', epoch=0)
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if log_file:
log_file_path = os.path.join(os.path.dirname(prefix), log_file)
if not os.path.exists(os.path.dirname(log_file_path)):
os.makedirs(os.path.dirname(log_file_path))
fh = logging.FileHandler(log_file_path)
logger.addHandler(fh)
train_iter = DetRecordIter(train_path, batch_size, data_shape, mean_pixels=mean_pixels,
label_pad_width=label_pad_width, path_imglist=train_list, **cfg.train)
if val_path:
val_iter = DetRecordIter(val_path, batch_size, data_shape, mean_pixels=mean_pixels,
label_pad_width=label_pad_width, path_imglist=val_list, **cfg.valid)
else:
val_iter = None
# val_iter = None
cls_loss = MaskSoftmaxCELoss(class_axis=-1, batch_axis=0)
# cls_loss = FocalLoss(class_axis=-1, batch_axis=0, alpha=0.25, gamma=2)
box_loss = SmoothL1Loss(batch_axis=0)
cls_metric = mx.metric.Accuracy(axis=-1)
box_metric = MaskMAE()
def train_net(net,
train_path,
num_classes,
batch_size,
data_shape,
mean_img,
mean_img_dir,
resume,
finetune,
pretrained,
epoch,
prefix,
ctx,
begin_epoch,
end_epoch,
frequent,
learning_rate,
momentum,
weight_decay,
lr_refactor_step,
lr_refactor_ratio,
convert_numpy=1,
freeze_layer_pattern='',
num_example=10000,
label_pad_width=350,
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
nms_topk=400,
force_suppress=False,
train_list="",
val_path="",
val_list="",
iter_monitor=0,
monitor_pattern=".*",
log_file=None,
summarywriter=0,
flush_secs=180):
"""
Wrapper for training phase.
Parameters:
----------
net : str
symbol name for the network structure
train_path : str
record file path for training
num_classes : int
number of object classes, not including background
batch_size : int
training batch-size
data_shape : int or tuple
width/height as integer or (3, height, width) tuple
mean_pixels : tuple of floats
mean pixel values for red, green and blue
resume : int
resume from previous checkpoint if > 0
finetune : int
fine-tune from previous checkpoint if > 0
pretrained : str
prefix of pretrained model, including path
epoch : int
load epoch of either resume/finetune/pretrained model
prefix : str
prefix for saving checkpoints
ctx : [mx.cpu()] or [mx.gpu(x)]
list of mxnet contexts
begin_epoch : int
starting epoch for training, should be 0 if not otherwise specified
end_epoch : int
end epoch of training
frequent : int
frequency to print out training status
learning_rate : float
training learning rate
momentum : float
trainig momentum
weight_decay : float
training weight decay param
lr_refactor_ratio : float
multiplier for reducing learning rate
lr_refactor_step : comma separated integers
at which epoch to rescale learning rate, e.g. '30, 60, 90'
freeze_layer_pattern : str
regex pattern for layers need to be fixed
num_example : int
number of training images
label_pad_width : int
force padding training and validation labels to sync their label widths
nms_thresh : float
non-maximum suppression threshold for validation
force_nms : boolean
suppress overlaped objects from different classes
train_list : str
list file path for training, this will replace the embeded labels in record
val_path : str
record file path for validation
val_list : str
list file path for validation, this will replace the embeded labels in record
iter_monitor : int
monitor internal stats in networks if > 0, specified by monitor_pattern
monitor_pattern : str
regex pattern for monitoring network stats
log_file : str
log to file if enabled
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
if log_file:
fh = logging.FileHandler(log_file)
logger.addHandler(fh)
# check args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
prefix += '_' + net + '_' + str(data_shape[1])
# if isinstance(mean_pixels, (int, float)):
# mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
# assert len(mean_pixels) == 3, "must provide all RGB mean values"
train_iter = DetRecordIter(train_path,
batch_size,
data_shape,
mean_img=mean_img,
label_pad_width=label_pad_width,
path_imglist=train_list,
**cfg.train)
if val_path:
val_iter = DetRecordIter(val_path,
batch_size,
data_shape,
mean_img=mean_img,
label_pad_width=label_pad_width,
path_imglist=val_list,
**cfg.valid)
else:
val_iter = None
# convert mean.bin to mean.npy
_convert_mean_numpy(convert_numpy, mean_img_dir, mean_img)
# load symbol
net = get_symbol_train(net,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_suppress,
nms_topk=nms_topk)
if summarywriter:
if os.path.exists('/opt/incubator-mxnet/example/ssd/logs'):
shutil.rmtree('/opt/incubator-mxnet/example/ssd/logs'
) # clear the previous logs
os.mkdir('/opt/incubator-mxnet/example/ssd/logs')
sw = SummaryWriter(logdir='/opt/incubator-mxnet/example/ssd/logs',
flush_secs=flush_secs)
sw.add_graph(net)
else:
sw = None
# mx.viz.plot_network(net, shape={"data":(64, 3, 320, 320)}, node_attrs={"shape":'rect',"fixedsize":'false'}).view()
# define layers with fixed weight/bias
if freeze_layer_pattern.strip():
re_prog = re.compile(freeze_layer_pattern)
fixed_param_names = [
name for name in net.list_arguments() if re_prog.match(name)
]
else:
fixed_param_names = None
# load pretrained or resume from previous state
ctx_str = '(' + ','.join([str(c) for c in ctx]) + ')'
if resume > 0:
logger.info("Resume training with {} from epoch {}".format(
ctx_str, resume))
_, args, auxs = mx.model.load_checkpoint(prefix, resume)
begin_epoch = resume
elif finetune > 0:
logger.info("Start finetuning with {} from epoch {}".format(
ctx_str, finetune))
_, args, auxs = mx.model.load_checkpoint(prefix, finetune)
begin_epoch = finetune
# the prediction convolution layers name starts with relu, so it's fine
fixed_param_names = [name for name in net.list_arguments() \
if name.startswith('conv')]
elif pretrained:
logger.info("Start training with {} from pretrained model {}".format(
ctx_str, pretrained))
_, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
args = convert_pretrained(pretrained, args)
else:
logger.info("Experimental: start training from scratch with {}".format(
ctx_str))
args = None
auxs = None
fixed_param_names = None
# helper information
if fixed_param_names:
logger.info("Freezed parameters: [" + ','.join(fixed_param_names) +
']')
# init training module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=fixed_param_names)
# fit parameters
if summarywriter:
# 增加可视化的回调函数，有多个回调函数时，除最后一个回调函数外不能进行准确率的清零操作(即auto_reset参数必须设置为False)
batch_end_callbacks = [
mx.callback.Speedometer(train_iter.batch_size,
frequent=frequent,
auto_reset=True),
summary_writter_callback.summary_writter_eval_metric(sw)
]
else:
batch_end_callbacks = [
mx.callback.Speedometer(train_iter.batch_size,
frequent=frequent,
auto_reset=False)
]
# batch_end_callback = mx.callback.Speedometer(train_iter.batch_size, frequent=frequent)
epoch_end_callback = mx.callback.do_checkpoint(prefix)
learning_rate, lr_scheduler = get_lr_scheduler(learning_rate,
lr_refactor_step,
lr_refactor_ratio,
num_example, batch_size,
begin_epoch)
optimizer_params = {
'learning_rate': learning_rate,
'momentum': momentum,
'wd': weight_decay,
'lr_scheduler': lr_scheduler,
'clip_gradient': None,
'rescale_grad': 1.0 / len(ctx) if len(ctx) > 0 else 1.0
}
monitor = mx.mon.Monitor(
iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
# run fit net, every n epochs we run evaluation network to get mAP
if voc07_metric:
valid_metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=3)
else:
valid_metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=3)
mod.fit(train_iter,
val_iter,
eval_metric=MultiBoxMetric(),
validation_metric=valid_metric,
batch_end_callback=batch_end_callbacks,
epoch_end_callback=epoch_end_callback,
optimizer='sgd',
optimizer_params=optimizer_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=True,
monitor=monitor)
if summarywriter:
sw.close()
def train_net(net,
train_path,
num_classes,
batch_size,
data_shape,
mean_pixels,
resume,
finetune,
pretrained,
epoch,
prefix,
ctx,
begin_epoch,
end_epoch,
frequent,
learning_rate,
momentum,
weight_decay,
lr_refactor_step,
lr_refactor_ratio,
freeze_layer_pattern='',
num_example=10000,
label_pad_width=350,
nms_thresh=0.45,
force_nms=False,
ovp_thresh=0.5,
use_difficult=False,
class_names=None,
voc07_metric=False,
nms_topk=2000,
force_suppress=False,
train_list="",
val_path="",
val_list="",
iter_monitor=0,
monitor_pattern=".*",
log_file=None):
"""
Wrapper for training phase.
Parameters:
----------
net : str
symbol name for the network structure
train_path : str
record file path for training
num_classes : int
number of object classes, not including background
batch_size : int
training batch-size
data_shape : int or tuple
width/height as integer or (3, height, width) tuple
mean_pixels : tuple of floats
mean pixel values for red, green and blue
resume : int
resume from previous checkpoint if > 0
finetune : int
fine-tune from previous checkpoint if > 0
pretrained : str
prefix of pretrained model, including path
epoch : int
load epoch of either resume/finetune/pretrained model
prefix : str
prefix for saving checkpoints
ctx : [mx.cpu()] or [mx.gpu(x)]
list of mxnet contexts
begin_epoch : int
starting epoch for training, should be 0 if not otherwise specified
end_epoch : int
end epoch of training
frequent : int
frequency to print out training status
learning_rate : float
training learning rate
momentum : float
trainig momentum
weight_decay : float
training weight decay param
lr_refactor_ratio : float
multiplier for reducing learning rate
lr_refactor_step : comma separated integers
at which epoch to rescale learning rate, e.g. '30, 60, 90'
freeze_layer_pattern : str
regex pattern for layers need to be fixed
num_example : int
number of training images
label_pad_width : int
force padding training and validation labels to sync their label widths
nms_thresh : float
non-maximum suppression threshold for validation
force_nms : boolean
suppress overlaped objects from different classes
train_list : str
list file path for training, this will replace the embeded labels in record
val_path : str
record file path for validation
val_list : str
list file path for validation, this will replace the embeded labels in record
iter_monitor : int
monitor internal stats in networks if > 0, specified by monitor_pattern
monitor_pattern : str
regex pattern for monitoring network stats
log_file : str
log to file if enabled
"""
# set up logger
logging.basicConfig()
logger = logging.getLogger()
logger.setLevel(logging.INFO)
# set a log
if log_file:
# crate a fileHandler
fh = logging.FileHandler(log_file)
logger.addHandler(fh)
# check args
if isinstance(data_shape, int):
data_shape = (3, data_shape, data_shape)
assert len(data_shape) == 3 and data_shape[0] == 3
prefix += '_' + net + '_' + str(data_shape[1])
# check the mean_pixels is list
if isinstance(mean_pixels, (int, float)):
mean_pixels = [mean_pixels, mean_pixels, mean_pixels]
assert len(mean_pixels) == 3, "must provide all RGB mean values"
train_iter = DetRecordIter(train_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=train_list,
**cfg.train)
# for c in range(12840):
# batch = train_iter.next()
# data=batch.data[0]
# label=batch.label[0]
# from matplotlib import pyplot as plt
# import numpy as np
# import cv2
# for i in range(2):
# plt.subplot(1,2,i+1)
# img = np.array(data[i].asnumpy().transpose(1,2,0).copy(), np.uint8)
# box = label[i].asnumpy()
# bbox = []
# print 'The', i, 'th image'
# for j in range(box.shape[0]):
# if box[j][0] == -1:
# break
# else:
# bbox.append(box[j][1:5])
# for k in range(len(bbox)):
# xmin = (bbox[k][0] * img.shape[0]).astype(np.int16)
# ymin = (bbox[k][1] * img.shape[0]).astype(np.int16)
# xmax = (bbox[k][2] * img.shape[0]).astype(np.int16)
# ymax = (bbox[k][3] * img.shape[0]).astype(np.int16)
# cv2.rectangle(img, (xmin,ymin), (xmax,ymax), (255,0,0),4)
#
# print 'xmin', xmin, 'ymin', ymin, 'xmax', xmax, 'ymax', ymax
# plt.imshow(img)
# plt.show()
# #path = 'crop_image/'+ str(c) + '.jpg'
# #plt.savefig(path)
# print batch
if val_path:
val_iter = DetRecordIter(val_path,
batch_size,
data_shape,
mean_pixels=mean_pixels,
label_pad_width=label_pad_width,
path_imglist=val_list,
**cfg.valid)
else:
val_iter = None
# load symbol
net = get_symbol_train(net,
data_shape[1],
num_classes=num_classes,
nms_thresh=nms_thresh,
force_suppress=force_suppress,
nms_topk=nms_topk)
# viz = mx.viz.plot_network(net)
# viz.view()
# define layers with fixed weight/bias
if freeze_layer_pattern.strip():
re_prog = re.compile(freeze_layer_pattern)
fixed_param_names = [
name for name in net.list_arguments() if re_prog.match(name)
]
else:
fixed_param_names = None
# load pretrained or resume from previous state
ctx_str = '(' + ','.join([str(c) for c in ctx]) + ')'
if resume > 0:
logger.info("Resume training with {} from epoch {}".format(
ctx_str, resume))
_, args, auxs = mx.model.load_checkpoint(prefix, resume)
begin_epoch = resume
elif finetune > 0:
logger.info("Start finetuning with {} from epoch {}".format(
ctx_str, finetune))
_, args, auxs = mx.model.load_checkpoint(prefix, finetune)
begin_epoch = finetune
# the prediction convolution layers name starts with relu, so it's fine
fixed_param_names = [name for name in net.list_arguments() \
if name.startswith('conv')]
elif pretrained:
logger.info("Start training with {} from pretrained model {}".format(
ctx_str, pretrained))
fixed_param_names = None
_, args, auxs = mx.model.load_checkpoint(pretrained, epoch)
args = convert_pretrained(pretrained, args)
else:
logger.info("Experimental: start training from scratch with {}".format(
ctx_str))
args = None
auxs = None
fixed_param_names = None
# helper information
if fixed_param_names:
logger.info("Freezed parameters: [" + ','.join(fixed_param_names) +
']')
# init training module
mod = mx.mod.Module(net,
label_names=('label', ),
logger=logger,
context=ctx,
fixed_param_names=fixed_param_names)
# fit parameters
batch_end_callback = mx.callback.Speedometer(train_iter.batch_size,
frequent=frequent)
epoch_end_callback = mx.callback.do_checkpoint(prefix)
learning_rate, lr_scheduler = get_lr_scheduler(learning_rate,
lr_refactor_step,
lr_refactor_ratio,
num_example, batch_size,
begin_epoch)
optimizer_params = {
'learning_rate': learning_rate,
'momentum': momentum,
'wd': weight_decay,
'lr_scheduler': lr_scheduler,
'clip_gradient': None,
'rescale_grad': 1.0 / len(ctx) if len(ctx) > 0 else 1.0
}
monitor = mx.mon.Monitor(
iter_monitor, pattern=monitor_pattern) if iter_monitor > 0 else None
# run fit net, every n epochs we run evaluation network to get mAP
if voc07_metric:
valid_metric = VOC07MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=3)
else:
valid_metric = MApMetric(ovp_thresh,
use_difficult,
class_names,
pred_idx=3)
mod.fit(
train_data=train_iter, #train_iter,
eval_data=val_iter,
eval_metric=MultiBoxMetric(),
validation_metric=valid_metric,
batch_end_callback=batch_end_callback,
epoch_end_callback=epoch_end_callback,
optimizer='sgd',
optimizer_params=optimizer_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
initializer=mx.init.Xavier(),
arg_params=args,
aux_params=auxs,
allow_missing=True,
monitor=monitor)
