class SolverWrapper(object):
"""
A wrapper class for the training process
"""
def __init__(self,
sess,
network,
imdb,
roidb,
valroidb,
output_dir,
tbdir,
pretrained_model=None):
self.net = network
self.imdb = imdb
self.roidb = roidb
self.valroidb = valroidb
self.output_dir = output_dir
self.tbdir = tbdir
# Simply put '_val' at the end to save the summaries from the validation set
self.tbvaldir = tbdir + '_val'
if not os.path.exists(self.tbvaldir):
os.makedirs(self.tbvaldir)
self.pretrained_model = pretrained_model
def snapshot(self, sess, iter):
net = self.net
if not os.path.exists(self.output_dir):
os.makedirs(self.output_dir)
# Store the model snapshot
filename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(
iter) + '.ckpt'
filename = os.path.join(self.output_dir, filename)
self.saver.save(sess, filename)
print('Wrote snapshot to: {:s}'.format(filename))
# Also store some meta information, random state, etc.
nfilename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(
iter) + '.pkl'
nfilename = os.path.join(self.output_dir, nfilename)
# current state of numpy random
st0 = np.random.get_state()
# current position in the database
cur = self.data_layer._cur
# current shuffled indeces of the database
perm = self.data_layer._perm
# current position in the validation database
cur_val = self.data_layer_val._cur
# current shuffled indeces of the validation database
perm_val = self.data_layer_val._perm
# Dump the meta info
with open(nfilename, 'wb') as fid:
pickle.dump(st0, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(iter, fid, pickle.HIGHEST_PROTOCOL)
return filename, nfilename
def get_variables_in_checkpoint_file(self, file_name):
try:
reader = pywrap_tensorflow.NewCheckpointReader(file_name)
var_to_shape_map = reader.get_variable_to_shape_map()
return var_to_shape_map
except Exception as e: # pylint: disable=broad-except
print(str(e))
if "corrupted compressed block contents" in str(e):
print(
"It's likely that your checkpoint file has been compressed "
"with SNAPPY.")
def train_model(self, sess, max_iters):
# Build data layers for both training and validation set
self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
self.data_layer_val = RoIDataLayer(self.valroidb,
self.imdb.num_classes,
random=True)
# Determine different scales for anchors, see paper
with sess.graph.as_default():
# Set the random seed for tensorflow
tf.set_random_seed(cfg.RNG_SEED)
# Build the main computation graph
layers = self.net.create_architecture(
sess,
'TRAIN',
self.imdb.num_classes,
tag='default',
anchor_scales=cfg.ANCHOR_SCALES,
anchor_ratios=cfg.ANCHOR_RATIOS)
# Define the loss
loss = layers['total_loss']
# Set learning rate and momentum
lr = tf.Variable(cfg.TRAIN.LEARNING_RATE, trainable=False)
momentum = cfg.TRAIN.MOMENTUM
self.optimizer = tf.train.MomentumOptimizer(lr, momentum)
# Compute the gradients wrt the loss
gvs = self.optimizer.compute_gradients(loss)
# Double the gradient of the bias if set
if cfg.TRAIN.DOUBLE_BIAS:
final_gvs = []
with tf.variable_scope('Gradient_Mult') as scope:
for grad, var in gvs:
scale = 1.
if cfg.TRAIN.DOUBLE_BIAS and '/biases:' in var.name:
scale *= 2.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs.append((grad, var))
train_op = self.optimizer.apply_gradients(final_gvs)
else:
train_op = self.optimizer.apply_gradients(gvs)
# We will handle the snapshots ourselves
self.saver = tf.train.Saver(max_to_keep=100000)
# Write the train and validation information to tensorboard
self.writer = tf.summary.FileWriter(self.tbdir, sess.graph)
self.valwriter = tf.summary.FileWriter(self.tbvaldir)
# Find previous snapshots if there is any to restore from
sfiles = os.path.join(self.output_dir,
cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.ckpt.meta')
sfiles = glob.glob(sfiles)
sfiles.sort(key=os.path.getmtime)
# Get the snapshot name in TensorFlow
redstr = '_iter_{:d}.'.format(cfg.TRAIN.STEPSIZE + 1)
sfiles = [ss.replace('.meta', '') for ss in sfiles]
sfiles = [ss for ss in sfiles if redstr not in ss]
nfiles = os.path.join(self.output_dir,
cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.pkl')
nfiles = glob.glob(nfiles)
nfiles.sort(key=os.path.getmtime)
nfiles = [nn for nn in nfiles if redstr not in nn]
lsf = len(sfiles)
assert len(nfiles) == lsf
np_paths = nfiles
ss_paths = sfiles
if lsf == 0:
# Fresh train directly from ImageNet weights
print('Loading initial model weights from {:s}'.format(
self.pretrained_model))
variables = tf.global_variables()
# Initialize all variables first
sess.run(tf.variables_initializer(variables, name='init'))
var_keep_dic = self.get_variables_in_checkpoint_file(
self.pretrained_model)
# Get the variables to restore, ignorizing the variables to fix
variables_to_restore = self.net.get_variables_to_restore(
variables, var_keep_dic)
restorer = tf.train.Saver(variables_to_restore)
restorer.restore(sess, self.pretrained_model)
print('Loaded.')
# Need to fix the variables before loading, so that the RGB weights are changed to BGR
# For VGG16 it also changes the convolutional weights fc6 and fc7 to
# fully connected weights
self.net.fix_variables(sess, self.pretrained_model)
print('Fixed.')
sess.run(tf.assign(lr, cfg.TRAIN.LEARNING_RATE))
last_snapshot_iter = 0
else:
# Get the most recent snapshot and restore
ss_paths = [ss_paths[-1]]
np_paths = [np_paths[-1]]
print('Restorining model snapshots from {:s}'.format(sfiles[-1]))
self.saver.restore(sess, str(sfiles[-1]))
print('Restored.')
# Needs to restore the other hyperparameters/states for training, (TODO xinlei) I have
# tried my best to find the random states so that it can be recovered exactly
# However the Tensorflow state is currently not available
with open(str(nfiles[-1]), 'rb') as fid:
st0 = pickle.load(fid)
cur = pickle.load(fid)
perm = pickle.load(fid)
cur_val = pickle.load(fid)
perm_val = pickle.load(fid)
last_snapshot_iter = pickle.load(fid)
np.random.set_state(st0)
self.data_layer._cur = cur
self.data_layer._perm = perm
self.data_layer_val._cur = cur_val
self.data_layer_val._perm = perm_val
# Set the learning rate, only reduce once
if last_snapshot_iter > cfg.TRAIN.STEPSIZE:
sess.run(
tf.assign(lr,
cfg.TRAIN.LEARNING_RATE * cfg.TRAIN.GAMMA))
else:
sess.run(tf.assign(lr, cfg.TRAIN.LEARNING_RATE))
timer = Timer()
iter = last_snapshot_iter + 1
last_summary_time = time.time()
while iter < max_iters + 1:
# Learning rate
if iter == cfg.TRAIN.STEPSIZE + 1:
# Add snapshot here before reducing the learning rate
self.snapshot(sess, iter)
sess.run(
tf.assign(lr, cfg.TRAIN.LEARNING_RATE * cfg.TRAIN.GAMMA))
timer.tic()
# Get training data, one batch at a time
blobs = self.data_layer.forward()
now = time.time()
if now - last_summary_time > cfg.TRAIN.SUMMARY_INTERVAL:
# Compute the graph with summary
rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, total_loss, summary = \
self.net.train_step_with_summary(sess, blobs, train_op)
self.writer.add_summary(summary, float(iter))
# Also check the summary on the validation set
blobs_val = self.data_layer_val.forward()
summary_val = self.net.get_summary(sess, blobs_val)
self.valwriter.add_summary(summary_val, float(iter))
last_summary_time = now
else:
# Compute the graph without summary
rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, total_loss = \
self.net.train_step(sess, blobs, train_op)
timer.toc()
# Display training information
if iter % (cfg.TRAIN.DISPLAY) == 0:
print('iter: %d / %d, total loss: %.6f\n >>> rpn_loss_cls: %.6f\n '
'>>> rpn_loss_box: %.6f\n >>> loss_cls: %.6f\n >>> loss_box: %.6f\n >>> lr: %f' % \
(iter, max_iters, total_loss, rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, lr.eval()))
print('speed: {:.3f}s / iter'.format(timer.average_time))
if iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:
last_snapshot_iter = iter
snapshot_path, np_path = self.snapshot(sess, iter)
np_paths.append(np_path)
ss_paths.append(snapshot_path)
# Remove the old snapshots if there are too many
if len(np_paths) > cfg.TRAIN.SNAPSHOT_KEPT:
to_remove = len(np_paths) - cfg.TRAIN.SNAPSHOT_KEPT
for c in range(to_remove):
nfile = np_paths[0]
os.remove(str(nfile))
np_paths.remove(nfile)
if len(ss_paths) > cfg.TRAIN.SNAPSHOT_KEPT:
to_remove = len(ss_paths) - cfg.TRAIN.SNAPSHOT_KEPT
for c in range(to_remove):
sfile = ss_paths[0]
# To make the code compatible to earlier versions of Tensorflow,
# where the naming tradition for checkpoints are different
if os.path.exists(str(sfile)):
os.remove(str(sfile))
else:
os.remove(str(sfile + '.data-00000-of-00001'))
os.remove(str(sfile + '.index'))
sfile_meta = sfile + '.meta'
os.remove(str(sfile_meta))
ss_paths.remove(sfile)
iter += 1
if last_snapshot_iter != iter - 1:
self.snapshot(sess, iter - 1)
self.writer.close()
self.valwriter.close()
class SolverWrapper(object):
def __init__(self,
network,
imdb,
valimdb,
roidb,
valroidb,
model_dir,
pretrained_model=None):
self.net = network
self.imdb = imdb
self.valimdb = valimdb
self.roidb = roidb
self.valroidb = valroidb
self.model_dir = model_dir
self.tbdir = os.path.join(model_dir, 'train_log')
if not os.path.exists(self.tbdir):
os.makedirs(self.tbdir)
self.pretrained_model = pretrained_model
def set_learn_strategy(self, learn_dict):
self._disp_interval = learn_dict['disp_interval']
self._valid_interval = learn_dict['disp_interval'] * 5
self._use_tensorboard = learn_dict['use_tensorboard']
self._use_valid = learn_dict['use_valid']
self._evaluate = learn_dict['evaluate']
self._save_point_interval = learn_dict['save_point_interval']
self._lr_decay_steps = learn_dict['lr_decay_steps']
if self._evaluate:
self._begin_eval_point = learn_dict['begin_eval_point']
self.evaluate_dir = os.path.join(self.model_dir, 'evaluate')
self._evaluate_thresh = 0.05
self._evaluate_max_per_image = 1
def train_model(self, resume=None, max_iters=100000):
# Build data layers for both training and validation set
self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
self.data_layer_val = RoIDataLayer(self.valroidb,
self.imdb.num_classes,
random=True)
# 载入存档点，初始化权重，设置优化函数，设置网络权重学习率
self.prepare_construct(resume)
net = self.net
# training
train_loss = 0
rpn_cls_loss = 0
rpn_bbox_loss = 0
fast_rcnn_cls_loss = 0
fast_rcnn_bbox_loss = 0
tp, tf, fg, bg = 0., 0., 0, 0
step_cnt = 0
re_cnt = False
t = Timer()
t.tic()
for step in range(self.start_step, max_iters + 1):
blobs = self.data_layer.forward()
if step % self._valid_interval == 0 and self._use_tensorboard:
loss_r, image_r = net.train_operation(
blobs,
self._optimizer,
image_if=True,
clip_parameters=self._parameters)
self._tensor_writer.add_image('Image', image_r, step)
else:
try:
loss_r, image_r = net.train_operation(
blobs,
self._optimizer,
image_if=False,
clip_parameters=self._parameters)
except:
print('=' * 40)
print('=' * 40)
print('=' * 40)
print(blobs['im_name'])
train_loss += loss_r[0]
rpn_cls_loss += loss_r[1]
rpn_bbox_loss += loss_r[2]
fast_rcnn_cls_loss += loss_r[3]
fast_rcnn_bbox_loss += loss_r[4]
# fg：物体 bg：背景 tp：真阳 tf：真阴
fg += net.metrics_dict['fg']
bg += net.metrics_dict['bg']
tp += net.metrics_dict['tp']
tf += net.metrics_dict['tf']
step_cnt += 1
if step % self._disp_interval == 0:
duration = t.toc(average=False)
fps = step_cnt / duration
log_text = 'step %d, image: %s, loss: %.4f, fps: %.2f (%.2fs per batch)' % (
step, blobs['im_name'], train_loss / step_cnt, fps,
1. / fps)
tp_text = 'step {}, tp: {}/{}, tf: {}/{}'.format(
step, int(tp / step_cnt), int(fg / step_cnt),
int(tf / step_cnt), int(bg / step_cnt))
pprint.pprint(log_text)
pprint.pprint(tp_text)
if self._use_tensorboard:
self._tensor_writer.add_text('Train',
log_text,
global_step=step)
# Train
avg_rpn_cls_loss = rpn_cls_loss / step_cnt
avg_rpn_bbox_loss = rpn_bbox_loss / step_cnt
avg_fast_rcnn_cls_loss = fast_rcnn_cls_loss / step_cnt
avg_fast_rcnn_bbox_loss = fast_rcnn_bbox_loss / step_cnt
self._tensor_writer.add_scalars(
'TrainSetLoss', {
'RPN_cls_loss': avg_rpn_cls_loss,
'RPN_bbox_loss': avg_rpn_bbox_loss,
'FastRcnn_cls_loss': avg_fast_rcnn_cls_loss,
'FastRcnn_bbox_loss': avg_fast_rcnn_bbox_loss
},
global_step=step)
self._tensor_writer.add_scalar('Learning_rate',
self._lr,
global_step=step)
re_cnt = True
if self._use_valid and step % self._valid_interval == 0 and step != 0:
total_valid_loss = 0.0
valid_rpn_cls_loss = 0.0
valid_rpn_bbox_loss = 0.0
valid_fast_rcnn_cls_loss = 0.0
valid_fast_rcnn_bbox_loss = 0.0
valid_step_cnt = 0
valid_tp, valid_tf, valid_fg, valid_bg = 0., 0., 0, 0
start_time = time.time()
valid_length = self._disp_interval
for valid_batch in range(valid_length):
# get one batch
blobs = self.data_layer_val.forward()
if self._use_tensorboard and valid_batch % valid_length == 0:
# 此处没传optimizer，不会更新网络，只计算loss
loss_r, image_r = net.train_operation(blobs,
None,
image_if=True)
self._tensor_writer.add_image('Image_Valid', image_r,
step)
else:
loss_r, image_r = net.train_operation(blobs,
None,
image_if=False)
total_valid_loss += loss_r[0]
valid_rpn_cls_loss += loss_r[1]
valid_rpn_bbox_loss += loss_r[2]
valid_fast_rcnn_cls_loss += loss_r[3]
valid_fast_rcnn_bbox_loss += loss_r[4]
valid_fg += net.metrics_dict['fg']
valid_bg += net.metrics_dict['bg']
valid_tp += net.metrics_dict['tp']
valid_tf += net.metrics_dict['tf']
valid_step_cnt += 1
duration = time.time() - start_time
fps = valid_step_cnt / duration
log_text = 'step %d, valid average loss: %.4f, fps: %.2f (%.2fs per batch)' % (
step, total_valid_loss / valid_step_cnt, fps, 1. / fps)
pprint.pprint(log_text)
if self._use_tensorboard:
# Valid
avg_rpn_cls_loss_valid = valid_rpn_cls_loss / valid_step_cnt
avg_rpn_bbox_loss_valid = valid_rpn_bbox_loss / valid_step_cnt
avg_fast_rcnn_cls_loss_valid = valid_fast_rcnn_cls_loss / valid_step_cnt
avg_fast_rcnn_bbox_loss_valid = valid_fast_rcnn_bbox_loss / valid_step_cnt
valid_tpr = valid_tp * 1.0 / valid_fg
valid_tfr = valid_tf * 1.0 / valid_bg
real_total_loss_valid = valid_rpn_cls_loss + valid_rpn_bbox_loss\
+ valid_fast_rcnn_cls_loss + valid_fast_rcnn_bbox_loss
# Train
avg_rpn_cls_loss = rpn_cls_loss / step_cnt
avg_rpn_bbox_loss = rpn_bbox_loss / step_cnt
avg_fast_rcnn_cls_loss = fast_rcnn_cls_loss / step_cnt
avg_fast_rcnn_bbox_loss = fast_rcnn_bbox_loss / step_cnt
tpr = tp * 1.0 / fg
tfr = tf * 1.0 / bg
real_total_loss = rpn_cls_loss + rpn_bbox_loss + fast_rcnn_cls_loss + fast_rcnn_bbox_loss
self._tensor_writer.add_text('Valid',
log_text,
global_step=step)
self._tensor_writer.add_scalars(
'Total_Loss', {
'train': train_loss / step_cnt,
'valid': total_valid_loss / valid_step_cnt
},
global_step=step)
self._tensor_writer.add_scalars(
'Real_loss', {
'train': real_total_loss / step_cnt,
'valid': real_total_loss_valid / valid_step_cnt
},
global_step=step)
self._tensor_writer.add_scalars(
'RPN_cls_loss', {
'train': avg_rpn_cls_loss,
'valid': avg_rpn_cls_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'RPN_bbox_loss', {
'train': avg_rpn_bbox_loss,
'valid': avg_rpn_bbox_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'FastRcnn_cls_loss', {
'train': avg_fast_rcnn_cls_loss,
'valid': avg_fast_rcnn_cls_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'FastRcnn_bbox_loss', {
'train': avg_fast_rcnn_bbox_loss,
'valid': avg_fast_rcnn_bbox_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars('tpr', {
'train': tpr,
'valid': valid_tpr
},
global_step=step)
self._tensor_writer.add_scalars('tfr', {
'train': tfr,
'valid': valid_tfr
},
global_step=step)
self._tensor_writer.add_scalars(
'ValidSetLoss', {
'RPN_cls_loss': avg_rpn_cls_loss_valid,
'RPN_bbox_loss': avg_rpn_bbox_loss_valid,
'FastRcnn_cls_loss': avg_fast_rcnn_cls_loss_valid,
'FastRcnn_bbox_loss': avg_fast_rcnn_bbox_loss_valid
},
global_step=step)
if (step % self._save_point_interval == 0) and step != 0:
save_name, _ = self.save_check_point(step)
print('save model: {}'.format(save_name))
if self._evaluate:
if step > self._begin_eval_point and step % cfg.TRAIN.EVALUATE_POINT == 0:
self.net.eval()
evaluate_solverwrapper = EvaluateSolverWrapper(
network=self.net,
imdb=self.valimdb,
model_dir=None,
output_dir=self.evaluate_dir)
metrics_cls, metrics_reg = evaluate_solverwrapper.\
eval_model(step, self._evaluate_max_per_image, self._evaluate_thresh)
self.after_model_mode()
del evaluate_solverwrapper
for key in metrics_cls.keys():
metrics_cls[key] = metrics_cls[key][0]
for key in metrics_reg.keys():
metrics_reg[key] = metrics_reg[key][0]
if self._use_tensorboard:
self._tensor_writer.add_scalars('metrics_cls',
metrics_cls,
global_step=step)
self._tensor_writer.add_scalars('metrics_reg',
metrics_reg,
global_step=step)
if step in self._lr_decay_steps:
self._lr *= self._lr_decay
self._optimizer = self._train_optimizer()
if re_cnt:
tp, tf, fg, bg = 0., 0., 0., 0.
train_loss = 0
rpn_cls_loss = 0
rpn_bbox_loss = 0
fast_rcnn_cls_loss = 0
fast_rcnn_bbox_loss = 0
step_cnt = 0
t.tic()
re_cnt = False
if self._use_tensorboard:
self._tensor_writer.export_scalars_to_json(
os.path.join(self.tbdir, 'all_scalars.json'))
def save_check_point(self, step):
net = self.net
if not os.path.exists(self.model_dir):
os.makedirs(self.model_dir)
# store the model snapshot
filename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.h5'.format(step))
h5f = h5py.File(filename, mode='w')
for k, v in net.state_dict().items():
h5f.create_dataset(k, data=v.cpu().numpy())
# store data information
nfilename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.pkl'.format(step))
# current state of numpy random
st0 = np.random.get_state()
# current position in the database
cur = self.data_layer._cur
# current shuffled indexes of the database
perm = self.data_layer._perm
# current position in the validation database
cur_val = self.data_layer_val._cur
# current shuffled indexes of the validation database
perm_val = self.data_layer_val._perm
# current learning rate
lr = self._lr
# Dump the meta info
with open(nfilename, 'wb') as fid:
pickle.dump(st0, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(lr, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(step, fid, pickle.HIGHEST_PROTOCOL)
return filename, nfilename
def load_check_point(self, step):
net = self.net
filename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.h5'.format(step))
nfilename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.pkl'.format(step))
print('Restoring model snapshots from {:s}'.format(filename))
if not os.path.exists(filename):
print('The checkPoint is not Right')
sys.exit(1)
# load model
h5f = h5py.File(filename, mode='r')
for k, v in net.state_dict().items():
param = torch.from_numpy(np.asarray(h5f[k]))
v.copy_(param)
# load data information
with open(nfilename, 'rb') as fid:
st0 = pickle.load(fid)
cur = pickle.load(fid)
perm = pickle.load(fid)
cur_val = pickle.load(fid)
perm_val = pickle.load(fid)
lr = pickle.load(fid)
last_snapshot_iter = pickle.load(fid)
np.random.set_state(st0)
self.data_layer._cur = cur
self.data_layer._perm = perm
self.data_layer_val._cur = cur_val
self.data_layer_val._perm = perm_val
self._lr = lr
if last_snapshot_iter == step:
print('Restore over ')
else:
print('The checkPoint is not Right')
raise ValueError
return last_snapshot_iter
#初始化网络权重
def weights_normal_init(self, model, dev=0.01):
import math
def _gaussian_init(m, dev):
m.weight.data.normal_(0.0, dev)
if hasattr(m.bias, 'data'):
m.bias.data.zero_()
def _xaiver_init(m):
nn.init.xavier_normal(m.weight.data)
if hasattr(m.bias, 'data'):
m.bias.data.zero_()
def _hekaiming_init(m):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
if hasattr(m.bias, 'data'):
m.bias.data.zero_()
def _resnet_init(model, dev):
if isinstance(model, list):
for m in model:
self.weights_normal_init(m, dev)
else:
for m in model.modules():
if isinstance(m, nn.Conv2d):
_hekaiming_init(m)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
_gaussian_init(m, dev)
def _vgg_init(model, dev):
if isinstance(model, list):
for m in model:
self.weights_normal_init(m, dev)
else:
for m in model.modules():
if isinstance(m, nn.Conv2d):
_gaussian_init(m, dev)
elif isinstance(m, nn.Linear):
_gaussian_init(m, dev)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
if cfg.TRAIN.INIT_WAY == 'resnet':
_vgg_init(model, dev)
elif cfg.TRAIN.INIT_WAY == 'vgg':
_vgg_init(model, dev)
else:
raise NotImplementedError
#载入存档点，初始化权重，设置优化函数，设置网络权重学习率
def prepare_construct(self, resume_iter):
# init network
self.net.init_fasterRCNN()
# Set the random seed
torch.manual_seed(cfg.RNG_SEED)
np.random.seed(cfg.RNG_SEED)
# Set learning rate and momentum
self._lr = cfg.TRAIN.LEARNING_RATE
self._lr_decay = 0.1
self._momentum = cfg.TRAIN.MOMENTUM
self._weight_decay = cfg.TRAIN.WEIGHT_DECAY
# load model
if resume_iter:
self.start_step = resume_iter + 1
self.load_check_point(resume_iter)
else:
self.start_step = 0
self.weights_normal_init(self.net, dev=0.01)
# refer to caffe faster RCNN
self.net.init_special_bbox_fc(dev=0.001)
if self.pretrained_model != None:
self.net._rpn._network._load_pre_trained_model(
self.pretrained_model)
print('Load parameters from Path: {}'.format(
self.pretrained_model))
else:
pass
if cfg.CUDA_IF:
self.net.cuda()
# BN should be fixed
self.after_model_mode()
# set optimizer
self._parameters = [
params for params in self.net.parameters()
if params.requires_grad == True
]
self._optimizer = self._train_optimizer()
# tensorboard
if self._use_tensorboard:
import tensorboardX as tbx
self._tensor_writer = tbx.SummaryWriter(log_dir=self.tbdir)
def after_model_mode(self):
# model
self.net.train()
# resnet fixed BN should be eval
if cfg.TRAIN.INIT_WAY == 'resnet':
self.net._rpn._network._bn_eval()
def _train_optimizer(self):
parameters = self._train_parameter()
optimizer = torch.optim.SGD(parameters, momentum=self._momentum)
return optimizer
def _train_parameter(self):
params = []
for key, value in self.net.named_parameters():
if value.requires_grad == True:
if 'bias' in key:
params += [{
'params': [value],
'lr': self._lr * (cfg.TRAIN.DOUBLE_BIAS + 1),
'weight_decay': 0
}]
else:
params += [{
'params': [value],
'lr': self._lr,
'weight_decay': self._weight_decay
}]
return params
class SolverWrapper(object):
"""
A wrapper class for the training process
"""
def __init__(self, sess, network, imdb, roidb, valroidb, output_dir, tbdir, pretrained_model=None):
self.net = network
self.imdb = imdb
self.roidb = roidb
self.valroidb = valroidb
self.output_dir = output_dir
self.tbdir = tbdir
# Simply put '_val' at the end to save the summaries from the validation set
self.tbvaldir = tbdir + '_val'
if not os.path.exists(self.tbvaldir):
os.makedirs(self.tbvaldir)
self.pretrained_model = pretrained_model
def snapshot(self, sess, iter):
net = self.net
if not os.path.exists(self.output_dir):
os.makedirs(self.output_dir)
# Store the model snapshot
filename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(iter) + '.ckpt'
filename = os.path.join(self.output_dir, filename)
self.saver.save(sess, filename)
print('Wrote snapshot to: {:s}'.format(filename))
# Also store some meta information, random state, etc.
nfilename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(iter) + '.pkl'
nfilename = os.path.join(self.output_dir, nfilename)
# current state of numpy random
st0 = np.random.get_state()
# current position in the database
cur = self.data_layer._cur
# current shuffled indexes of the database
perm = self.data_layer._perm
# current position in the validation database
cur_val = self.data_layer_val._cur
# current shuffled indexes of the validation database
perm_val = self.data_layer_val._perm
# Dump the meta info
with open(nfilename, 'wb') as fid:
pickle.dump(st0, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(iter, fid, pickle.HIGHEST_PROTOCOL)
return filename, nfilename
def from_snapshot(self, sess, sfile, nfile):
print('Restoring model snapshots from {:s}'.format(sfile))
self.saver.restore(sess, sfile)
print('Restored.')
# Needs to restore the other hyper-parameters/states for training, (TODO xinlei) I have
# tried my best to find the random states so that it can be recovered exactly
# However the Tensorflow state is currently not available
with open(nfile, 'rb') as fid:
st0 = pickle.load(fid)
cur = pickle.load(fid)
perm = pickle.load(fid)
cur_val = pickle.load(fid)
perm_val = pickle.load(fid)
last_snapshot_iter = pickle.load(fid)
np.random.set_state(st0)
self.data_layer._cur = cur
self.data_layer._perm = perm
self.data_layer_val._cur = cur_val
self.data_layer_val._perm = perm_val
return last_snapshot_iter
def get_variables_in_checkpoint_file(self, file_name):
try:
print('&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&')
# reader = tf.train.NewCheckpointReader(file_name)
reader = pywrap_tensorflow.NewCheckpointReader(file_name)
var_to_shape_map = reader.get_variable_to_shape_map()
return var_to_shape_map
except Exception as e: # pylint: disable=broad-except
print(str(e))
if "corrupted compressed block contents" in str(e):
print("It's likely that your checkpoint file has been compressed "
"with SNAPPY.")
def construct_graph(self, sess):
with sess.graph.as_default():
# Set the random seed for tensorflow
tf.set_random_seed(cfg.RNG_SEED)
# Build the main computation graph
layers = self.net.create_architecture('TRAIN', self.imdb.num_classes, tag='default',
anchor_scales=cfg.ANCHOR_SCALES,
anchor_ratios=cfg.ANCHOR_RATIOS)
# Define the loss
losses = layers['all_losses']
loss = losses['total_loss']
m1_loss = losses['M1']['total_loss']
m2_loss = losses['M2']['total_loss']
m3_loss = losses['M3']['total_loss']
# Set learning rate and momentum
lr = tf.Variable(cfg.TRAIN.LEARNING_RATE, trainable=False)
self.optimizer = tf.train.MomentumOptimizer(lr, cfg.TRAIN.MOMENTUM)
self.optimizer_m1 = tf.train.MomentumOptimizer(lr, cfg.TRAIN.MOMENTUM)
self.optimizer_m2 = tf.train.MomentumOptimizer(lr, cfg.TRAIN.MOMENTUM)
self.optimizer_m3 = tf.train.MomentumOptimizer(lr, cfg.TRAIN.MOMENTUM)
# Compute the gradients with regard to the loss
gvs = self.optimizer.compute_gradients(loss)
gvs_m1 = self.optimizer_m1.compute_gradients(m1_loss)
gvs_m2 = self.optimizer_m2.compute_gradients(m2_loss)
gvs_m3 = self.optimizer_m3.compute_gradients(m3_loss)
# Double the gradient of the bias if set
if cfg.TRAIN.DOUBLE_BIAS:
final_gvs = []
final_gvs_m1 = []
final_gvs_m2 = []
final_gvs_m3 = []
with tf.variable_scope('Gradient_Mult') as scope:
for grad, var in gvs:
# print("grad, var:", grad, var)
scale = 1.
if cfg.TRAIN.DOUBLE_BIAS and '/biases:' in var.name:
scale *= 2.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs.append((grad, var))
train_op = self.optimizer.apply_gradients(final_gvs)
with tf.variable_scope('Gradient_Mult_m1') as scope:
for grad, var in gvs_m1:
scale = 1.
if cfg.TRAIN.DOUBLE_BIAS and '/biases:' in var.name:
scale *= 2.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs_m1.append((grad, var))
train_m1_op = self.optimizer.apply_gradients(final_gvs_m1)
with tf.variable_scope('Gradient_Mult_m2') as scope:
for grad, var in gvs_m2:
scale = 1.
if cfg.TRAIN.DOUBLE_BIAS and '/biases:' in var.name:
scale *= 2.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs_m2.append((grad, var))
train_m2_op = self.optimizer.apply_gradients(final_gvs_m2)
with tf.variable_scope('Gradient_Mult_m3') as scope:
for grad, var in gvs_m3:
scale = 1.
if cfg.TRAIN.DOUBLE_BIAS and '/biases:' in var.name:
scale *= 2.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs_m3.append((grad, var))
train_m3_op = self.optimizer.apply_gradients(final_gvs_m3)
else:
train_op = self.optimizer.apply_gradients(gvs)
train_m1_op = self.optimizer_m1.apply_gradients(gvs_m1)
train_m2_op = self.optimizer_m2.apply_gradients(gvs_m2)
train_m3_op = self.optimizer_m3.apply_gradients(gvs_m3)
# group the three independent train_op
final_train_op = tf.group(train_m1_op, train_m2_op, train_m3_op)
# We will handle the snapshots ourselves
self.saver = tf.train.Saver(max_to_keep=100000)
# Write the train and validation information to tensorboard
self.writer = tf.summary.FileWriter(self.tbdir, sess.graph)
self.valwriter = tf.summary.FileWriter(self.tbvaldir)
return lr, train_op, train_m1_op, train_m2_op, train_m3_op, final_train_op
def find_previous(self):
sfiles = os.path.join(self.output_dir, cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.ckpt.meta')
sfiles = glob.glob(sfiles)
sfiles.sort(key=os.path.getmtime)
# Get the snapshot name in TensorFlow
redfiles = []
for stepsize in cfg.TRAIN.STEPSIZE:
redfiles.append(os.path.join(self.output_dir,
cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}.ckpt.meta'.format(stepsize + 1)))
sfiles = [ss.replace('.meta', '') for ss in sfiles if ss not in redfiles]
nfiles = os.path.join(self.output_dir, cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.pkl')
nfiles = glob.glob(nfiles)
nfiles.sort(key=os.path.getmtime)
redfiles = [redfile.replace('.ckpt.meta', '.pkl') for redfile in redfiles]
nfiles = [nn for nn in nfiles if nn not in redfiles]
lsf = len(sfiles)
assert len(nfiles) == lsf
return lsf, nfiles, sfiles
def initialize(self, sess):
# Initial file lists are empty
np_paths = []
ss_paths = []
# Fresh train directly from ImageNet weights
print('Loading initial model weights from {:s}'.format(self.pretrained_model))
variables = tf.global_variables()
print("variables:", variables)
if 'darknet53' in self.pretrained_model:
print('the base network is Darknet53!!!')
sess.run(tf.variables_initializer(variables, name='init'))
self.net.restored_from_npz(sess)
print('Loaded.')
# print('>>>>>>>', variables[0].eval())
last_snapshot_iter = 0
rate = cfg.TRAIN.LEARNING_RATE
stepsizes = list(cfg.TRAIN.STEPSIZE)
else:
sess.run(tf.variables_initializer(variables, name='init'))
var_keep_dic = self.get_variables_in_checkpoint_file(self.pretrained_model)
# Get the variables to restore, ignoring the variables to fix
variables_to_restore = self.net.get_variables_to_restore(variables, var_keep_dic)
restorer = tf.train.Saver(variables_to_restore)
restorer.restore(sess, self.pretrained_model)
print('Loaded.')
# Need to fix the variables before loading, so that the RGB weights are changed to BGR
# For VGG16 it also changes the convolutional weights
self.net.fix_variables(sess, self.pretrained_model)
print('Fixed.')
last_snapshot_iter = 0
rate = cfg.TRAIN.LEARNING_RATE
stepsizes = list(cfg.TRAIN.STEPSIZE)
return rate, last_snapshot_iter, stepsizes, np_paths, ss_paths
def restore(self, sess, sfile, nfile):
# Get the most recent snapshot and restore
variables = tf.global_variables()
print("variables:", variables)
np_paths = [nfile]
ss_paths = [sfile]
# Restore model from snapshots
last_snapshot_iter = self.from_snapshot(sess, sfile, nfile)
# Set the learning rate
rate = cfg.TRAIN.LEARNING_RATE
stepsizes = []
for stepsize in cfg.TRAIN.STEPSIZE:
if last_snapshot_iter > stepsize:
rate *= cfg.TRAIN.GAMMA
else:
stepsizes.append(stepsize)
return rate, last_snapshot_iter, stepsizes, np_paths, ss_paths
def remove_snapshot(self, np_paths, ss_paths):
to_remove = len(np_paths) - cfg.TRAIN.SNAPSHOT_KEPT
for c in range(to_remove):
nfile = np_paths[0]
os.remove(str(nfile))
np_paths.remove(nfile)
to_remove = len(ss_paths) - cfg.TRAIN.SNAPSHOT_KEPT
for c in range(to_remove):
sfile = ss_paths[0]
# To make the code compatible to earlier versions of Tensorflow,
# where the naming tradition for checkpoints are different
if os.path.exists(str(sfile)):
os.remove(str(sfile))
else:
os.remove(str(sfile + '.data-00000-of-00001'))
os.remove(str(sfile + '.index'))
sfile_meta = sfile + '.meta'
os.remove(str(sfile_meta))
ss_paths.remove(sfile)
def train_model(self, sess, max_iters):
# Build data layers for both training and validation set
self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
self.data_layer_val = RoIDataLayer(self.valroidb, self.imdb.num_classes, random=True)
# Construct the computation graph
# lr, train_op = self.construct_graph(sess)
lr, train_op, train_m1_op, train_m2_op, train_m3_op, final_train_op = self.construct_graph(sess)
# Find previous snapshots if there is any to restore from
lsf, nfiles, sfiles = self.find_previous()
# Initialize the variables or restore them from the last snapshot
if lsf == 0:
rate, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.initialize(sess)
else:
rate, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.restore(sess,
str(sfiles[-1]),
str(nfiles[-1]))
timer = Timer()
iter = last_snapshot_iter + 1
last_summary_time = time.time()
# Make sure the lists are not empty
stepsizes.append(max_iters)
stepsizes.reverse()
next_stepsize = stepsizes.pop()
while iter < max_iters + 1:
# Learning rate
if iter == next_stepsize + 1:
# Add snapshot here before reducing the learning rate
self.snapshot(sess, iter)
rate *= cfg.TRAIN.GAMMA
sess.run(tf.assign(lr, rate))
next_stepsize = stepsizes.pop()
timer.tic()
# Get training data, one batch at a time
blobs = self.data_layer.forward()
now = time.time()
if iter == 1 or now - last_summary_time > cfg.TRAIN.SUMMARY_INTERVAL:
# Compute the graph with summary
losses, summary = self.net.train_step_with_summary(sess, blobs, final_train_op)
self.writer.add_summary(summary, float(iter))
# Also check the summary on the validation set
blobs_val = self.data_layer_val.forward()
summary_val = self.net.get_summary(sess, blobs_val)
self.valwriter.add_summary(summary_val, float(iter))
last_summary_time = now
else:
# Compute the graph without summary
losses = self.net.train_step(sess, blobs, final_train_op)
timer.toc()
# get the corresponding loss to show
m1_cls_loss = losses['M1']['rpn_cross_entropy']
m1_box_loss = losses['M1']['rpn_loss_box']
m1_kp_loss = losses['M1']['kpoints_loss']
m1_total_loss = losses['M1']['total_loss']
m2_cls_loss = losses['M2']['rpn_cross_entropy']
m2_box_loss = losses['M2']['rpn_loss_box']
m2_kp_loss = losses['M2']['kpoints_loss']
m2_total_loss = losses['M2']['total_loss']
m3_cls_loss = losses['M3']['rpn_cross_entropy']
m3_box_loss = losses['M3']['rpn_loss_box']
m3_kp_loss = losses['M3']['kpoints_loss']
m3_total_loss = losses['M3']['total_loss']
total_loss = losses['total_loss']
# Display training information
if iter % (cfg.TRAIN.DISPLAY) == 0:
print('iter: %d / %d \n >>> m1_cls_loss: %.6f, m1_box_loss: %.6f, m1_kp_loss: %.6f, m1_total_loss: %.6f\n '
'>>> m2_cls_loss: %.6f, m2_box_loss: %.6f, m2_kp_loss: %.6f, m2_total_loss: %.6f\n '
'>>> m3_cls_loss: %.6f, m3_box_loss: %.6f, m3_kp_loss: %.6f, m3_total_loss: %.6f\n '
'>>> total_loss: %.6f, lr: %f' % \
(iter, max_iters, m1_cls_loss, m1_box_loss, m1_kp_loss, m1_total_loss, m2_cls_loss, m2_box_loss, m2_kp_loss, m2_total_loss,
m3_cls_loss, m3_box_loss, m3_kp_loss, m3_total_loss, total_loss, lr.eval()))
print('speed: {:.3f}s / iter'.format(timer.average_time))
# Snapshotting
if iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:
last_snapshot_iter = iter
ss_path, np_path = self.snapshot(sess, iter)
np_paths.append(np_path)
ss_paths.append(ss_path)
# Remove the old snapshots if there are too many
if len(np_paths) > cfg.TRAIN.SNAPSHOT_KEPT:
self.remove_snapshot(np_paths, ss_paths)
iter += 1
if last_snapshot_iter != iter - 1:
self.snapshot(sess, iter - 1)
self.writer.close()
self.valwriter.close()
def train_model_old(self, sess, max_iters):
# Build data layers for both training and validation set
self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
self.data_layer_val = RoIDataLayer(self.valroidb, self.imdb.num_classes, random=True)
# Construct the computation graph
lr, train_op, train_m1_op, train_m2_op, train_m3_op, _ = self.construct_graph(sess)
# Find previous snapshots if there is any to restore from
lsf, nfiles, sfiles = self.find_previous()
# Initialize the variables or restore them from the last snapshot
if lsf == 0:
rate, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.initialize(sess)
else:
rate, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.restore(sess,
str(sfiles[-1]),
str(nfiles[-1]))
timer = Timer()
iter = last_snapshot_iter + 1
last_summary_time = time.time()
# Make sure the lists are not empty
stepsizes.append(max_iters)
stepsizes.reverse()
next_stepsize = stepsizes.pop()
m1_iters = 0
m2_iters = 0
m3_iters = 0
while iter < max_iters + 1:
random_seed = np.random.rand()
if random_seed < 0.33:
module = "M1"
train_op = train_m1_op
m1_iters += 1
elif 0.33 <= random_seed < 0.67:
module = "M2"
train_op = train_m2_op
m2_iters += 1
else:
module = "M3"
train_op = train_m3_op
m3_iters += 1
# Learning rate
if iter == next_stepsize + 1:
# Add snapshot here before reducing the learning rate
self.snapshot(sess, iter)
rate *= cfg.TRAIN.GAMMA
sess.run(tf.assign(lr, rate))
next_stepsize = stepsizes.pop()
timer.tic()
# Get training data, one batch at a time
blobs = self.data_layer.forward()
now = time.time()
if iter == 1 or now - last_summary_time > cfg.TRAIN.SUMMARY_INTERVAL:
# Compute the graph with summary
rpn_loss_cls, rpn_loss_box, landmarks_loss, total_loss, summary = self.net.train_step_with_summary_old(sess, blobs,
train_op, module)
self.writer.add_summary(summary, float(iter))
# Also check the summary on the validation set
blobs_val = self.data_layer_val.forward()
summary_val = self.net.get_summary(sess, blobs_val)
self.valwriter.add_summary(summary_val, float(iter))
last_summary_time = now
else:
# Compute the graph without summary
rpn_loss_cls, rpn_loss_box, kpoints_loss, total_loss = self.net.train_step_old(sess, blobs, train_op, module)
timer.toc()
# Display training information
if iter % (cfg.TRAIN.DISPLAY) == 0:
if module == 'M1':
iters = m1_iters
elif module == 'M2':
iters = m2_iters
else:
iters = m3_iters
print('iter: %d / %d, now training module: %s, iters: %d,\n >>> total loss: %.6f\n >>> rpn_loss_cls: %.6f\n '
'>>> rpn_loss_box: %.6f\n >>> kpoints_loss: %.6f\n >>> lr: %f' % \
(iter, max_iters, module, iters, total_loss, rpn_loss_cls, rpn_loss_box, kpoints_loss, lr.eval()))
print('speed: {:.3f}s / iter'.format(timer.average_time))
# Snapshotting
if iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:
last_snapshot_iter = iter
ss_path, np_path = self.snapshot(sess, iter)
np_paths.append(np_path)
ss_paths.append(ss_path)
# Remove the old snapshots if there are too many
if len(np_paths) > cfg.TRAIN.SNAPSHOT_KEPT:
self.remove_snapshot(np_paths, ss_paths)
iter += 1
if last_snapshot_iter != iter - 1:
self.snapshot(sess, iter - 1)
self.writer.close()
self.valwriter.close()
class SolverWrapper(object):
def __init__(self,
network,
imdb,
roidb,
valroidb,
model_dir,
pretrained_model=None):
self.net = network
self.imdb = imdb
self.roidb = roidb
self.valroidb = valroidb
self.model_dir = model_dir
self.tbdir = os.path.join(model_dir, 'train_log')
if not os.path.exists(self.tbdir):
os.makedirs(self.tbdir)
self.pretrained_model = pretrained_model
def set_learn_strategy(self, learn_dict):
self._disp_interval = learn_dict['disp_interval']
self._valid_interval = learn_dict['disp_interval'] * 5
self._use_tensorboard = learn_dict['use_tensorboard']
self._use_valid = learn_dict['use_valid']
self._save_point_interval = learn_dict['save_point_interval']
self._lr_decay_steps = learn_dict['lr_decay_steps']
def train_model(self, resume=None, max_iters=100000):
# Build data layers for both training and validation set
self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
self.data_layer_val = RoIDataLayer(self.valroidb,
self.imdb.num_classes,
random=True)
self.prepare_construct(resume)
net = self.net
# training
train_loss = 0
rpn_cls_loss = 0
rpn_bbox_loss = 0
fast_rcnn_cls_loss = 0
fast_rcnn_bbox_loss = 0
tp, tf, fg, bg = 0., 0., 0, 0
step_cnt = 0
re_cnt = False
t = Timer()
t.tic()
for step in range(self.start_step, max_iters + 1):
blobs = self.data_layer.forward()
im_data = blobs['data']
im_info = blobs['im_info']
gt_boxes = blobs['gt_boxes']
# forward
result_cls_prob, result_bbox_pred, result_rois = net(
im_data, im_info, gt_boxes)
loss = net.loss + net._rpn.loss
train_loss += loss.data.cpu()[0]
rpn_cls_loss += net._rpn.cross_entropy.data.cpu()[0]
rpn_bbox_loss += net._rpn.loss_box.data.cpu()[0]
fast_rcnn_cls_loss += net.cross_entropy.data.cpu()[0]
fast_rcnn_bbox_loss += net.loss_box.data.cpu()[0]
step_cnt += 1
# backward
self._optimizer.zero_grad()
loss.backward()
nn.utils.clip_grad_norm(self._parameters, max_norm=10)
self._optimizer.step()
# clear middle memory
net._delete_cache()
if step % self._disp_interval == 0:
duration = t.toc(average=False)
fps = step_cnt / duration
log_text = 'step %d, image: %s, loss: %.4f, fps: %.2f (%.2fs per batch)' % (
step, blobs['im_name'], train_loss / step_cnt, fps,
1. / fps)
pprint.pprint(log_text)
if self._use_tensorboard:
self._tensor_writer.add_text('Train',
log_text,
global_step=step)
# Train
avg_rpn_cls_loss = rpn_cls_loss / step_cnt
avg_rpn_bbox_loss = rpn_bbox_loss / step_cnt
avg_fast_rcnn_cls_loss = fast_rcnn_cls_loss / step_cnt
avg_fast_rcnn_bbox_loss = fast_rcnn_bbox_loss / step_cnt
self._tensor_writer.add_scalars(
'TrainSetLoss', {
'RPN_cls_loss': avg_rpn_cls_loss,
'RPN_bbox_loss': avg_rpn_bbox_loss,
'FastRcnn_cls_loss': avg_fast_rcnn_cls_loss,
'FastRcnn_bbox_loss': avg_fast_rcnn_bbox_loss
},
global_step=step)
self._tensor_writer.add_scalar('Learning_rate',
self._lr,
global_step=step)
re_cnt = True
if self._use_tensorboard and step % self._valid_interval == 0:
new_gt_boxes = gt_boxes.copy()
new_gt_boxes[:, :4] = new_gt_boxes[:, :4]
image = self.back_to_image(blobs['data']).astype(np.uint8)
im_shape = image.shape
pred_boxes, scores, classes = net.interpret_faster_rcnn_scale(
result_cls_prob,
result_bbox_pred,
result_rois,
im_shape,
min_score=0.1)
image = self.draw_photo(image, pred_boxes, scores, classes,
new_gt_boxes)
image = torchtrans.ToTensor()(image)
image = vutils.make_grid([image])
self._tensor_writer.add_image('Image', image, step)
if self._use_valid and step % self._valid_interval == 0:
total_valid_loss = 0.0
valid_rpn_cls_loss = 0.0
valid_rpn_bbox_loss = 0.0
valid_fast_rcnn_cls_loss = 0.0
valid_fast_rcnn_bbox_loss = 0.0
valid_step_cnt = 0
start_time = time.time()
valid_length = self._disp_interval
net.eval()
for valid_batch in range(valid_length):
# get one batch
blobs = self.data_layer_val.forward()
im_data = blobs['data']
im_info = blobs['im_info']
gt_boxes = blobs['gt_boxes']
# forward
result_cls_prob, result_bbox_pred, result_rois = net(
im_data, im_info, gt_boxes)
valid_loss = net.loss + net._rpn.loss
total_valid_loss += valid_loss.data.cpu()[0]
valid_rpn_cls_loss += net._rpn.cross_entropy.data.cpu()[0]
valid_rpn_bbox_loss += net._rpn.loss_box.data.cpu()[0]
valid_fast_rcnn_cls_loss += net.cross_entropy.data.cpu()[0]
valid_fast_rcnn_bbox_loss += net.loss_box.data.cpu()[0]
valid_step_cnt += 1
net.train()
duration = time.time() - start_time
fps = valid_step_cnt / duration
log_text = 'step %d, valid average loss: %.4f, fps: %.2f (%.2fs per batch)' % (
step, total_valid_loss / valid_step_cnt, fps, 1. / fps)
pprint.pprint(log_text)
if self._use_tensorboard:
self._tensor_writer.add_text('Valid',
log_text,
global_step=step)
new_gt_boxes = gt_boxes.copy()
new_gt_boxes[:, :4] = new_gt_boxes[:, :4]
image = self.back_to_image(blobs['data']).astype(np.uint8)
im_shape = image.shape
pred_boxes, scores, classes = net.interpret_faster_rcnn_scale(
result_cls_prob,
result_bbox_pred,
result_rois,
im_shape,
min_score=0.1)
image = self.draw_photo(image, pred_boxes, scores, classes,
new_gt_boxes)
image = torchtrans.ToTensor()(image)
image = vutils.make_grid([image])
self._tensor_writer.add_image('Image_Valid', image, step)
if self._use_tensorboard:
# Valid
avg_rpn_cls_loss_valid = valid_rpn_cls_loss / valid_step_cnt
avg_rpn_bbox_loss_valid = valid_rpn_bbox_loss / valid_step_cnt
avg_fast_rcnn_cls_loss_valid = valid_fast_rcnn_cls_loss / valid_step_cnt
avg_fast_rcnn_bbox_loss_valid = valid_fast_rcnn_bbox_loss / valid_step_cnt
real_total_loss_valid = valid_rpn_cls_loss + valid_rpn_bbox_loss + valid_fast_rcnn_cls_loss + valid_fast_rcnn_bbox_loss
# Train
avg_rpn_cls_loss = rpn_cls_loss / step_cnt
avg_rpn_bbox_loss = rpn_bbox_loss / step_cnt
avg_fast_rcnn_cls_loss = fast_rcnn_cls_loss / step_cnt
avg_fast_rcnn_bbox_loss = fast_rcnn_bbox_loss / step_cnt
real_total_loss = rpn_cls_loss + rpn_bbox_loss + fast_rcnn_cls_loss + fast_rcnn_bbox_loss
self._tensor_writer.add_scalars(
'Total_Loss', {
'train': train_loss / step_cnt,
'valid': total_valid_loss / valid_step_cnt
},
global_step=step)
self._tensor_writer.add_scalars(
'Real_loss', {
'train': real_total_loss / step_cnt,
'valid': real_total_loss_valid / valid_step_cnt
},
global_step=step)
self._tensor_writer.add_scalars(
'RPN_cls_loss', {
'train': avg_rpn_cls_loss,
'valid': avg_rpn_cls_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'RPN_bbox_loss', {
'train': avg_rpn_bbox_loss,
'valid': avg_rpn_bbox_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'FastRcnn_cls_loss', {
'train': avg_fast_rcnn_cls_loss,
'valid': avg_fast_rcnn_cls_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'FastRcnn_bbox_loss', {
'train': avg_fast_rcnn_bbox_loss,
'valid': avg_fast_rcnn_bbox_loss_valid
},
global_step=step)
self._tensor_writer.add_scalars(
'ValidSetLoss', {
'RPN_cls_loss': avg_rpn_cls_loss_valid,
'RPN_bbox_loss': avg_rpn_bbox_loss_valid,
'FastRcnn_cls_loss': avg_fast_rcnn_cls_loss_valid,
'FastRcnn_bbox_loss': avg_fast_rcnn_bbox_loss_valid
},
global_step=step)
# self._tensor_writer.add_scalars('TrainSetLoss', {
# 'RPN_cls_loss': avg_rpn_cls_loss,
# 'RPN_bbox_loss': avg_rpn_bbox_loss,
# 'FastRcnn_cls_loss': avg_fast_rcnn_cls_loss,
# 'FastRcnn_bbox_loss': avg_fast_rcnn_bbox_loss
# }, global_step=step)
# self._tensor_writer.add_scalar('Learning_rate', self._lr, global_step=step)
if (step % self._save_point_interval == 0) and step > 0:
save_name, _ = self.save_check_point(step)
print('save model: {}'.format(save_name))
if step in self._lr_decay_steps:
self._lr *= self._lr_decay
self._optimizer = self._train_optimizer()
if re_cnt:
tp, tf, fg, bg = 0., 0., 0, 0
train_loss = 0
rpn_cls_loss = 0
rpn_bbox_loss = 0
fast_rcnn_cls_loss = 0
fast_rcnn_bbox_loss = 0
step_cnt = 0
t.tic()
re_cnt = False
if self._use_tensorboard:
self._tensor_writer.export_scalars_to_json(
os.path.join(self.tbdir, 'all_scalars.json'))
def draw_photo(self, image, dets, scores, classes, gt_boxes):
# im2show = np.copy(image)
im2show = image
# color_b = (0, 191, 255)
for i, det in enumerate(dets):
det = tuple(int(x) for x in det)
r = min(0 + i * 10, 255)
r_i = i / 5
g = min(150 + r_i * 10, 255)
g_i = r_i / 5
b = min(200 + g_i, 255)
color_b_c = (r, g, b)
cv2.rectangle(im2show, det[0:2], det[2:4], color_b_c, 2)
cv2.putText(im2show,
'%s: %.3f' % (classes[i], scores[i]),
(det[0], det[1] + 15),
cv2.FONT_HERSHEY_PLAIN,
1.0, (0, 0, 255),
thickness=1)
for i, det in enumerate(gt_boxes):
det = tuple(int(x) for x in det)
gt_class = self.net._classes[det[-1]]
cv2.rectangle(im2show, det[0:2], det[2:4], (255, 0, 0), 2)
cv2.putText(im2show,
'%s' % (gt_class), (det[0], det[1] + 15),
cv2.FONT_HERSHEY_PLAIN,
1.0, (0, 0, 255),
thickness=1)
return im2show
def back_to_image(self, img):
image = img[0] + cfg.PIXEL_MEANS
image = image[:, :, ::-1].copy(order='C')
return image
def save_check_point(self, step):
net = self.net
if not os.path.exists(self.model_dir):
os.makedirs(self.model_dir)
# store the model snapshot
filename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.h5'.format(step))
h5f = h5py.File(filename, mode='w')
for k, v in net.state_dict().items():
h5f.create_dataset(k, data=v.cpu().numpy())
# store data information
nfilename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.pkl'.format(step))
# current state of numpy random
st0 = np.random.get_state()
# current position in the database
cur = self.data_layer._cur
# current shuffled indexes of the database
perm = self.data_layer._perm
# current position in the validation database
cur_val = self.data_layer_val._cur
# current shuffled indexes of the validation database
perm_val = self.data_layer_val._perm
# current learning rate
lr = self._lr
# Dump the meta info
with open(nfilename, 'wb') as fid:
pickle.dump(st0, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(cur_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(perm_val, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(lr, fid, pickle.HIGHEST_PROTOCOL)
pickle.dump(step, fid, pickle.HIGHEST_PROTOCOL)
return filename, nfilename
def load_check_point(self, step):
net = self.net
filename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.h5'.format(step))
nfilename = os.path.join(self.model_dir,
'fasterRcnn_iter_{}.pkl'.format(step))
print('Restoring model snapshots from {:s}'.format(filename))
if not os.path.exists(filename):
print('The checkPoint is not Right')
sys.exit(1)
# load model
h5f = h5py.File(filename, mode='r')
for k, v in net.state_dict().items():
param = torch.from_numpy(np.asarray(h5f[k]))
v.copy_(param)
# load data information
with open(nfilename, 'rb') as fid:
st0 = pickle.load(fid)
cur = pickle.load(fid)
perm = pickle.load(fid)
cur_val = pickle.load(fid)
perm_val = pickle.load(fid)
lr = pickle.load(fid)
last_snapshot_iter = pickle.load(fid)
np.random.set_state(st0)
self.data_layer._cur = cur
self.data_layer._perm = perm
self.data_layer_val._cur = cur_val
self.data_layer_val._perm = perm_val
self._lr = lr
if last_snapshot_iter == step:
print('Restore over ')
else:
print('The checkPoint is not Right')
raise ValueError
return last_snapshot_iter
def weights_normal_init(self, model, dev=0.01):
import math
def _gaussian_init(m, dev):
m.weight.data.normal_(0.0, dev)
if hasattr(m.bias, 'data'):
m.bias.data.zero_()
def _xaiver_init(m):
nn.init.xavier_normal(m.weight.data)
if hasattr(m.bias, 'data'):
m.bias.data.zero_()
def _hekaiming_init(m):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
if hasattr(m.bias, 'data'):
m.bias.data.zero_()
def _resnet_init(model, dev):
if isinstance(model, list):
for m in model:
self.weights_normal_init(m, dev)
else:
for m in model.modules():
if isinstance(m, nn.Conv2d):
_hekaiming_init(m)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
elif isinstance(m, nn.Linear):
_gaussian_init(m, dev)
def _vgg_init(model, dev):
if isinstance(model, list):
for m in model:
self.weights_normal_init(m, dev)
else:
for m in model.modules():
if isinstance(m, nn.Conv2d):
_gaussian_init(m, dev)
elif isinstance(m, nn.Linear):
_gaussian_init(m, dev)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
if cfg.TRAIN.INIT_WAY == 'resnet':
_vgg_init(model, dev)
elif cfg.TRAIN.INIT_WAY == 'vgg':
_vgg_init(model, dev)
else:
raise NotImplementedError
def prepare_construct(self, resume_iter):
# init network
self.net.init_fasterRCNN()
# Set the random seed
torch.manual_seed(cfg.RNG_SEED)
np.random.seed(cfg.RNG_SEED)
# Set learning rate and momentum
self._lr = cfg.TRAIN.LEARNING_RATE
self._lr_decay = 0.1
self._momentum = cfg.TRAIN.MOMENTUM
self._weight_decay = cfg.TRAIN.WEIGHT_DECAY
# load model
if resume_iter:
self.start_step = resume_iter + 1
self.load_check_point(resume_iter)
else:
self.start_step = 0
self.weights_normal_init(self.net, dev=0.01)
# refer to caffe faster RCNN
self.net.init_special_bbox_fc(dev=0.001)
if self.pretrained_model != None:
self.net._rpn._network._load_pre_trained_model(
self.pretrained_model)
print('Load parameters from Path: {}'.format(
self.pretrained_model))
else:
pass
# model
self.net.train()
if cfg.CUDA_IF:
self.net.cuda()
# resnet fixed BN should be eval
if cfg.TRAIN.INIT_WAY == 'resnet':
self.net._rpn._network._bn_eval()
# set optimizer
self._parameters = [
params for params in self.net.parameters()
if params.requires_grad == True
]
self._optimizer = self._train_optimizer()
# tensorboard
if self._use_tensorboard:
import tensorboardX as tbx
self._tensor_writer = tbx.SummaryWriter(log_dir=self.tbdir)
def _train_optimizer(self):
parameters = self._train_parameter()
optimizer = torch.optim.SGD(parameters, momentum=self._momentum)
return optimizer
def _train_parameter(self):
params = []
for key, value in self.net.named_parameters():
if value.requires_grad == True:
if 'bias' in key:
params += [{
'params': [value],
'lr': self._lr * (cfg.TRAIN.DOUBLE_BIAS + 1),
'weight_decay': 0
}]
else:
params += [{
'params': [value],
'lr': self._lr,
'weight_decay': self._weight_decay
}]
return params
class SolverWrapper(object):
"""
A wrapper class for the training process
据作者的说法，这个类就是为了方便自己使用Python代码来控制训练过程中的相关东西
"""
def __init__(self, sess, network, imdb, roidb, valroidb, output_dir,
pretrained_model):
self.net = network #网络 vgg 或者 resnet
self.imdb = imdb #数据库
self.roidb = roidb #region of insterest
self.valroidb = valroidb #tensorboard 输出文件
self.output_dir = output_dir #结果输出文件
#self.tbdir = tbdir
# Simply put '_val' at the end to save the summaries from the validation set
#self.tbvaldir = tbdir + '_val'
#if not os.path.exists(self.tbvaldir):
#os.makedirs(self.tbvaldir)
self.pretrained_model = pretrained_model
def snapshot(self, sess, iter): #这个函数做一些存储 存储
net = self.net
if not os.path.exists(self.output_dir):
os.makedirs(self.output_dir)
# Store the model snapshot
filename = cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_{:d}'.format(
iter
) + '.ckpt' #__C.TRAIN.SNAPSHOT_PREFIX = 'res101_faster_rcnn' 默认输出模型
filename = os.path.join(self.output_dir, filename) #存在output下面
self.saver.save(sess, filename) #存储tensor变量
print('Wrote snapshot to: {:s}'.format(filename))
return filename
def from_snapshot(self, sess, sfile, nfile):
print('Restoring model snapshots from {:s}'.format(sfile))
self.saver.restore(sess, sfile)
print('Restored.')
# Needs to restore the other hyper-parameters/states for training, (TODO xinlei) I have
# tried my best to find the random states so that it can be recovered exactly
# However the Tensorflow state is currently not available
with open(nfile, 'rb') as fid:
st0 = pickle.load(fid)
cur = pickle.load(fid)
perm = pickle.load(fid)
cur_val = pickle.load(fid)
perm_val = pickle.load(fid)
last_snapshot_iter = pickle.load(fid)
np.random.set_state(st0)
self.data_layer._cur = cur
self.data_layer._perm = perm
self.data_layer_val._cur = cur_val
self.data_layer_val._perm = perm_val
return last_snapshot_iter
def get_variables_in_checkpoint_file(self, file_name): #初始化模型 文件
try:
reader = pywrap_tensorflow.NewCheckpointReader(file_name)
var_to_shape_map = reader.get_variable_to_shape_map()
return var_to_shape_map
except Exception as e: # pylint: disable=broad-except
print(str(e))
if "corrupted compressed block contents" in str(e):
print(
"It's likely that your checkpoint file has been compressed "
"with SNAPPY.")
def construct_graph(self, sess): #该函数构建计算图
with sess.graph.as_default():
# Set the random seed for tensorflow
tf.set_random_seed(cfg.RNG_SEED) #随机种子
# Build the main computation graph
layers = self.net.create_architecture(
'TRAIN',
self.imdb.num_classes,
anchor_scales=cfg.ANCHOR_SCALES,
anchor_ratios=cfg.ANCHOR_RATIOS)
# Define the loss
loss = layers['total_loss']
# Set learning rate and momentum
lr = tf.Variable(cfg.TRAIN.LEARNING_RATE, trainable=False) #学习率
self.optimizer = tf.train.MomentumOptimizer(
lr, cfg.TRAIN.MOMENTUM) #梯度优化器
# Compute the gradients with regard to the loss
gvs = self.optimizer.compute_gradients(loss)
# Double the gradient of the bias if set
if cfg.TRAIN.DOUBLE_BIAS:
final_gvs = []
with tf.variable_scope('Gradient_Mult') as scope:
for grad, var in gvs:
scale = 1.
if cfg.TRAIN.DOUBLE_BIAS and '/biases:' in var.name:
scale *= 2.
if not np.allclose(scale, 1.0):
grad = tf.multiply(grad, scale)
final_gvs.append((grad, var))
train_op = self.optimizer.apply_gradients(final_gvs)
else:
train_op = self.optimizer.apply_gradients(gvs)
# We will handle the snapshots ourselves
self.saver = tf.train.Saver(max_to_keep=100000)
# Write the train and validation information to tensorboard
#self.writer = tf.summary.FileWriter(self.tbdir, sess.graph)
#self.valwriter = tf.summary.FileWriter(self.tbvaldir)
print("构建网络通过")
return lr, train_op
def find_previous(self):
sfiles = os.path.join(self.output_dir,
cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.ckpt.meta')
sfiles = glob.glob(sfiles)
sfiles.sort(key=os.path.getmtime)
# Get the snapshot name in TensorFlow
redfiles = []
for stepsize in cfg.TRAIN.STEPSIZE:
redfiles.append(
os.path.join(
self.output_dir, cfg.TRAIN.SNAPSHOT_PREFIX +
'_iter_{:d}.ckpt.meta'.format(stepsize + 1)))
sfiles = [
ss.replace('.meta', '') for ss in sfiles if ss not in redfiles
]
nfiles = os.path.join(self.output_dir,
cfg.TRAIN.SNAPSHOT_PREFIX + '_iter_*.pkl')
nfiles = glob.glob(nfiles)
nfiles.sort(key=os.path.getmtime)
redfiles = [
redfile.replace('.ckpt.meta', '.pkl') for redfile in redfiles
]
nfiles = [nn for nn in nfiles if nn not in redfiles]
lsf = len(sfiles)
assert len(nfiles) == lsf
return lsf, nfiles, sfiles
def initialize(self, sess):
# Initial file lists are empty
np_paths = []
ss_paths = []
# Fresh train directly from ImageNet weights
print('Loading initial model weights from {:s}'.format(
self.pretrained_model))
variables = tf.global_variables()
# Initialize all variables first
sess.run(tf.variables_initializer(variables, name='init'))
var_keep_dic = self.get_variables_in_checkpoint_file(
self.pretrained_model) #把预训练网络参数 拿出来
# Get the variables to restore, ignoring the variables to fix
variables_to_restore = self.net.get_variables_to_restore(
variables, var_keep_dic)
restorer = tf.train.Saver(variables_to_restore)
restorer.restore(sess, self.pretrained_model)
print('Loaded.')
# Need to fix the variables before loading, so that the RGB weights are changed to BGR
# For VGG16 it also changes the convolutional weights fc6 and fc7 to
# fully connected weights
self.net.fix_variables(sess, self.pretrained_model)
print('Fixed.')
last_snapshot_iter = 0
rate = cfg.TRAIN.LEARNING_RATE
stepsizes = list(cfg.TRAIN.STEPSIZE)
return rate, last_snapshot_iter, stepsizes, np_paths, ss_paths
def restore(self, sess, sfile, nfile):
# Get the most recent snapshot and restore
np_paths = [nfile]
ss_paths = [sfile]
# Restore model from snapshots
last_snapshot_iter = self.from_snapshot(sess, sfile, nfile)
# Set the learning rate
rate = cfg.TRAIN.LEARNING_RATE
stepsizes = []
for stepsize in cfg.TRAIN.STEPSIZE:
if last_snapshot_iter > stepsize:
rate *= cfg.TRAIN.GAMMA
else:
stepsizes.append(stepsize)
return rate, last_snapshot_iter, stepsizes, np_paths, ss_paths
def remove_snapshot(self, np_paths, ss_paths):
to_remove = len(np_paths) - cfg.TRAIN.SNAPSHOT_KEPT
for c in range(to_remove):
nfile = np_paths[0]
os.remove(str(nfile))
np_paths.remove(nfile)
to_remove = len(ss_paths) - cfg.TRAIN.SNAPSHOT_KEPT
for c in range(to_remove):
sfile = ss_paths[0]
# To make the code compatible to earlier versions of Tensorflow,
# where the naming tradition for checkpoints are different
if os.path.exists(str(sfile)):
os.remove(str(sfile))
else:
os.remove(str(sfile + '.data-00000-of-00001'))
os.remove(str(sfile + '.index'))
sfile_meta = sfile + '.meta'
os.remove(str(sfile_meta))
ss_paths.remove(sfile)
def train_model(self, sess, max_iters): #这个是 训练的核心函数
# Build data layers for both training and validation set
self.data_layer = RoIDataLayer(self.roidb, self.imdb.num_classes)
print("训练准备数据通过")
self.data_layer_val = RoIDataLayer(self.valroidb,
self.imdb.num_classes,
random=True)
print("测试数据通过")
# Construct the computation graph
lr, train_op = self.construct_graph(sess) #构建网络通过
# Find previous snapshots if there is any to restore from
lsf, nfiles, sfiles = self.find_previous()
# Initialize the variables or restore them from the last snapshot
if lsf == 0:
rate, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.initialize(
sess)
else:
rate, last_snapshot_iter, stepsizes, np_paths, ss_paths = self.restore(
sess, str(sfiles[-1]), str(nfiles[-1]))
timer = Timer()
iter = last_snapshot_iter + 1
last_summary_time = time.time()
# Make sure the lists are not empty
stepsizes.append(max_iters)
stepsizes.reverse()
next_stepsize = stepsizes.pop()
while iter < max_iters + 1:
# Learning rate
if iter == next_stepsize + 1:
# Add snapshot here before reducing the learning rate
self.snapshot(sess, iter)
rate *= cfg.TRAIN.GAMMA
sess.run(tf.assign(lr, rate))
next_stepsize = stepsizes.pop()
timer.tic()
# Get training data, one batch at a time
blobs = self.data_layer.forward(
) #这里开始报错 解决（原因是忘了加image 属性 导致没读出数据） #验证集不需要数据扩增 ！！前向运算
now = time.time()
if iter == 1 or now - last_summary_time > cfg.TRAIN.SUMMARY_INTERVAL:
# Compute the graph with summary
rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, total_loss, summary = \
self.net.train_step_with_summary(sess, blobs, train_op)
# Also check the summary on the validation set
blobs_val = self.data_layer_val.forward() #验证集的运算
last_summary_time = now
else:
# Compute the graph without summary
rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, total_loss = \
self.net.train_step(sess, blobs, train_op)
timer.toc()
# Display training information
if iter % (cfg.TRAIN.DISPLAY) == 0:
print('iter: %d / %d, total loss: %.6f\n >>> rpn_loss_cls: %.6f\n '
'>>> rpn_loss_box: %.6f\n >>> loss_cls: %.6f\n >>> loss_box: %.6f\n >>> lr: %f' % \
(iter, max_iters, total_loss, rpn_loss_cls, rpn_loss_box, loss_cls, loss_box, lr.eval()))
print('speed: {:.3f}s / iter'.format(timer.average_time))
# Snapshotting
if iter % cfg.TRAIN.SNAPSHOT_ITERS == 0:
last_snapshot_iter = iter
ss_path, np_path = self.snapshot(sess, iter)
np_paths.append(np_path)
ss_paths.append(ss_path)
# Remove the old snapshots if there are too many
if len(np_paths) > cfg.TRAIN.SNAPSHOT_KEPT:
self.remove_snapshot(np_paths, ss_paths)
iter += 1
if last_snapshot_iter != iter - 1:
self.snapshot(sess, iter - 1)
self.writer.close()
self.valwriter.close()
