def run(self, session):
self.model.setup_testing(session)
if self.args.use_dataset:
test_data= data_parser(self.args)
n_data = len(test_data[1])
else:
test_data=get_single_image(self.args)
n_data = len(test_data)
print_info('Writing PNGs at {}'.format(self.args.base_dir_results))
if self.args.batch_size_test==1 and self.args.use_dataset:
for i in range(n_data):
im, em, file_name = get_testing_batch(self.args,
[test_data[0][test_data[1][i]], test_data[1][i]], use_batch=False)
self.img_info = file_name
edgemap = session.run(self.model.predictions, feed_dict={self.model.images: [im]})
self.save_egdemaps(edgemap, single_image=True)
print_info('Done testing {}, {}'.format(self.img_info[0], self.img_info[1]))
# for individual images
elif self.args.batch_size_test==1 and not self.args.use_dataset:
for i in range(n_data):
im, file_name = get_single_image(self.args,file_path=test_data[i])
self.img_info = file_name
edgemap = session.run(self.model.predictions, feed_dict={self.model.images: [im]})
self.save_egdemaps(edgemap, single_image=True)
print_info('Done testing {}, {}'.format(self.img_info[0], self.img_info[1]))
def run(self, session):
self.model.setup_testing(session)
if self.args.use_dataset:
test_data= data_parser(self.args)
n_data = len(test_data[1])
else:
test_data=get_single_image(self.args)
n_data = len(test_data)
print_info('Writing PNGs at {}'.format(self.args.base_dir_results))
if self.args.batch_size_test==1 and self.args.use_dataset:
for i in range(n_data):
im, em, file_name = get_testing_batch(self.args,
[test_data[0][test_data[1][i]], test_data[1][i]], use_batch=False)
self.img_info = file_name
#Dexi Start Time
startDexi = time.time()
#Edge map creation from the pretrained model
edgemap = session.run(self.model.predictions, feed_dict={self.model.images: [im]})
#Dexi End Time
endDexi = time.time()
secondsDexi = endDexi - startDexi
print_info('Time taken for DexiNED: {} seconds'.format(secondsDexi))
self.save_egdemaps(edgemap, single_image=True)
print_info('Done testing {}, {}'.format(self.img_info[0], self.img_info[1]))
# for individual images
elif self.args.batch_size_test==1 and not self.args.use_dataset:
for i in range(n_data):
im, file_name = get_single_image(self.args,file_path=test_data[i])
self.img_info = file_name
#Dexi Start Time
startDexi = time.time()
edgemap = session.run(self.model.predictions, feed_dict={self.model.images: [im]})
#Dexi End Time
endDexi = time.time()
secondsDexi = endDexi - startDexi
print_info('Time taken for DexiNED: {} seconds'.format(secondsDexi))
self.save_egdemaps(edgemap, single_image=True)
print_info('Done testing {}, {}'.format(self.img_info[0], self.img_info[1]))
def run(self, sess):
if not self.init:
return
#This will return the Files Path, Number of Files, Train Indices, Validation Indices using cache_info
train_data = data_parser(self.args)
self.model.setup_training(sess)
if self.args.lr_scheduler is not None:
global_step = tf.Variable(0, trainable=False, dtype=tf.int64)
if self.args.lr_scheduler is None:
learning_rate = tf.constant(self.args.learning_rate,
dtype=tf.float16)
else:
raise NotImplementedError(
'Learning rate scheduler type [%s] is not implemented',
self.args.lr_scheduler)
opt = tf.train.AdamOptimizer(learning_rate)
trainG = opt.minimize(self.model.loss) # like hed
saver = tf.train.Saver(max_to_keep=7)
sess.run(tf.global_variables_initializer())
# here to recovery previous training
if self.args.use_previous_trained:
if self.args.dataset_name.lower(
) != 'biped': # using biped pretrained to use in other dataset
model_path = os.path.join(
'checkpoints',
self.args.model_name + '_' + self.args.train_dataset,
'train')
else:
model_path = os.path.join(
self.args.checkpoint_dir,
self.args.model_name + '_' + self.args.train_dataset)
model_path = os.path.join(model_path, 'train')
if not os.path.exists(model_path) or len(
os.listdir(model_path)) == 0: # :
ini = 0
maxi = self.args.max_iterations + 1
print_warning(
'There is not previous trained data for the current model... and'
)
print_warning(
'*** The training process is starting from scratch ***')
else:
# restoring using the last checkpoint
assert (
len(os.listdir(model_path)) != 0
), 'There is not previous trained data for the current model...'
last_ckpt = tf.train.latest_checkpoint(model_path)
saver.restore(sess, last_ckpt)
ini = self.args.max_iterations
maxi = ini + self.args.max_iterations + 1 # check
print_info(
'--> Previous model restored successfully: {}'.format(
last_ckpt))
else:
print_warning(
'*** The training process is starting from scratch ***')
ini = 0
maxi = ini + self.args.max_iterations
prev_loss = 1000.
prev_val = None
#Checkpoint Directory for CDIBD will be /checkpoints/DXN_CDIBD/train
checkpoint_dir = os.path.join(
self.args.checkpoint_dir,
self.args.model_name + '_' + self.args.train_dataset,
self.args.model_state)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
fig = plt.figure()
for idx in range(ini, maxi):
x_batch, y_batch, _ = get_training_batch(self.args, train_data)
run_metadata = tf.RunMetadata()
_, summary, loss, pred_maps = sess.run([
trainG, self.model.merged_summary, self.model.loss,
self.model.predictions
],
feed_dict={
self.model.images:
x_batch,
self.model.edgemaps:
y_batch
})
if idx % 5 == 0:
self.model.train_writer.add_run_metadata(
run_metadata, 'step{:06}'.format(idx))
self.model.train_writer.add_summary(summary, idx)
print(time.ctime(), '[{}/{}]'.format(idx, maxi),
' TRAINING loss: %.5f' % loss,
'prev_loss: %.5f' % prev_loss)
# saving trained parameters
save_inter = ini + self.args.save_interval
if prev_loss > loss:
saver.save(sess,
os.path.join(checkpoint_dir, self.args.model_name),
global_step=idx)
prev_loss = loss
print("Weights saved in the lowest loss", idx, " Current Loss",
prev_loss)
if idx % self.args.save_interval == 0:
saver.save(sess,
os.path.join(checkpoint_dir, self.args.model_name),
global_step=idx)
prev_loss = loss
print("Weights saved in the interval", idx, " Current Loss",
prev_loss)
# ********* for validation **********
if (idx + 1) % self.args.val_interval == 0:
pause_show = 0.01
imgs_list = []
img = x_batch[2][:, :, 0:3]
gt_mp = y_batch[2]
imgs_list.append(img)
imgs_list.append(gt_mp)
for i in range(len(pred_maps)):
tmp = pred_maps[i][2, ...]
imgs_list.append(tmp)
vis_imgs = visualize_result(imgs_list, self.args)
fig.suptitle("Iterac:" + str(idx + 1) + " Loss:" +
'%.5f' % loss + " training")
fig.add_subplot(1, 1, 1)
plt.imshow(np.uint8(vis_imgs))
print("Evaluation in progress...")
plt.draw()
plt.pause(pause_show)
im, em, _ = get_validation_batch(self.args, train_data)
summary, error, pred_val = sess.run([
self.model.merged_summary, self.model.error,
self.model.fuse_output
],
feed_dict={
self.model.images: im,
self.model.edgemaps: em
})
if error <= 0.08:
saver.save(sess,
os.path.join(checkpoint_dir,
self.args.model_name),
global_step=idx)
prev_loss = loss
print(
"Parameters saved in the validation stage when its error is <=0.08::",
error)
self.model.val_writer.add_summary(summary, idx)
print_info(('[{}/{}]'.format(idx, self.args.max_iterations),
'VALIDATION error: %0.5f' % error,
'pError: %.5f' % prev_loss))
if (idx + 1) % (self.args.val_interval * 150) == 0:
print('updating visualisation')
plt.close()
fig = plt.figure()
saver.save(sess,
os.path.join(checkpoint_dir, self.args.model_name),
global_step=idx)
print("Final Weights saved", idx, " Current Loss", loss)
self.model.train_writer.close()
