add moving average

.gitignore

@@ -1,4 +1,3 @@
-scripts/*.sh
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

eval_seglink.py

@@ -19,24 +19,35 @@
 # model threshold parameters
 # =========================================================================== #
 tf.app.flags.DEFINE_string('train_with_ignored', False, 
-                           'whether to use ignored bbox (in ic15) in training.')
+       'whether to use ignored bbox (in ic15) in training.')
 tf.app.flags.DEFINE_boolean('do_grid_search', False, 
-                           'whether to do grid search to find a best combinations of seg_conf_threshold and link_conf_threshold.')
-tf.app.flags.DEFINE_float('seg_loc_loss_weight', 1.0, 'the loss weight of segment localization')
-tf.app.flags.DEFINE_float('link_cls_loss_weight', 1.0, 'the loss weight of linkage classification loss')
+       'whether to do grid search to find a best combinations of \
+       seg_conf_threshold and link_conf_threshold.')
+tf.app.flags.DEFINE_float('seg_loc_loss_weight', 1.0, 
+      'the loss weight of segment localization')
+tf.app.flags.DEFINE_float('link_cls_loss_weight', 1.0, 
+      'the loss weight of linkage classification loss')
 
 tf.app.flags.DEFINE_float('seg_conf_threshold', 0.9, 
-                          'the threshold on the confidence of segment')
+      'the threshold on the confidence of segment')
 tf.app.flags.DEFINE_float('link_conf_threshold', 0.7, 
-                          'the threshold on the confidence of linkage')
+      'the threshold on the confidence of linkage')
 
 
 # =========================================================================== #
 # Checkpoint and running Flags
 # =========================================================================== #
 tf.app.flags.DEFINE_string('checkpoint_path', None, 
-   'the path of checkpoint to be evaluated. If it is a directory containing many checkpoints, the lastest will be evaluated.')
-tf.app.flags.DEFINE_float('gpu_memory_fraction', 0.1, 'the gpu memory fraction to be used. If less than 0, allow_growth = True is used.')
+   'the path of checkpoint to be evaluated. \
+   If it is a directory containing many checkpoints, \
+   the lastest will be evaluated.')
+tf.app.flags.DEFINE_float('gpu_memory_fraction', 0.1, 
+  'the gpu memory fraction to be used. \
+   If less than 0, allow_growth = True is used.')
+tf.app.flags.DEFINE_bool('using_moving_average', False, 
+   'Whether to use ExponentionalMovingAverage')
+tf.app.flags.DEFINE_float('moving_average_decay', 0.9999, 
+    'The decay rate of ExponentionalMovingAverage')
 
 # =========================================================================== #
 # I/O and preprocessing Flags.
@@ -218,18 +229,30 @@ def eval(dataset):
     elif FLAGS.gpu_memory_fraction > 0:
         sess_config.gpu_options.per_process_gpu_memory_fraction = FLAGS.gpu_memory_fraction;
 
+    # Variables to restore: moving avg. or normal weights.
+    if FLAGS.using_moving_average:
+        variable_averages = tf.train.ExponentialMovingAverage(
+                FLAGS.moving_average_decay)
+        variables_to_restore = variable_averages.variables_to_restore(
+                slim.get_model_variables())
+        variables_to_restore[global_step.op.name] = global_step
+    else:
+        variables_to_restore = slim.get_variables_to_restore()
+
     if util.io.is_dir(FLAGS.checkpoint_path):
         slim.evaluation.evaluation_loop(
             master = '',
             eval_op=list(names_to_updates.values()),
             num_evals=dataset.num_samples,
+            variables_to_restore=variables_to_restore,
             checkpoint_dir = checkpoint_dir,
             logdir = logdir,
             session_config=sess_config)
     else:
         slim.evaluation.evaluate_once(
             master = '',
             eval_op=list(names_to_updates.values()),
+            variables_to_restore=variables_to_restore,
             num_evals=2,#dataset.num_samples,
             checkpoint_path = FLAGS.checkpoint_path,
             logdir = logdir,

push.sh

@@ -1,4 +1,4 @@
 git add . --all
-git commit -m "add train without ignored gt."
+git commit -m "add moving average"
 git push -u origin master
 

scripts/eval.sh

@@ -0,0 +1,45 @@
+set -x
+set -e
+# ./scripts/eval.sh 1 icdar2013 train 384 384 ckpt 
+# ./scripts/eval.sh 1 icdar2013 train 512 512 ckpt 
+
+export CUDA_VISIBLE_DEVICES=$1
+DATASET=$2
+SPLIT=$3
+WIDTH=$4
+HEIGHT=$5
+CHECKPOINT_PATH=$6
+
+if [ $DATASET == 'synthtext' ]
+then
+    DATA_PATH=SynthText
+elif [ $DATASET == 'scut' ]
+then
+    DATA_PATH=SCUT
+elif [ $DATASET == 'icdar2013' ]
+then
+    DATA_PATH=ICDAR
+elif [ $DATASET == 'icdar2015' ]
+then
+    DATA_PATH=ICDAR
+else
+    echo invalid dataset: $DATASET
+    exit
+fi
+
+DATASET_DIR=$HOME/dataset/SSD-tf/${DATA_PATH}
+
+python eval_seglink.py \
+			--checkpoint_path=${CHECKPOINT_PATH} \
+            --dataset_dir=${DATASET_DIR} \
+            --dataset_name=${DATASET} \
+            --dataset_split_name=$SPLIT \
+			--eval_image_width=${WIDTH} \
+			--eval_image_height=${HEIGHT} \
+			--gpu_memory_fraction=0.4 \
+			--do_grid_search=$7 \
+			--using_moving_average=0
+
+
+
+

scripts/test.sh

@@ -0,0 +1,22 @@
+set -x
+set -e
+# ./scripts/test.sh 1 icdar2013 train 384 384 ckpt 
+# ./scripts/test.sh 1 icdar2013 train 512 512 ckpt 
+# ./scripts/test.sh 1 icdar2015 train 1280 768 ckpt 
+
+export CUDA_VISIBLE_DEVICES=$1
+CHECKPOINT_PATH=$2
+
+
+
+python test_seglink.py \
+			--checkpoint_path=${CHECKPOINT_PATH} \
+			--gpu_memory_fraction=-1 \
+			--seg_conf_threshold=0.8 \
+			--link_conf_threshold=0.5
+
+
+
+
+
+

scripts/train.sh

@@ -0,0 +1,57 @@
+set -x
+set -e
+# ./scripts/train.sh 0 18 synthtext
+export CUDA_VISIBLE_DEVICES=$1
+IMG_PER_GPU=$2
+DATASET=$3
+
+CHKPT_PATH=${HOME}/models/seglink/seglink_synthtext
+TRAIN_DIR=${HOME}/models/seglink/seglink_icdar2015_384
+#TRAIN_DIR=${HOME}/temp/no-use/seglink/seglink_icdar2015_384
+#CHKPT_PATH=${HOME}/models/ssd-pretrain/seglink
+
+# get the number of gpus
+OLD_IFS="$IFS" 
+IFS="," 
+gpus=($CUDA_VISIBLE_DEVICES) 
+IFS="$OLD_IFS"
+NUM_GPUS=${#gpus[@]}
+
+# batch_size = num_gpus * IMG_PER_GPU
+BATCH_SIZE=`expr $NUM_GPUS \* $IMG_PER_GPU`
+
+#dataset
+if [ $DATASET == 'synthtext' ]
+then
+    DATA_PATH=SynthText
+elif [ $DATASET == 'scut' ]
+then
+    DATA_PATH=SCUT
+elif [ $DATASET == 'icdar2013' ]
+then
+    DATA_PATH=ICDAR
+elif [ $DATASET == 'icdar2015' ]
+then
+    DATA_PATH=ICDAR
+else
+    echo invalid dataset: $DATASET
+    exit
+fi
+
+DATASET_DIR=$HOME/dataset/SSD-tf/${DATA_PATH}
+
+python train_seglink.py \
+			--train_dir=${TRAIN_DIR} \
+			--num_gpus=${NUM_GPUS} \
+			--learning_rate=0.0001 \
+			--gpu_memory_fraction=-1 \
+			--train_image_width=384 \
+			--train_image_height=384 \
+			--batch_size=${BATCH_SIZE}\
+            --dataset_dir=${DATASET_DIR} \
+            --dataset_name=${DATASET} \
+            --dataset_split_name=train \
+            --train_with_ignored=0 \
+			--checkpoint_path=${CHKPT_PATH} \
+			--using_moving_average=0
+

scripts/vis.sh

@@ -0,0 +1,19 @@
+#python visualize_detection_result.py \
+#	--image=~/dataset/ICDAR2015/Challenge2.Task123/Challenge2_Test_Task12_Images/ \
+#	--gt=~/dataset/ICDAR2015/Challenge2.Task123/Challenge2_Test_Task1_GT/ \
+#	--det=~/temp/no-use/seglink_debug_icdar2013/eval/icdar2013_test/model.ckpt-48176/txt/ \
+#	--output=~/temp/no-use/seglink_result
+
+#python visualize_detection_result.py \
+#	--image=~/dataset/ICDAR2015/Challenge2.Task123/Challenge2_Training_Task12_Images/ \
+#	--gt=~/dataset/ICDAR2015/Challenge2.Task123/Challenge2_Training_Task1_GT/ \
+#	--det=~/temp/no-use/seglink_debug_icdar2013/eval/icdar2013_train/model.ckpt-48176/txt/ \
+#	--output=~/temp/no-use/seglink_result
+
+
+python visualize_detection_result.py \
+	--image=~/dataset/ICDAR2015/Challenge4/ch4_training_images/ \
+	--gt=~/dataset/ICDAR2015/Challenge4/ch4_training_localization_transcription_gt/ \
+	--det=~/models/seglink/seglink_icdar2015_without_ignored/eval/icdar2015_train/model.ckpt-72885/seg_link_conf_th_0.900000_0.700000/txt \
+	--output=~/temp/no-use/seglink_result_512_train
+

train_seglink.py

@@ -44,7 +44,8 @@
 tf.app.flags.DEFINE_float('learning_rate', 0.001, 'learning rate.')
 tf.app.flags.DEFINE_float('momentum', 0.9, 'The momentum for the MomentumOptimizer')
 tf.app.flags.DEFINE_float('weight_decay', 0.0005, 'The weight decay on the model weights.')
-
+tf.app.flags.DEFINE_bool('using_moving_average', False, 'Whether to use ExponentionalMovingAverage')
+tf.app.flags.DEFINE_float('moving_average_decay', 0.9999, 'The decay rate of ExponentionalMovingAverage')
 
 # =========================================================================== #
 # I/O and preprocessing Flags.
@@ -219,7 +220,19 @@ def create_clones(batch_queue):
     averaged_gradients = sum_gradients(gradients)
 
     update_op = optimizer.apply_gradients(averaged_gradients, global_step=global_step)
-    train_op = control_flow_ops.with_dependencies([update_op], seglink_loss, name='train_op')
+
+    train_ops = [update_op]
+
+    # moving average
+    if FLAGS.using_moving_average:
+        tf.logging.info('using moving average in training, \
+        with decay = %f'%(FLAGS.moving_average_decay))
+        ema = tf.train.ExponentialMovingAverage(FLAGS.moving_average_decay)
+        ema_op = ema.apply(tf.trainable_variables())
+        with tf.control_dependencies([update_op]):# ema after updating
+            train_ops.append(tf.group(ema_op))
+
+    train_op = control_flow_ops.with_dependencies(train_ops, seglink_loss, name='train_op')
     return train_op
 
 
@@ -240,7 +253,7 @@ def train(train_op):
             logdir = FLAGS.train_dir,
             init_fn = init_fn,
             summary_op = summary_op,
-            number_of_steps = FLAGS.max_number_of_steps,
+            number_of_steps = 100,#, FLAGS.max_number_of_steps,
             log_every_n_steps = FLAGS.log_every_n_steps,
             save_summaries_secs = 60,
             saver = saver,