fix generate_sub_average

fix generate_sub_final_ensemble
update asanakoy bash scripts
fix main bash scritps

asanakoy/generate_sub_average.py

@@ -42,7 +42,6 @@ def average_from_files(test_image_paths, probs_dirs, output_dir, should_save_mas
             raise ValueError('{} not found'.format(dir_path))
     if should_save_masks:
         output_dir.mkdir(exist_ok=True)
-    print 'Number of samples:', len(test_image_paths)
 
     all_rles = []
     all_img_filenames = []
@@ -93,6 +92,7 @@ def main():
                         help='load pregenerated probs from folder?')
     parser.add_argument('--net_name', choices=['scratch', 'vgg11v1'])
     args = parser.parse_args()
+    print 'config.submissions_dir', config.submissions_dir
 
     if args.net_name == 'vgg11v1':
         probs_dirs = list()
@@ -118,6 +118,7 @@ def main():
     else:
         raise ValueError('Unknown net_name {}'.format(args.net_name))
 
+    probs_dirs = map(Path, probs_dirs)
     with open(str(output_dir) + '.txt', mode='w') as f:
         f.write('Following models were averaged:\n')
         for l in probs_dirs:

asanakoy/predict_scratch.sh

@@ -1,6 +1,9 @@
 #!/usr/bin/env bash
 set -e # abort if any command fails
 
+echo "PREDICT SCRATCH"
+echo "---"
+
 models_dir=$(python -c "import json, os; print os.path.expanduser(json.load(open('../config/config.json', 'r'))['models_dir'])")
 
 BATCH=2 # batch size
@@ -18,4 +21,4 @@ for ((FOLD=0; FOLD<7; FOLD++)); do
     python generate_sub.py -b=$BATCH -o "${o_dir}" --no_cudnn
 done
 
-python generate_sub_average.py --net_name=scratch -j=32
+python generate_sub_average.py --net_name=scratch -j=4

asanakoy/predict_vgg11v1.sh

@@ -1,6 +1,9 @@
 #!/usr/bin/env bash
 set -e # abort if any command fails
 
+echo "PREDICT VGG11v1"
+echo "---"
+
 models_dir=$(python -c "import json, os; print os.path.expanduser(json.load(open('../config/config.json', 'r'))['models_dir'])")
 
 BATCH=2 # batch size
@@ -17,4 +20,4 @@ for ((FOLD=0; FOLD<7; FOLD++)); do
     python generate_sub.py -b=$BATCH -o "${o_dir}" -net=vgg11v1 --no_cudnn
 done
 
-python generate_sub_average.py --net_name=vgg11v1 -j=32
+python generate_sub_average.py --net_name=vgg11v1 -j=4

asanakoy/train_scratch.sh

@@ -1,18 +1,16 @@
 #!/usr/bin/env bash
 set -e # abort if any command fails
 
+echo "TRAIN SCRATCH"
+echo "---"
+
 epochs=250
-should_gen_sub=1
 models_dir=$(python -c "import json, os; print os.path.expanduser(json.load(open('../config/config.json', 'r'))['models_dir'])")
 
-usage() { echo "Usage: $0 -f FOLD -b BATCH_SIZE -g GACC -e EPOCHS -n\n -n: to skip submission generation" 1>&2; exit 1; }
+usage() { echo "Usage: $0 -b BATCH_SIZE -g GACC -e EPOCHS" 1>&2; exit 1; }
 
 while getopts ":f:b:g:e:n" o; do
     case "${o}" in
-        f)
-            i=${OPTARG}
-            ((s >= 0 || s <= 6)) || usage
-            ;;
         b)
             BATCH=${OPTARG}
             ;;
@@ -22,33 +20,31 @@ while getopts ":f:b:g:e:n" o; do
         e)
             epochs=${OPTARG}
             ;;
-        n)
-            should_gen_sub=0
-            ;;
         *)
             echo "Unknown argument ${o}"
             usage
             ;;
     esac
 done
 shift $((OPTIND-1))
-if [ -z "${i}" ] || [ -z "${BATCH}" ] || [ -z "${gacc}" ] ; then
+if [ -z "${BATCH}" ] || [ -z "${gacc}" ] ; then
     echo "Not all reauired arguments were specified"
     usage
 fi
 
+((BATCH == 1 || BATCH == 2 || BATCH == 4)) ||
+{ echo "Wrong batch size! Only size 1, 2 or 4 is possible"; exit 1; }
 ((BATCH*gacc == 4)) ||
 { echo "Wrong batch size and GACC! BATCH * GACC must be equal 4! "; exit 1; }
 
 
-for ((FOLD=i; FOLD<i+1; FOLD++)); do
+for ((FOLD=0; FOLD<7; FOLD++)); do
 
     echo "=========="
     echo "FOLD $FOLD"
     echo "BATCH $BATCH"
     echo "gacc $gacc"
     echo "epochs $epochs"
-    echo "should_gen_sub $should_gen_sub"
     echo "=========="
     echo ""
 
@@ -58,7 +54,4 @@ for ((FOLD=i; FOLD<i+1; FOLD++)); do
         --lr=0.005 -opt=sgd --decay_step=100 --decay_gamma=0.5 \
         -aug=2 --weight_decay=0.0005 \
         -o="${o_dir}" --epochs=$epochs --no_cudnn
-
-    ((should_gen_sub)) && python generate_sub.py -b=$BATCH \
-        -o "${o_dir}" --no_cudnn
 done

asanakoy/train_vgg11v1.sh

@@ -1,15 +1,41 @@
 #!/usr/bin/env bash
 set -e # abort if any command fails
 
-BATCH=2
-gacc=$((4/BATCH))
+echo "TRAIN VGG11v1"
+echo "---"
+
 models_dir=$(python -c "import json, os; print os.path.expanduser(json.load(open('../config/config.json', 'r'))['models_dir'])")
 
+usage() { echo "Usage: $0 -b BATCH_SIZE -g GACC"; exit 1; }
+
+while getopts ":f:b:g:e:n" o; do
+    case "${o}" in
+        b)
+            BATCH=${OPTARG}
+            ;;
+        g)
+            gacc=${OPTARG}
+            ;;
+        *)
+            echo "Unknown argument ${o}"
+            usage
+            ;;
+    esac
+done
+shift $((OPTIND-1))
+if [ -z "${BATCH}" ] || [ -z "${gacc}" ] ; then
+    echo "Not all reauired arguments were specified"
+    usage
+fi
+
 ((BATCH == 1 || BATCH == 2 || BATCH == 4)) ||
 { echo "Wrong batch size! Only size 1, 2 or 4 is possible"; exit 1; }
+((BATCH*gacc == 4)) ||
+{ echo "Wrong batch size and GACC! BATCH * GACC must be equal 4! "; exit 1; }
+
 
 # Calculate train car areas. To split on folds according to area
-python calc_car_areas.py -j=2
+python calc_car_area.py -j=2
 
 for ((FOLD=0; FOLD<7; FOLD++)); do
     echo "=========="
@@ -25,6 +51,8 @@ for ((FOLD=0; FOLD<7; FOLD++)); do
         --lr=0.0001 -opt=adam --cyclic_lr=20 -aug=1  --no_cudnn --weight_decay=0 \
     -o "$o_dir" --epochs=15
 
+    echo "Train after 15th epoch"
+    echo "-"
     python run_train.py -net=vgg11v1 -imsize=1024 -b=$BATCH -gacc=$gacc -f=$FOLD -nf=7 -fv=2 \
         --lr=0.0001 -opt=adam --cyclic_lr=20 -aug=0  --no_cudnn --weight_decay=0 \
         -o "$o_dir" --epochs=60

generate_sub_final_ensemble.py

@@ -11,9 +11,9 @@
 from sklearn.utils import gen_even_slices
 
 import config
-from data_utils import rle_encode
-from data_utils import rle_to_string
-from dataset import CARVANA
+from asanakoy.data_utils import rle_encode
+from asanakoy.data_utils import rle_to_string
+from asanakoy.dataset import CARVANA
 
 
 def load_from_files(test_image_paths, output_dir=None, is_quiet=False):
@@ -41,7 +41,6 @@ def average_from_files(test_image_paths, probs_dirs, output_dir, should_save_mas
         if not dir_path.exists():
             raise ValueError('{} not found'.format(dir_path))
     output_dir.mkdir(exist_ok=True)
-    print 'Number of samples:', len(test_image_paths)
 
     all_rles = []
     all_img_filenames = []
@@ -101,7 +100,7 @@ def main():
 
     probs_dirs = [
         ('test_scratch2', 1.0),
-        ('test_vgg11v1_7folds_final', 1.0),
+        ('test_vgg11v1_final', 1.0),
         ('albu/albu_sep27', 1.0),
         ('ternaus/ternaus_sep27', 1.0),
     ]

predict.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 # This script will run predict on test for all the networks and assemble a final submission file.
-
+set -e # abort if any command fails
 export PYTHONPATH=$(pwd):$PYTHONPATH
 
 pushd asanakoy
@@ -10,6 +10,7 @@ popd
 
 # TODO: albu, ternaus
 
+echo "Generate final ensemble"
 python generate_sub_final_ensemble.py -j=4
 # TODO: fix van in final prediction using albu's script
 # or include this fix in generate_sub_final_ensemble.py

train.sh

@@ -1,8 +1,11 @@
 #!/usr/bin/env bash
 # This script will run training for all the networks.
+set -e # abort if any command fails
+export PYTHONPATH=$(pwd):$PYTHONPATH
 
-
-bash asanakoy/train_scratch.sh
-bash asanakoy/train_vgg11v1.sh
+pushd asanakoy
+bash train_scratch.sh -b1 -g4
+bash train_vgg11v1.sh -b1 -g4
+popd
 
 # TODO: albu, ternaus

train_and_predict.sh

@@ -1,5 +1,5 @@
 #!/usr/bin/env bash
 # This script will run all trainings, run all predicts and assemble a final submission file.
 
-bash train.sh
-bash predict.sh
+bash train.sh || exit 1
+bash predict.sh || exit 1