Merge pull request #17 from RohanBhandari/master

Edit to allow functionality with Octave

model_zalando_mask_content_test.py

@@ -119,7 +119,7 @@ def main(unused_argv):
     pass
 
   # batch inference, can also be done one image per time.
-  batch_size = 16
+  batch_size = 1
   image_holder = tf.placeholder(tf.float32, shape=[batch_size, 256, 192, 3])
   prod_image_holder = tf.placeholder(
       tf.float32, shape=[batch_size, 256, 192, 3])

model_zalando_refine_test.py

@@ -123,7 +123,7 @@ def main(unused_argv):
   except:
     pass
 
-  batch_size = 4
+  batch_size = 1
 
   # Feed into the refine module
   image_holder = tf.placeholder(tf.float32, shape=[batch_size,256,192,3])

prepare_data/build_viton.py

@@ -142,13 +142,7 @@ def _extract_pose_map(pose_keypoints, h, w):
 
 def _process_segment_map(segment, h, w):
   """Extract segment maps."""
-  # Segment is a reversed 641x641 heat map (20 values). 
   segment = np.asarray(segment, dtype=np.uint8)
-  segment = np.transpose(segment)
-  if h > w:
-    segment = segment[:, :int(641.0 * float(w) / float(h))]
-  else:
-    segment = segment[:int(641.0 * float(h) / float(w)), :]
   segment = imresize(segment, (h, w), interp='nearest')
   return segment.tostring()
 

shape_context_warp.m

@@ -3,19 +3,27 @@
 % Code for shape context matching (under shape_context folder) are modified from:
 % https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/shape/sc_digits.html
 
-
 addpath('shape_context')
 DATA_ROOT='data/women_top/';
 MASK_DIR='results/stage1/tps/00015000_';
 
+% Check if using MATLAB or Octave
+isOctave = exist('OCTAVE_VERSION', 'builtin') ~= 0;
+if(isOctave)
+  % Load image package for resizing images
+  pkg load image;
+  % Turn off warning
+  warning('off', 'Octave:possible-matlab-short-circuit-operator');
+end
+
 [image1, image2] = textread('data/viton_test_pairs.txt', '%s %s\n');
 % using a smaller height and width for the shape context matching
 % can save time without hurting the perform too much.
 h = 256/4;
 w = 192/4;
 % we use 10x10 control_points
 n_control = 10;
-for i = 1:50 %length(image1) % only run for first 10 images
+for i = 1:1 %length(image1) % only run over 1 image (for now)
     image_name1 = image1{i};
     image_name2 = image2{i};
     if exist([MASK_DIR, image_name1, '_', image_name2, '_tps.mat'])
@@ -54,5 +62,5 @@
     control_points = [x(point_index), y(point_index)]';
     control_points = [control_points(1,:) / w; control_points(2,:) / h];
     control_points = reshape(control_points, [2,n_control,n_control]);
-    save([MASK_DIR, image_name1, '_', image_name2, '_tps.mat'], 'control_points');
-end
+    save([MASK_DIR, image_name1, '_', image_name2, '_tps.mat'], 'control_points', '-v6');
+end

test_stage1.sh

@@ -4,4 +4,4 @@ python model_zalando_mask_content_test.py \
   --mode test \
   --result_dir results/stage1/ \
   --begin 0 \
-  --end 50
+  --end 1 #exclusive

test_stage2.sh

@@ -5,4 +5,4 @@ python model_zalando_refine_test.py \
   --mode test \
   --result_dir results/stage2/ \
   --begin 0 \
-  --end 50
+  --end 1 #exclusive

utils.py

@@ -84,13 +84,7 @@ def extract_segmentation(segment):
 
 def process_segment_map(segment, h, w):
   """Extract segment maps."""
-  # Segment is a reversed 641x641 heat map (20 values).
   segment = np.asarray(segment, dtype=np.uint8)
-  segment = np.transpose(segment)
-  if h > w:
-    segment = segment[:, :int(641.0 * float(w) / float(h))]
-  else:
-    segment = segment[:int(641.0 * float(h) / float(w)), :]
   segment = imresize(segment, (h, w), interp='nearest')
   return segment
 
@@ -107,9 +101,6 @@ def extract_pose_representation(pose_keypoints, h, w, resize_h, resize_w):
   pose_keypoints = tf.cast(pose_keypoints, tf.int64)
   # pose_representation = tf.zeros((h, w, 18))
 
-  # Fucking hard assignment!
-  # tf.nn.dilation2d
-
   pose_representation = tf.one_hot(pose_keypoints, 3)
   return pose_representation