Fix 3D bug in display+larcv interface

CHANGES.txt

@@ -1,3 +1,5 @@
 v0.1.1, 3/30/2018 -- Initial release.
 
 v0.2.0, 4/9/2018 -- 3D version of PPN initial release.
+
+v0.2.1, 4/12/2018 -- Correct 3D bugs in display and larcv interface.

faster_particles/config.py

@@ -5,6 +5,10 @@
 
 os.environ['CUDA_DEVICE_ORDER'] = 'PCI_BUS_ID'
 os.environ['CUDA_VISIBLE_DEVICES'] = '0,1,2,3'
+# Control Tensorflow verbose level with TF_CPP_MIN_LOG_LEVEL
+# it defaults to 0 (all logs shown), but can be set to 1 to filter out INFO logs,
+# 2 to additionally filter out WARNING logs, and 3 to additionally filter out ERROR logs
+os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
 
 class PPNConfig(object):
     IMAGE_SIZE = 768 # 512

faster_particles/demo_ppn.py

@@ -76,7 +76,7 @@ def display(blob, cfg, im_proposals=None, rois=None, im_labels=None, im_scores=N
     # Display original image
     if cfg.DATA_3D:
         for voxel in blob['voxels']:
-            draw_voxel(voxel[2], voxel[1], voxel[0], 1, ax, facecolors='blue', alpha=0.3, linewidths=0.1, edgecolors='black')
+            draw_voxel(voxel[0], voxel[1], voxel[2], 1, ax, facecolors='blue', alpha=0.3, linewidths=0.1, edgecolors='black')
     else:
         ax.imshow(blob['data'][0,...,0], cmap='jet', interpolation='none', origin='lower', vmin=0, vmax=400)
 
@@ -137,9 +137,9 @@ def display(blob, cfg, im_proposals=None, rois=None, im_labels=None, im_scores=N
     # Display original image
     if cfg.DATA_3D:
         for voxel in blob['voxels']:
-            draw_voxel(voxel[2], voxel[1], voxel[0], 1, ax2, facecolors='blue', alpha=0.3, linewidths=0.1, edgecolors='black')
+            draw_voxel(voxel[0], voxel[1], voxel[2], 1, ax2, facecolors='blue', alpha=0.3, linewidths=0.1, edgecolors='black')
     else:
-        ax.imshow(blob['data'][0,...,0], cmap='jet', interpolation='none', origin='lower', vmin=0, vmax=400)
+        ax2.imshow(blob['data'][0,...,0], cmap='jet', interpolation='none', origin='lower', vmin=0, vmax=400)
 
     if im_proposals is not None and im_scores is not None:
         if len(im_proposals) > 0:
@@ -170,6 +170,7 @@ def display(blob, cfg, im_proposals=None, rois=None, im_labels=None, im_scores=N
     # Use dpi=1000 for high resolution
     plt.savefig(os.path.join(cfg.DISPLAY_DIR, name + '_predictions_%d.png' % index))
     plt.close(fig2)
+    print(im_proposals)
     return im_proposals
 
     """if cfg.DATA_3D and im_proposals is not None and im_scores is not None:
@@ -282,7 +283,7 @@ def inference(cfg):
     im_proposals, im_labels, im_scores, distances = [], [], [], []
     with tf.Session() as sess:
         saver.restore(sess, cfg.WEIGHTS_FILE)
-        for i in range(100):
+        for i in range(10):
             blob = data.forward()
             summary, results = net.test_image(sess, blob)
             if cfg.NET == 'ppn':

faster_particles/larcvdata/larcvdata_generator.py

@@ -203,13 +203,13 @@ def forward(self):
                     y = t_points[ 3*pt_index + 1 ]
                     z = t_points[ 3*pt_index + 2 ]
                     if x < 0: break
-                    gt_pixels.append([x, y, z, 1])
+                    gt_pixels.append([z, y, x, 1])
                 for pt_index in np.arange(int(len(s_points)/3)):
                     x = s_points[ 3*pt_index     ]
                     y = s_points[ 3*pt_index + 1 ]
                     z = s_points[ 3*pt_index + 2 ]
                     if x < 0: break
-                    gt_pixels.append([x, y, z, 2])
+                    gt_pixels.append([z, y, x, 2])
             else:
                 for pt_index in np.arange(int(len(t_points)/2)):
                     x = t_points[ 2*pt_index     ]

faster_particles/ppn.py

@@ -136,9 +136,7 @@ def create_architecture(self, is_training=True, reuse=None, scope="ppn"):
 
                 # Build PPN1
                 rois = self.build_ppn1(net2)
-                #print("rois", rois.get_shape().as_list())
                 rois = slice_rois(rois, self.dim2)
-                #print("rois", rois.get_shape().as_list())
 
                 if self.is_training:
                     # During training time, check if all ground truth pixels are covered by ROIs
@@ -239,7 +237,6 @@ def build_ppn1(self, net2):
             # Compute softmax
             # Shape of ppn1_cls_prob = 1, 16, 16, 2
             ppn1_cls_prob = tf.nn.softmax(ppn1_cls_score)
-            print("ppn1_cls_score ", ppn1_cls_score.get_shape().as_list())
 
             # Step 3) Get a (meaningful) subset of rois and associated scores
             # Generate anchors = pixel centers of the last feature map.
@@ -251,7 +248,6 @@ def build_ppn1(self, net2):
             # and anchors. Take the first R proposed pixels which contain an object.
             proposals, scores = predicted_pixels(ppn1_cls_prob, ppn1_pixel_pred, anchors, (self.N2,) * self.dim)
             rois, roi_scores = top_R_pixels(proposals, scores, R=20, threshold=self.ppn1_score_threshold)
-            print("final proposals ", proposals.get_shape().as_list())
             assert proposals.get_shape().as_list() == [self.N3**self.dim, self.dim]
             assert scores.get_shape().as_list() == [self.N3**self.dim, 1]
             #assert rois.get_shape().as_list() == [None, 2]
@@ -315,7 +311,6 @@ def build_ppn2(self, rpn_pooling, rois):
             # Step 0) Convolution for PPN2 intermediate layer
             # Based on F3 feature map (ie after 3 max-pool layers in VGG)
             # Shape = nb_rois, 1, 1, 512
-            print("rpn_pooling", rpn_pooling)
             ppn2 = self.conv(rpn_pooling,
                               self.ppn2_channels, # RPN Channels = num_outputs
                               3, # RPN Kernels FIXME change this to (1, 1)?

setup.py

@@ -3,7 +3,7 @@
 
 setup(
     name='faster-particles',
-    version='0.2.0',
+    version='0.2.1',
     author='Laura Domine, Ji Won Park, Kazuhiro Terao',
     author_email='temigo@gmx.com',
     packages=find_packages(),