Remove unused parameters and code. Clean up point cloud functions.

dpc/resources/default_config.yaml

@@ -4,15 +4,13 @@
 config: "" # .yaml config file
 
 inp_dir: ""  # Directory path containing the input data (tfrecords).
-dataset_name: "shapenet_single_category"  # Dataset name that is to be used for training and evaluation.
 synth_set: "03001627"  # Class of Shapenet data
 num_views: 5  # Number of viewpoints in the input data.
 num_views_to_use: -1  # Num of views to actually use.
 tfrecords_gzip_compressed: true  # Compress TF record files.
 image_size: 128  # Input images dimension (pixels) - width & height.
 saved_camera: true
 saved_depth: false
-saved_voxels: false
 
 # Model general
 encoder_name: "img_encoder"  # Name of the encoder network being used.
@@ -40,7 +38,6 @@ pose_predict_num_candidates: 1
 pose_candidates_num_layers: 3
 pose_predictor_student: true
 pose_predictor_student_loss_weight: 1.0
-pose_student_spherical_loss: true
 pose_student_align_loss: false
 
 predict_translation: false
@@ -52,8 +49,7 @@ predict_translation_init_stddev: 0.05
 pc_relative_sigma: 1.0
 pc_relative_sigma_end: 0.2  # specify -1.0 for constant sigma
 pc_normalise_gauss: false
-pc_normalise_gauss_analytical: false
-pc_fix_bug_sigma: false
+pc_normalise_gauss_analytical: true
 pc_unit_cube: true
 pc_fast: true
 pc_gauss_kernel_size: 11
@@ -64,8 +60,6 @@ pc_occupancy_scaling_maximum: 1.0
 # RGB
 pc_rgb: false
 pc_rgb_stop_points_gradient: false
-pc_clip_after_conv: false
-pc_do_not_clip: false
 pc_rgb_clip_after_conv: false
 pc_rgb_divide_by_occupancies: false
 pc_rgb_divide_by_occupancies_epsilon: 0.01
@@ -129,16 +123,9 @@ max_number_of_steps: 600000  # Maximum number of steps for training.
 compute_validation_loss: true
 variable_num_views: false
 
-save_snapshot_interval: -1
 validation_interval: 1000  # Do validation every so often.
 save_intermediate_pcs: false
 save_intermediate_pcs_interval: 5000
-save_early_frequent_snapshots: false
-
-master: ""  # The address of the tensorflow master
-
-save_summaries_secs: 120  # Seconds interval for dumping TF summaries.
-save_interval_secs: 300  # Seconds interval to save models.
 
 # Predict and visualise
 num_dataset_samples: -1

dpc/util/point_cloud.py

@@ -36,17 +36,13 @@ def pointcloud2voxels(cfg, input_pc, sigma):  # [B,N,3]
     sq_distance = tf.square(x_big - xg) + tf.square(y_big - yg) + tf.square(z_big - zg)
 
     # compute gaussian
-    if cfg.pc_fix_bug_sigma:
-        func = tf.exp(-sq_distance / (2.0 * sigma * sigma))  # [B,N,G,G,G]
-    else:
-        func = tf.exp(-sq_distance/(sigma*sigma))
+    func = tf.exp(-sq_distance / (2.0 * sigma * sigma))  # [B,N,G,G,G]
 
     # normalise gaussian
     if cfg.pc_normalise_gauss:
         normaliser = tf.reduce_sum(func, [2, 3, 4], keep_dims=True)
         func /= normaliser
     elif cfg.pc_normalise_gauss_analytical:
-        assert(cfg.pc_fix_bug_sigma)
         # should work with any grid sizes
         magic_factor = 1.78984352254  # see estimate_gauss_normaliser
         sigma_normalised = sigma * vox_size
@@ -67,7 +63,6 @@ def pointcloud2voxels3d_fast(cfg, pc, rgb):  # [B,N,3]
     else:
         vox_size_z = vox_size
 
-    filter_outsiders = True
     batch_size = pc.shape[0]
     num_points = tf.shape(pc)[1]
 
@@ -76,6 +71,7 @@ def pointcloud2voxels3d_fast(cfg, pc, rgb):  # [B,N,3]
     grid_size = 1.0
     half_size = grid_size / 2
 
+    filter_outliers = True
     valid = tf.logical_and(pc >= -half_size, pc <= half_size)
     valid = tf.reduce_all(valid, axis=-1)
 
@@ -94,14 +90,14 @@ def pointcloud2voxels3d_fast(cfg, pc, rgb):  # [B,N,3]
     r = pc_grid - indices_floor  # fractional part
     rr = [1.0 - r, r]
 
-    if filter_outsiders:
+    if filter_outliers:
         valid = tf.reshape(valid, [-1])
         indices = tf.boolean_mask(indices, valid)
 
     def interpolate_scatter3d(pos):
         updates_raw = rr[pos[0]][:, :, 0] * rr[pos[1]][:, :, 1] * rr[pos[2]][:, :, 2]
         updates = tf.reshape(updates_raw, [-1])
-        if filter_outsiders:
+        if filter_outliers:
             updates = tf.boolean_mask(updates, valid)
 
         indices_loc = indices
@@ -116,7 +112,7 @@ def interpolate_scatter3d(pos):
                 updates_raw = tf.stop_gradient(updates_raw)
             updates_rgb = tf.expand_dims(updates_raw, axis=-1) * rgb
             updates_rgb = tf.reshape(updates_rgb, [-1, 3])
-            if filter_outsiders:
+            if filter_outliers:
                 updates_rgb = tf.boolean_mask(updates_rgb, valid)
             voxels_rgb = tf.scatter_nd(indices_loc, updates_rgb, [batch_size, vox_size_z, vox_size, vox_size, 3])
         else:
@@ -134,13 +130,7 @@ def interpolate_scatter3d(pos):
                 voxels_rgb.append(vx_rgb)
 
     voxels = tf.add_n(voxels)
-
-    if has_rgb:
-        voxels_rgb = tf.add_n(voxels_rgb)
-        if not cfg.pc_rgb_clip_after_conv:
-            voxels_rgb = tf.clip_by_value(voxels_rgb, 0.0, 1.0)
-    else:
-        voxels_rgb = None
+    voxels_rgb = tf.add_n(voxels_rgb) if has_rgb else None
 
     return voxels, voxels_rgb
 
@@ -151,9 +141,6 @@ def smoothen_voxels3d(cfg, voxels, kernel):
             voxels = tf.nn.conv3d(voxels, krnl, [1, 1, 1, 1, 1], padding="SAME")
     else:
         voxels = tf.nn.conv3d(voxels, kernel, [1, 1, 1, 1, 1], padding="SAME")
-    if cfg.pc_clip_after_conv and not cfg.pc_do_not_clip:
-        print("clipping occupancies after convolution")
-        voxels = tf.clip_by_value(voxels, 0.0, 1.0)
     return voxels
 
 
@@ -249,12 +236,13 @@ def pointcloud_project_fast(cfg, point_cloud, transform, predicted_translation,
     voxels = tf.expand_dims(voxels, axis=-1)
     voxels_raw = voxels
 
-    if not cfg.pc_clip_after_conv and not cfg.pc_do_not_clip:
-        voxels = tf.clip_by_value(voxels, 0.0, 1.0)
+    voxels = tf.clip_by_value(voxels, 0.0, 1.0)
 
     if kernel is not None:
         voxels = smoothen_voxels3d(cfg, voxels, kernel)
         if has_rgb:
+            if not cfg.pc_rgb_clip_after_conv:
+                voxels_rgb = tf.clip_by_value(voxels_rgb, 0.0, 1.0)
             voxels_rgb = convolve_rgb(cfg, voxels_rgb, kernel)
 
     if scaling_factor is not None:

dpc/util/train.py

@@ -1,29 +1,7 @@
 import tensorflow as tf
 
 
-def get_train_op_for_scope(cfg, loss, optimizer, scopes):
-    """Train operation function for the given scope used file training."""
-    is_trainable = lambda x: x in tf.trainable_variables()
-
-    var_list = []
-    update_ops = []
-
-    for scope in scopes:
-        var_list_raw = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=scope)
-        var_list_scope = list(filter(is_trainable, var_list_raw))
-        var_list.extend(var_list_scope)
-        update_ops.extend(tf.get_collection(tf.GraphKeys.UPDATE_OPS, scope))
-
-    return tf.contrib.slim.learning.create_train_op(
-        loss,
-        optimizer,
-        update_ops=update_ops,
-        variables_to_train=var_list,
-        clip_gradient_norm=cfg.clip_gradient_norm)
-
-
 def get_trainable_variables(scopes):
-    """Train operation function for the given scope used file training."""
     is_trainable = lambda x: x in tf.trainable_variables()
 
     var_list = []