Fix errors due to API change in Tensorflow 2.20 (#691)

Continued from copilot PR #690 

- Fix SharedMLP.call in randlanet.py - make training keyword-only
- Fix LocalSpatialEncoding.call in randlanet.py - make training and
relative_features keyword-only
- Fix AttentivePooling.call in randlanet.py - make training keyword-only
- Fix LocalFeatureAggregation.call in randlanet.py - make training
keyword-only
- Fix BatchNormBlock.call call sites in network_blocks.py - pass
training as keyword
- Fix UnaryBlock.call in network_blocks.py - make batch and training
keyword-only
- Fix SimpleBlock.call in network_blocks.py - make batch keyword-only;
fix internal training calls
- Fix ResnetBottleneckBlock.call in network_blocks.py - make batch
keyword-only; fix all internal training/batch calls
- Fix NearestUpsampleBlock.call in network_blocks.py - make batch
keyword-only
- Fix KPFCNN.call in kpconv.py - pass batch as keyword in all block
calls; fix head_mlp/head_softmax calls
- Fix PointPillarsScatter.call in point_pillars.py - make batch_size
keyword-only; fix call site in extract_feats
- Fix tf_batch_neighbors in kpconv.py - pass radius, points_row_splits,
queries_row_splits as keywords to FixedRadiusSearch
- Fix SECONDFPN in point_pillars.py - add Permute layers to convert
NCHW↔NHWC around Conv2DTranspose/Conv2D to avoid channels_first
unsupported on CPU

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: Benjamin Ummenhofer <benjamin.ummenhofer@intel.com>

Author: 3 people

ml3d/tf/models/kpconv.py

@@ -293,15 +293,15 @@ def call(self, flat_inputs, training=False):
         for block_i, block_op in enumerate(self.encoder_blocks):
             if block_i in self.encoder_skips:
                 skip_conn.append(x)
-            x = block_op(x, inputs, training=training)
+            x = block_op(x, batch=inputs, training=training)
 
         for block_i, block_op in enumerate(self.decoder_blocks):
             if block_i in self.decoder_concats:
                 x = tf.concat([x, skip_conn.pop()], axis=1)
-            x = block_op(x, inputs, training=training)
+            x = block_op(x, batch=inputs, training=training)
 
-        x = self.head_mlp(x, inputs, training)
-        x = self.head_softmax(x, inputs, training)
+        x = self.head_mlp(x, batch=inputs, training=training)
+        x = self.head_softmax(x, batch=inputs, training=training)
 
         return x
 
@@ -1049,7 +1049,11 @@ def tf_batch_neighbors(queries, points, q_batches, p_batches, r):
         tf.concat([tf.constant([0]), tf.cumsum(q_batches)], axis=0), tf.int64)
     p_splits = tf.cast(
         tf.concat([tf.constant([0]), tf.cumsum(p_batches)], axis=0), tf.int64)
-    result = nns(points, queries, r, p_splits, q_splits)
+    result = nns(points,
+                 queries,
+                 radius=r,
+                 points_row_splits=p_splits,
+                 queries_row_splits=q_splits)
 
     idx = result.neighbors_index
     splits = result.neighbors_row_splits

ml3d/tf/models/network_blocks.py

@@ -222,8 +222,8 @@ def init_KP(self):
 
     def regular_loss(self):
 
-        fitting_loss = 0
-        repulsive_loss = 0
+        fitting_loss = tf.constant(0.0)
+        repulsive_loss = tf.constant(0.0)
 
         if self.deformable:
             KP_min_d2 = self.min_d2 / (self.KP_extent**2)
@@ -432,7 +432,7 @@ def __init__(self, in_dim, use_bn, bn_momentum):
 
     def call(self, x, training=False):
         if (self.use_bn):
-            return self.batch_norm(x, training)
+            return self.batch_norm(x, training=training)
         else:
             return x + self.bias
 
@@ -470,9 +470,9 @@ def __init__(self,
         if not no_relu:
             self.leaky_relu = tf.keras.layers.LeakyReLU(l_relu)
 
-    def call(self, x, batch=None, training=False):
+    def call(self, x, *, batch=None, training=False):
         x = self.mlp(x)
-        x = self.batch_norm(x, training)
+        x = self.batch_norm(x, training=training)
         if not self.no_relu:
             x = self.leaky_relu(x)
         return x
@@ -517,7 +517,7 @@ def __init__(self, block_name, in_dim, out_dim, radius, layer_ind, cfg):
                                          self.bn_momentum)
         self.leaky_relu = tf.keras.layers.LeakyReLU(cfg.get('l_relu', 0.2))
 
-    def call(self, x, batch, training=False):
+    def call(self, x, *, batch, training=False):
 
         # TODO : check x, batch
         if 'strided' in self.block_name:
@@ -530,7 +530,7 @@ def call(self, x, batch, training=False):
             neighb_inds = batch['neighbors'][self.layer_ind]
 
         x = self.KPConv(q_pts, s_pts, neighb_inds, x)
-        x = self.batch_norm(x, training)
+        x = self.batch_norm(x, training=training)
         x = self.leaky_relu(x)
         return x
 
@@ -621,7 +621,7 @@ def __init__(self, block_name, in_dim, out_dim, radius, layer_ind, cfg):
 
         return
 
-    def call(self, features, batch, training=False):
+    def call(self, features, *, batch, training=False):
 
         if 'strided' in self.block_name:
             q_pts = batch['points'][self.layer_ind + 1]
@@ -633,21 +633,21 @@ def call(self, features, batch, training=False):
             neighb_inds = batch['neighbors'][self.layer_ind]
 
         # First downscaling mlp
-        x = self.unary1(features, training)
+        x = self.unary1(features, training=training)
 
         # Convolution
         x = self.KPConv(q_pts, s_pts, neighb_inds, x)
-        x = self.leaky_relu(self.batch_norm_conv(x, training))
+        x = self.leaky_relu(self.batch_norm_conv(x, training=training))
 
         # Second upscaling mlp
-        x = self.unary2(x, training)
+        x = self.unary2(x, training=training)
 
         # Shortcut
         if 'strided' in self.block_name:
             shortcut = max_pool(features, neighb_inds)  # TODO : test max_pool
         else:
             shortcut = features
-        shortcut = self.unary_shortcut(shortcut, training)
+        shortcut = self.unary_shortcut(shortcut, training=training)
 
         return self.leaky_relu(x + shortcut)
 
@@ -664,7 +664,7 @@ def __init__(self, layer_ind):
         self.layer_ind = layer_ind
         return
 
-    def call(self, x, batch):
+    def call(self, x, *, batch):
         return closest_pool(x, batch['upsamples'][self.layer_ind - 1])
 
     def __repr__(self):

ml3d/tf/models/point_pillars.py

@@ -96,11 +96,11 @@ def extract_feats(self, points, training=False):
                                             coors,
                                             training=training)
 
-        batch_size = int(coors[-1, 0].numpy()) + 1
+        batch_size = tf.cast(coors[-1, 0], tf.int32) + 1
 
         x = self.middle_encoder(voxel_features,
                                 coors,
-                                batch_size,
+                                batch_size=batch_size,
                                 training=training)
         x = self.backbone(x, training=training)
         x = self.neck(x, training=training)
@@ -693,19 +693,19 @@ def __init__(self, in_channels=64, output_shape=[496, 432]):
         self.fp16_enabled = False
 
     #@auto_fp16(apply_to=('voxel_features', ))
-    def call(self, voxel_features, coors, batch_size, training=False):
+    def call(self, voxel_features, coors, *, batch_size, training=False):
         """Scatter features of single sample.
 
         Args:
             voxel_features (tf.Tensor): Voxel features in shape (N, M, C).
             coors (tf.Tensor): Coordinates of each voxel in shape (N, 4).
                 The first column indicates the sample ID.
-            batch_size (int): Number of samples in the current batch.
+            batch_size (int or tf.Tensor): Number of samples in the current batch.
             training (bool): Whether we are training or not?
         """
         # batch_canvas will be the final output.
         batch_canvas = []
-        for batch_itt in range(batch_size):
+        for batch_itt in range(int(batch_size)):
             # Create the canvas for this sample
             canvas_shape = (self.nx * self.ny, self.in_channels)
 
@@ -851,20 +851,22 @@ def __init__(self,
                     kernel_size=upsample_strides[i],
                     strides=upsample_strides[i],
                     use_bias=False,
-                    data_format='channels_first',
+                    data_format='channels_last',
                 )
             else:
                 stride = np.round(1 / stride).astype(np.int64)
-                upsample_layer = tf.keras.layers.Conv2D(  # TODO : convert to channels last.
+                upsample_layer = tf.keras.layers.Conv2D(
                     filters=out_channels[i],
                     kernel_size=stride,
-                    data_format='channels_first',
+                    data_format='channels_last',
                     use_bias=False,
                     strides=stride,
                     kernel_initializer='he_normal')
 
             deblock = tf.keras.Sequential()
+            deblock.add(tf.keras.layers.Permute((2, 3, 1)))  # NCHW -> NHWC
             deblock.add(upsample_layer)
+            deblock.add(tf.keras.layers.Permute((3, 1, 2)))  # NHWC -> NCHW
             deblock.add(
                 tf.keras.layers.BatchNormalization(axis=1,
                                                    epsilon=1e-3,

ml3d/tf/models/randlanet.py

@@ -536,7 +536,7 @@ def __init__(self,
             axis=-1, momentum=0.99, epsilon=1e-6) if bn else None
         self.activation_fn = activation_fn
 
-    def call(self, input, training):
+    def call(self, input, *, training=None):
         """Forward pass of the Module.
 
         Args:
@@ -593,7 +593,8 @@ def call(self,
              coords,
              features,
              neighbor_indices,
-             training,
+             *,
+             training=None,
              relative_features=None):
         """Forward pass of the Module.
 
@@ -654,7 +655,7 @@ def __init__(self, in_channels, out_channels):
                              out_channels,
                              activation_fn=tf.keras.layers.LeakyReLU(0.2))
 
-    def call(self, x, training):
+    def call(self, x, *, training=None):
         """Forward pass of the Module.
 
         Args:
@@ -700,7 +701,7 @@ def __init__(self, d_in, d_out, num_neighbors):
         self.shortcut = SharedMLP(d_in, 2 * d_out)
         self.lrelu = tf.keras.layers.LeakyReLU(0.2)
 
-    def call(self, coords, feat, neighbor_indices, training):
+    def call(self, coords, feat, neighbor_indices, *, training=None):
         """Forward pass of the Module.
 
         Args:

ml3d/tf/utils/tf_utils.py

@@ -7,7 +7,7 @@ def atoi(text):
 
 
 def natural_keys(text):
-    return [atoi(c) for c in re.split('(\d+)', text)]
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
 
 
 def gen_CNN(channels,

ml3d/torch/utils/torch_utils.py

@@ -9,7 +9,7 @@ def atoi(text):
 
 
 def natural_keys(text):
-    return [atoi(c) for c in re.split('(\d+)', text)]
+    return [atoi(c) for c in re.split(r'(\d+)', text)]
 
 
 def latest_torch_ckpt(train_ckpt_dir):