Properly set reuse in tf.layers to avoid some surprise

examples/basics/cifar-convnet.py

@@ -9,6 +9,7 @@ import os
 from tensorpack import *
 from tensorpack.tfutils.summary import *
 from tensorpack.dataflow import dataset
+from tensorpack.utils.gpu import get_num_gpu
 
 """
 A small convnet model for Cifar10 or Cifar100 dataset.
@@ -145,7 +146,7 @@ if __name__ == '__main__':
         if args.load:
             config.session_init = SaverRestore(args.load)
 
-        nr_gpu = len(args.gpu.split(','))
-        trainer = QueueInputTrainer() if nr_gpu <= 1 \
-            else SyncMultiGPUTrainerParameterServer(nr_gpu)
+        num_gpu = get_num_gpu()
+        trainer = QueueInputTrainer() if num_gpu <= 1 \
+            else SyncMultiGPUTrainerParameterServer(num_gpu)
         launch_train_with_config(config, trainer)

tensorpack/models/batch_norm.py

@@ -110,7 +110,8 @@ def BatchNorm(inputs, axis=None, training=None, momentum=0.9, epsilon=1e-5,
                 beta_initializer=beta_initializer,
                 gamma_initializer=gamma_initializer,
                 virtual_batch_size=virtual_batch_size,
-                fused=True
+                fused=True,
+                _reuse=tf.get_variable_scope().reuse
             )
         else:
             assert virtual_batch_size is None, "Feature not supported in this version of TF!"
@@ -120,7 +121,8 @@ def BatchNorm(inputs, axis=None, training=None, momentum=0.9, epsilon=1e-5,
                 center=center, scale=scale,
                 beta_initializer=beta_initializer,
                 gamma_initializer=gamma_initializer,
-                fused=True
+                fused=True,
+                _reuse=tf.get_variable_scope().reuse
             )
         xn = layer.apply(inputs, training=training, scope=tf.get_variable_scope())
 
@@ -206,7 +208,8 @@ def BatchRenorm(x, rmax, dmax, momentum=0.9, epsilon=1e-5,
             'dmax': dmax},
         renorm_momentum=0.99,
         gamma_initializer=gamma_initializer,
-        fused=False)
+        fused=False,
+        _reuse=tf.get_variable_scope().reuse)
     xn = layer.apply(x, training=ctx.is_training, scope=tf.get_variable_scope())
 
     if ctx.is_main_training_tower:

tensorpack/models/conv2d.py

@@ -63,7 +63,8 @@ def Conv2D(
                 bias_initializer=bias_initializer,
                 kernel_regularizer=kernel_regularizer,
                 bias_regularizer=bias_regularizer,
-                activity_regularizer=activity_regularizer)
+                activity_regularizer=activity_regularizer,
+                _reuse=tf.get_variable_scope().reuse)
             ret = layer.apply(inputs, scope=tf.get_variable_scope())
             ret = tf.identity(ret, name='output')
 
@@ -164,7 +165,8 @@ def Conv2DTranspose(
             bias_initializer=bias_initializer,
             kernel_regularizer=kernel_regularizer,
             bias_regularizer=bias_regularizer,
-            activity_regularizer=activity_regularizer)
+            activity_regularizer=activity_regularizer,
+            _reuse=tf.get_variable_scope().reuse)
         ret = layer.apply(inputs, scope=tf.get_variable_scope())
         ret = tf.identity(ret, name='output')
 

tensorpack/models/fc.py

@@ -46,7 +46,8 @@ def FullyConnected(
             bias_initializer=bias_initializer,
             kernel_regularizer=kernel_regularizer,
             bias_regularizer=bias_regularizer,
-            activity_regularizer=activity_regularizer)
+            activity_regularizer=activity_regularizer,
+            _reuse=tf.get_variable_scope().reuse)
         ret = layer.apply(inputs, scope=tf.get_variable_scope())
         ret = tf.identity(ret, name='output')