use sparse softmax

tensorpack · Apr 14, 2016 · 9a4e6d9 · 9a4e6d9
1 parent d646972
commit 9a4e6d9
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 11 deletions.
diff --git a/examples/cifar10_convnet.py b/examples/cifar10_convnet.py
@@ -19,7 +19,8 @@
 from tensorpack.dataflow import imgaug
 
 """
-CIFAR10 90% validation accuracy after 40k step.
+A small cifar10 convnet model.
+90% validation accuracy after 40k step.
 """
 
 BATCH_SIZE = 128
@@ -62,8 +63,7 @@ def _get_cost(self, input_vars, is_training):
         # fc will have activation summary by default. disable for the output layer
         logits = FullyConnected('linear', l, out_dim=10, nl=tf.identity)
 
-        y = one_hot(label, 10)
-        cost = tf.nn.softmax_cross_entropy_with_logits(logits, y)
+        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, label)
         cost = tf.reduce_mean(cost, name='cross_entropy_loss')
         tf.add_to_collection(MOVING_SUMMARY_VARS_KEY, cost)
 

diff --git a/examples/cifar10_resnet.py b/examples/cifar10_resnet.py
@@ -105,8 +105,7 @@ def residual(name, l, increase_dim=False, first=False):
         logits = FullyConnected('linear', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
-        y = one_hot(label, 10)
-        cost = tf.nn.softmax_cross_entropy_with_logits(logits, y)
+        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, label)
         cost = tf.reduce_mean(cost, name='cross_entropy_loss')
         tf.add_to_collection(MOVING_SUMMARY_VARS_KEY, cost)
 

diff --git a/examples/mnist_convnet.py b/examples/mnist_convnet.py
@@ -22,7 +22,7 @@
 
 """
 MNIST ConvNet example.
-about 0.6% validation error after 50 epochs.
+about 0.6% validation error after 30 epochs.
 """
 
 BATCH_SIZE = 128
@@ -58,8 +58,7 @@ def _get_cost(self, input_vars, is_training):
         logits = FullyConnected('fc1', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='prob')
 
-        y = one_hot(label, 10)
-        cost = tf.nn.softmax_cross_entropy_with_logits(logits, y)
+        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, label)
         cost = tf.reduce_mean(cost, name='cross_entropy_loss')
         tf.add_to_collection(MOVING_SUMMARY_VARS_KEY, cost)
 
@@ -97,7 +96,7 @@ def get_config():
         learning_rate=1e-3,
         global_step=get_global_step_var(),
         decay_steps=dataset_train.size() * 10,
-        decay_rate=0.5, staircase=True, name='learning_rate')
+        decay_rate=0.3, staircase=True, name='learning_rate')
     tf.scalar_summary('learning_rate', lr)
 
     return TrainConfig(

diff --git a/examples/svhn_digit_convnet.py b/examples/svhn_digit_convnet.py
@@ -48,8 +48,7 @@ def _get_cost(self, input_vars, is_training):
         logits = FullyConnected('linear', l, out_dim=10, nl=tf.identity)
         prob = tf.nn.softmax(logits, name='output')
 
-        y = one_hot(label, 10)
-        cost = tf.nn.softmax_cross_entropy_with_logits(logits, y)
+        cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, label)
         cost = tf.reduce_mean(cost, name='cross_entropy_loss')
         tf.add_to_collection(MOVING_SUMMARY_VARS_KEY, cost)