change name for .print

examples/DisturbLabel/mnist-disturb.py

@@ -25,9 +25,7 @@ mnist_example.get_data = get_data
 IMAGE_SIZE = 28
 
 class Model(mnist_example.Model):
-    def _build_graph(self, input_vars, is_training):
-        is_training = bool(is_training)
-        keep_prob = tf.constant(0.5 if is_training else 1.0)
+    def _build_graph(self, input_vars):
 
         image, label = input_vars
         image = tf.expand_dims(image, 3)    # add a single channel

examples/DoReFa-Net/alexnet-dorefa.py

@@ -72,7 +72,7 @@ class Model(ModelDesc):
         return [InputVar(tf.float32, [None, 224, 224, 3], 'input'),
                 InputVar(tf.int32, [None], 'label') ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         image, label = input_vars
         image = image / 255.0
 
@@ -97,7 +97,7 @@ class Model(ModelDesc):
         def activate(x):
             return fa(nonlin(x))
 
-        with argscope(BatchNorm, decay=0.9, epsilon=1e-4, use_local_stat=is_training), \
+        with argscope(BatchNorm, decay=0.9, epsilon=1e-4), \
                 argscope([Conv2D, FullyConnected], use_bias=False, nl=tf.identity):
             logits = (LinearWrap(image)
                 .Conv2D('conv0', 96, 12, stride=4, padding='VALID')

examples/DoReFa-Net/svhn-digit-dorefa.py

@@ -69,7 +69,7 @@ class Model(ModelDesc):
 
         image = image / 256.0
 
-        with argscope(BatchNorm, decay=0.9, epsilon=1e-4, use_local_stat=is_training), \
+        with argscope(BatchNorm, decay=0.9, epsilon=1e-4), \
                 argscope(Conv2D, use_bias=False, nl=tf.identity):
             logits = (LinearWrap(image)
                 .Conv2D('conv0', 48, 5, padding='VALID', use_bias=True)

examples/char-rnn/char-rnn.py

@@ -74,7 +74,7 @@ class Model(ModelDesc):
         return [InputVar(tf.int32, (None, param.seq_len), 'input'),
                 InputVar(tf.int32, (None, param.seq_len), 'nextinput') ]
 
-    def _build_graph(self, input_vars, is_training):
+    def _build_graph(self, input_vars):
         input, nextinput = input_vars
 
         cell = rnn_cell.BasicLSTMCell(num_units=param.rnn_size)

examples/load-alexnet.py

@@ -28,10 +28,8 @@ class Model(ModelDesc):
         return [InputVar(tf.float32, (None, 227, 227, 3), 'input'),
             InputVar(tf.int32, (None,), 'label') ]
 
-    def _build_graph(self, inputs, is_training):
+    def _build_graph(self, inputs):
         # img: 227x227x3
-        is_training = bool(is_training)
-        keep_prob = tf.constant(0.5 if is_training else 1.0)
 
         image, label = inputs
 

examples/load-vgg16.py

@@ -8,7 +8,7 @@ import tensorflow as tf
 import numpy as np
 import os
 import argparse
-import cPickle as pkl
+import pickle as pkl
 
 from tensorpack.train import TrainConfig
 from tensorpack.predict import PredictConfig, get_predict_func
@@ -33,8 +33,6 @@ class Model(ModelDesc):
                 InputVar(tf.int32, (None,), 'label') ]
 
     def _build_graph(self, inputs, is_training):
-        is_training = bool(is_training)
-        keep_prob = tf.constant(0.5 if is_training else 1.0)
 
         image, label = inputs
 
@@ -65,9 +63,10 @@ class Model(ModelDesc):
                 .MaxPooling('pool5', 2)
                  # 7
                 .FullyConnected('fc6', 4096)
-                .tf.nn.dropout(keep_prob)
+                .Dropout('drop0', 0.5)
+                .print_tensor()
                 .FullyConnected('fc7', 4096)
-                .tf.nn.dropout(keep_prob)
+                .Dropout('drop1', 0.5)
                 .FullyConnected('fc8', out_dim=1000, nl=tf.identity)())
         prob = tf.nn.softmax(logits, name='output')
 
@@ -93,10 +92,10 @@ def run_test(path, input):
     outputs = predict_func([im])[0]
     prob = outputs[0]
     ret = prob.argsort()[-10:][::-1]
-    print ret
+    print(ret)
 
     meta = ILSVRCMeta().get_synset_words_1000()
-    print [meta[k] for k in ret]
+    print([meta[k] for k in ret])
 
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()

tensorpack/models/__init__.py

@@ -49,9 +49,14 @@ class LinearWrap(object):
         layer = eval(layer_name)
         if hasattr(layer, 'f'):
             # this is a registered tensorpack layer
-            def f(name, *args, **kwargs):
-                ret = layer(name, self._t, *args, **kwargs)
-                return LinearWrap(ret)
+            if layer.use_scope:
+                def f(name, *args, **kwargs):
+                    ret = layer(name, self._t, *args, **kwargs)
+                    return LinearWrap(ret)
+            else:
+                def f(*args, **kwargs):
+                    ret = layer(self._t, *args, **kwargs)
+                    return LinearWrap(ret)
             return f
         else:
             if layer_name != 'tf':
@@ -70,7 +75,7 @@ class LinearWrap(object):
     def tensor(self):
         return self._t
 
-    def print(self):
+    def print_tensor(self):
         print(self._t)
         return self
 

tensorpack/models/_common.py

@@ -22,7 +22,10 @@ def disable_layer_logging():
     # can use nonlocal in python3, but how
     globals()['_layer_logged'] = ContainEverything()
 
-def layer_register(summary_activation=False, log_shape=True):
+def layer_register(
+        summary_activation=False,
+        log_shape=True,
+        use_scope=True):
     """
     Register a layer.
     :param summary_activation: Define the default behavior of whether to
@@ -33,40 +36,52 @@ def layer_register(summary_activation=False, log_shape=True):
 
     def wrapper(func):
         @wraps(func)
-        def wrapped_func(name, inputs, *args, **kwargs):
-            assert isinstance(name, six.string_types), name
+        def wrapped_func(*args, **kwargs):
+            if use_scope:
+                name, inputs = args[0], args[1]
+                args = args[1:] # actual positional args used to call func
+                assert isinstance(name, six.string_types), name
+            else:
+                assert not log_shape and not summary_activation
+                inputs = args[0]
+                name = None
             do_summary = kwargs.pop(
                 'summary_activation', summary_activation)
-            args = (inputs,) + args
 
             # TODO use inspect.getcallargs to enhance?
             # update from current argument scope
             actual_args = copy.copy(get_arg_scope()[func.__name__])
             actual_args.update(kwargs)
 
-            with tf.variable_scope(name) as scope:
-                do_log_shape = log_shape and scope.name not in _layer_logged
-                do_summary = do_summary and scope.name not in _layer_logged
-                if do_log_shape:
-                    logger.info("{} input: {}".format(scope.name, get_shape_str(inputs)))
+            if name is not None:
+                with tf.variable_scope(name) as scope:
+                    do_log_shape = log_shape and scope.name not in _layer_logged
+                    do_summary = do_summary and scope.name not in _layer_logged
+                    if do_log_shape:
+                        logger.info("{} input: {}".format(scope.name, get_shape_str(inputs)))
 
+                    # run the actual function
+                    outputs = func(*args, **actual_args)
+
+                    if do_log_shape:
+                        # log shape info and add activation
+                        logger.info("{} output: {}".format(
+                            scope.name, get_shape_str(outputs)))
+                        _layer_logged.add(scope.name)
+
+                    if do_summary:
+                        if isinstance(outputs, list):
+                            for x in outputs:
+                                add_activation_summary(x, scope.name)
+                        else:
+                            add_activation_summary(outputs, scope.name)
+            else:
                 # run the actual function
                 outputs = func(*args, **actual_args)
+            return outputs
 
-                if do_log_shape:
-                    # log shape info and add activation
-                    logger.info("{} output: {}".format(
-                        scope.name, get_shape_str(outputs)))
-                    _layer_logged.add(scope.name)
-
-                if do_summary:
-                    if isinstance(outputs, list):
-                        for x in outputs:
-                            add_activation_summary(x, scope.name)
-                    else:
-                        add_activation_summary(outputs, scope.name)
-                return outputs
         wrapped_func.f = func   # attribute to access the underlining function object
+        wrapped_func.use_scope = use_scope
         return wrapped_func
 
     # need some special handling for sphinx to work with the arguments

tensorpack/models/nonlin.py

@@ -11,7 +11,7 @@ from .batch_norm import BatchNorm
 
 __all__ = ['Maxout', 'PReLU', 'LeakyReLU', 'BNReLU']
 
-@layer_register(log_shape=False)
+@layer_register()
 def Maxout(x, num_unit):
     """
     Maxout as in `Maxout Networks <http://arxiv.org/abs/1302.4389>`_.

tensorpack/models/regularize.py

@@ -7,8 +7,10 @@ import re
 
 from ..utils import logger
 from ..utils.utils import *
+from .model_desc import get_current_tower_context
+from ._common import layer_register
 
-__all__ = ['regularize_cost', 'l2_regularizer', 'l1_regularizer']
+__all__ = ['regularize_cost', 'l2_regularizer', 'l1_regularizer', 'Dropout']
 
 @memoized
 def _log_regularizer(name):
@@ -36,3 +38,10 @@ def regularize_cost(regex, func, name=None):
         return 0
     return tf.add_n(costs, name=name)
 
+
+@layer_register(log_shape=False)
+def Dropout(x, prob=0.5):
+    is_training = get_current_tower_context().is_training
+    keep_prob = tf.constant(prob if is_training else 1.0)
+    return tf.nn.dropout(x, keep_prob)
+