example docs and improvements

README.md

@@ -41,7 +41,7 @@ Multi-GPU training is off-the-shelf by simply switching the trainer.
 ## Dependencies:
 
 + Python 2 or 3
-+ TensorFlow >= 0.8
++ TensorFlow >= 0.10
 + Python bindings for OpenCV
 + other requirements:
 ```

examples/HED/README.md

@@ -27,6 +27,7 @@ To start training:
 ```bash
 ./hed.py --load vgg16.npy
 ```
+It takes about 100k steps (~10 hour on a TitanX) to reach a reasonable performance.
 
 To inference (produce a heatmap at each level at out*.png):
 ```bash

examples/Inception/inceptionv3.py

@@ -21,8 +21,8 @@ See "Rethinking the Inception Architecture for Computer Vision", arxiv:1512.0056
 
 This config follows the official inceptionv3 setup (https://github.com/tensorflow/models/tree/master/inception/inception)
 with much much fewer lines of code.
-It reaches 74.5% single-crop validation accuracy, slightly better than the official code,
-and has the same running speed as well.
+It reaches 74% single-crop validation accuracy,
+and has the same running speed as the official code.
 The hyperparameters here are for 8 GPUs, so the effective batch size is 8*64 = 512.
 With 8 TitanX it runs about 0.45 it/s.
 """

examples/cifar-convnet.py

@@ -51,11 +51,9 @@ class Model(ModelDesc):
                     .MaxPooling('pool2', 3, stride=2, padding='SAME') \
                     .Conv2D('conv3.1', out_channel=128, padding='VALID') \
                     .Conv2D('conv3.2', out_channel=128, padding='VALID') \
-                    .FullyConnected('fc0', 1024 + 512, nl=tf.nn.relu,
-                           b_init=tf.constant_initializer(0.1)) \
+                    .FullyConnected('fc0', 1024 + 512, nl=tf.nn.relu) \
                     .tf.nn.dropout(keep_prob) \
-                    .FullyConnected('fc1', 512, nl=tf.nn.relu,
-                           b_init=tf.constant_initializer(0.1)) \
+                    .FullyConnected('fc1', 512, nl=tf.nn.relu) \
                     .FullyConnected('linear', out_dim=self.cifar_classnum, nl=tf.identity)()
 
         cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits, label)

examples/svhn-digit-convnet.py

@@ -14,9 +14,10 @@ from tensorpack.tfutils.summary import *
 
 """
 A very small SVHN convnet model (only 0.8m parameters).
-About 3.0% validation error after 70 epoch. 2.5% after 130 epoch.
+About 2.3% validation error after 70 epochs. 2.15% after 150 epochs.
 
-Each epoch is set to 4721 iterations. The speed is about 44 it/s on a Tesla M40
+Each epoch iterates over the whole training set (4721 iterations).
+Speed is about 43 it/s on TitanX.
 """
 
 class Model(ModelDesc):
@@ -29,7 +30,7 @@ class Model(ModelDesc):
 
         image = image / 128.0 - 1
 
-        with argscope(Conv2D, nl=tf.nn.relu):
+        with argscope(Conv2D, nl=BNReLU, use_bias=False):
             logits = (LinearWrap(image)
                     .Conv2D('conv1', 24, 5, padding='VALID')
                     .MaxPooling('pool1', 2, padding='SAME')
@@ -62,7 +63,7 @@ def get_data():
     d1 = dataset.SVHNDigit('train')
     d2 = dataset.SVHNDigit('extra')
     data_train = RandomMixData([d1, d2])
-    data_test = dataset.SVHNDigit('test')
+    data_test = dataset.SVHNDigit('test', shuffle=False)
 
     augmentors = [
         imgaug.Resize((40, 40)),

scripts/dump-model-params.py

@@ -24,7 +24,7 @@ with tf.Graph().as_default() as G:
     if args.config:
         MODEL = imp.load_source('config_script', args.config).Model
         M = MODEL()
-        M.build_graph(M.get_input_vars(), is_training=False)
+        M.build_graph(M.get_input_vars())
     else:
         M = ModelFromMetaGraph(args.meta)
 

tensorpack/models/conv2d.py

@@ -60,8 +60,7 @@ def Conv2D(x, out_channel, kernel_shape,
                    for i, k in zip(inputs, kernels)]
         conv = tf.concat(3, outputs)
     if nl is None:
-        logger.warn("[DEPRECATED] Default nonlinearity for Conv2D and FullyConnected will be deprecated.")
-        logger.warn("[DEPRECATED] Please use argscope instead.")
+        logger.warn("[DEPRECATED] Default ReLU nonlinearity for Conv2D and FullyConnected will be deprecated. Please use argscope instead.")
         nl = tf.nn.relu
     return nl(tf.nn.bias_add(conv, b) if use_bias else conv, name='output')
 

tensorpack/models/fc.py

@@ -40,7 +40,6 @@ def FullyConnected(x, out_dim,
         b = tf.get_variable('b', [out_dim], initializer=b_init)
     prod = tf.nn.xw_plus_b(x, W, b) if use_bias else tf.matmul(x, W)
     if nl is None:
-        logger.warn("[DEPRECATED] Default nonlinearity for Conv2D and FullyConnected will be deprecated.")
-        logger.warn("[DEPRECATED] Please use argscope instead.")
+        logger.warn("[DEPRECATED] Default ReLU nonlinearity for Conv2D and FullyConnected will be deprecated. Please use argscope instead.")
         nl = tf.nn.relu
     return nl(prod, name='output')

tensorpack/train/config.py

@@ -50,13 +50,25 @@ class TrainConfig(object):
         assert_type(self.session_config, tf.ConfigProto)
         self.session_init = kwargs.pop('session_init', JustCurrentSession())
         assert_type(self.session_init, SessionInit)
-        self.step_per_epoch = int(kwargs.pop('step_per_epoch'))
+
+        self.step_per_epoch = kwargs.pop('step_per_epoch', None)
+        if self.step_per_epoch is None:
+            try:
+                self.step_per_epoch = self.dataset.size()
+            except NotImplementedError:
+                logger.exception("You must set `step_per_epoch` if dataset.size() is not implemented.")
+        else:
+            self.step_per_epoch = int(self.step_per_epoch)
+
         self.starting_epoch = int(kwargs.pop('starting_epoch', 1))
         self.max_epoch = int(kwargs.pop('max_epoch', 99999))
         assert self.step_per_epoch >= 0 and self.max_epoch > 0
 
-        if 'nr_tower' in kwargs or 'tower' in kwargs:
-            self.set_tower(**kwargs)
+        if 'nr_tower' in kwargs:
+            assert 'tower' not in kwargs, "Cannot set both nr_tower and tower in TrainConfig!"
+            self.nr_tower = kwargs.pop('nr_tower')
+        elif 'tower' in kwargs:
+            self.tower = kwargs.pop('tower')
         else:
             self.tower = [0]
 
@@ -64,8 +76,8 @@ class TrainConfig(object):
         assert len(kwargs) == 0, 'Unknown arguments: {}'.format(str(kwargs.keys()))
 
     def set_tower(self, nr_tower=None, tower=None):
-        logger.warn("config.set_tower is deprecated. set config.tower or config.nr_tower directly")
         # this is a deprecated function
+        logger.warn("config.set_tower is deprecated. set config.tower or config.nr_tower directly")
         assert nr_tower is None or tower is None, "Cannot set both nr_tower and tower!"
         if nr_tower:
             tower = list(range(nr_tower))