add variable names to docs of models/

examples/DoReFa-Net/README.md

@@ -11,13 +11,16 @@ Pretrained model for (1,4,32)-ResNet18 and (1,2,6)-AlexNet are available at
 They're provided in the format of numpy dictionary, so it should be very easy to port into other applications.
 The __binary-weight 4-bit-activation ResNet-18__ model has 59.2% top-1 validation accuracy.
 
+Note that when (W,A,G) is set to (1,32,32), this code is also an implementation of [Binary Weight Network](https://arxiv.org/abs/1511.00363).
+But with (W,A,G) set to (1,1,32), it is not equivalent to [XNOR-Net](https://arxiv.org/abs/1603.05279), although it won't be hard to implement it.
+
 Alternative link to this page: [http://dorefa.net](http://dorefa.net)
 
 ## Preparation:
 
 To use the script. You'll need:
 
-+ TensorFlow >= 0.12
++ TensorFlow >= 0.12.1
 
 + OpenCV bindings for Python
 

tensorpack/models/batch_norm.py

@@ -7,17 +7,17 @@ import tensorflow as tf
 from tensorflow.contrib.framework import add_model_variable
 from tensorflow.python.training import moving_averages
 
-from ..tfutils.common import get_tf_version
 from ..tfutils.tower import get_current_tower_context
-from ..utils import logger, building_rtfd
+from ..utils import logger
 from .common import layer_register
 
-__all__ = ['BatchNorm', 'BatchNormV1', 'BatchNormV2']
+__all__ = ['BatchNorm']
 
 # decay: being too close to 1 leads to slow start-up. torch use 0.9.
 # eps: torch: 1e-5. Lasagne: 1e-4
 
 
+# Deprecated. Only kept for future reference.
 @layer_register(log_shape=False)
 def BatchNormV1(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
     shape = x.get_shape().as_list()
@@ -110,6 +110,16 @@ def BatchNormV2(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
         decay (float): decay rate of moving average.
         epsilon (float): epsilon to avoid divide-by-zero.
 
+    Returns:
+        tf.Tensor: a tensor named ``output`` with the same shape of x.
+
+    Variable Names:
+
+    * ``beta``: the bias term.
+    * ``gamma``: the scale term. Input will be transformed by ``x * gamma + beta``.
+    * ``mean/EMA``: the moving average of mean.
+    * ``variance/EMA``: the moving average of variance.
+
     Note:
         * In multi-tower training, only the first training tower maintains a moving average.
           This is consistent with most frameworks.
@@ -171,6 +181,7 @@ def BatchNormV2(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
             x, moving_mean, moving_var, beta, gamma, epsilon)
 
     # TODO for other towers, maybe can make it depend some op later
+    # TODO update it later (similar to slim) might be faster?
     if ctx.is_main_training_tower:
         with tf.control_dependencies([update_op1, update_op2]):
             return tf.identity(xn, name='output')
@@ -178,8 +189,4 @@ def BatchNormV2(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
         return tf.identity(xn, name='output')
 
 
-if building_rtfd() or get_tf_version() >= 12:
-    BatchNorm = BatchNormV2
-else:
-    logger.warn("BatchNorm might be faster if you update TensorFlow")
-    BatchNorm = BatchNormV1
+BatchNorm = BatchNormV2

tensorpack/models/conv2d.py

@@ -30,6 +30,14 @@ def Conv2D(x, out_channel, kernel_shape,
         b_init: initializer for b. Defaults to zero.
         nl: a nonlinearity function.
         use_bias (bool): whether to use bias.
+
+    Returns:
+        tf.Tensor: a NHWC tensor named ``output``.
+
+    Variable Names:
+
+    * ``W``: weights
+    * ``b``: bias
     """
     in_shape = x.get_shape().as_list()
     in_channel = in_shape[-1]
@@ -95,6 +103,14 @@ def Deconv2D(x, out_shape, kernel_shape,
         b_init: initializer for b. Defaults to zero.
         nl: a nonlinearity function.
         use_bias (bool): whether to use bias.
+
+    Returns:
+        tf.Tensor: a NHWC tensor named ``output``.
+
+    Variable Names:
+
+    * ``W``: weights
+    * ``b``: bias
     """
     in_shape = x.get_shape().as_list()[1:]
     in_channel = in_shape[-1]

tensorpack/models/fc.py

@@ -25,6 +25,14 @@ def FullyConnected(x, out_dim,
         b_init: initializer for b. Defaults to zero.
         nl: a nonlinearity function
         use_bias (bool): whether to use bias.
+
+    Returns:
+        tf.Tensor: a NC tensor named ``output``.
+
+    Variable Names:
+
+    * ``W``: weights
+    * ``b``: bias
     """
     x = symbf.batch_flatten(x)
     in_dim = x.get_shape().as_list()[1]

tensorpack/models/image_sample.py

@@ -59,7 +59,7 @@ def ImageSample(inputs, borderMode='repeat'):
         borderMode: either "repeat" or "constant" (zero-filled)
 
     Returns:
-        a (N,H',W',C) tensor.
+        tf.Tensor: a tensor named ``output`` of shape (N,H',W',C).
     """
     # TODO borderValue
     template, mapping = inputs
@@ -102,7 +102,7 @@ def ImageSample(inputs, borderMode='repeat'):
         mask = tf.reduce_all(mask, [3])  # bxh2xw2 boolean
         mask = tf.expand_dims(mask, 3)
         ret = ret * tf.cast(mask, tf.float32)
-    return ret
+    return tf.identity(ret, name='output')
 
 
 class TestSample(TestModel):

tensorpack/models/nonlin.py

@@ -21,7 +21,7 @@ def Maxout(x, num_unit):
         num_unit (int): a int. Must be divisible by C.
 
     Returns:
-        tf.Tensor: of shape NHW(C/num_unit).
+        tf.Tensor: of shape NHW(C/num_unit) named ``output``.
     """
     input_shape = x.get_shape().as_list()
     ndim = len(input_shape)
@@ -46,6 +46,10 @@ def PReLU(x, init=0.001, name='output'):
         x (tf.Tensor): input
         init (float): initial value for the learnable slope.
         name (str): name of the output.
+
+    Variable Names:
+
+    * ``alpha``: learnable slope.
     """
     init = tf.constant_initializer(init)
     alpha = tf.get_variable('alpha', [], initializer=init)

tensorpack/models/pool.py

@@ -25,6 +25,9 @@ def MaxPooling(x, shape, stride=None, padding='VALID'):
         shape: int or (h, w) tuple
         stride: int or (h, w) tuple. Defaults to be the same as shape.
         padding (str): 'valid' or 'same'.
+
+    Returns:
+        tf.Tensor: a NHWC tensor named ``output``.
     """
     padding = padding.upper()
     shape = shape4d(shape)
@@ -48,6 +51,9 @@ def AvgPooling(x, shape, stride=None, padding='VALID'):
         shape: int or (h, w) tuple
         stride: int or (h, w) tuple. Defaults to be the same as shape.
         padding (str): 'valid' or 'same'.
+
+    Returns:
+        tf.Tensor: a NHWC tensor named ``output``.
     """
     padding = padding.upper()
     shape = shape4d(shape)
@@ -69,7 +75,7 @@ def GlobalAvgPooling(x):
     Args:
         x (tf.Tensor): a NHWC tensor.
     Returns:
-        tf.Tensor: a NC tensor.
+        tf.Tensor: a NC tensor named ``output``.
     """
     assert x.get_shape().ndims == 4
     return tf.reduce_mean(x, [1, 2], name='output')
@@ -101,6 +107,9 @@ def FixedUnPooling(x, shape, unpool_mat=None):
         shape: int or (h, w) tuple
         unpool_mat: a tf.Tensor or np.ndarray 2D matrix with size=shape.
             If is None, will use a matrix with 1 at top-left corner.
+
+    Returns:
+        tf.Tensor: a NHWC tensor.
     """
     shape = shape2d(shape)
 
@@ -138,6 +147,9 @@ def BilinearUpSample(x, shape):
     Args:
         x (tf.Tensor): a NHWC tensor
         shape (int): the upsample factor
+
+    Returns:
+        tf.Tensor: a NHWC tensor.
     """
     # inp_shape = tf.shape(x)
     # return tf.image.resize_bilinear(x,

tensorpack/models/shapes.py

@@ -19,6 +19,7 @@ def ConcatWith(x, dim, tensor):
         dim (int): the dimension along which to concatenate
         tensor (list[tf.Tensor]): a tensor or list of tensors to concatenate with x.
             x will be at the beginning
+
     Returns:
         tf.Tensor: ``tf.concat_v2([x] + tensor, dim)``
     """

tensorpack/models/softmax.py

 #!/usr/bin/env python
-# -*- coding: utf-8 -*-
-# File: softmax.py
+# -*- coding: utf-8 -*- File: softmax.py
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
 import tensorflow as tf
@@ -17,9 +16,17 @@ def SoftMax(x, use_temperature=False, temperature_init=1.0):
     <https://arxiv.org/abs/1503.02531>`_.
 
     Args:
-        x (tf.Tensor): input
+        x (tf.Tensor): input of any dimension. Softmax will be performed on
+            the last dimension.
         use_temperature (bool): use a learnable temperature or not.
         temperature_init (float): initial value of the temperature.
+
+    Returns:
+        tf.Tensor: a tensor of the same shape named ``output``.
+
+    Variable Names:
+
+    * ``invtemp``: 1.0/temperature.
     """
     if use_temperature:
         t = tf.get_variable('invtemp', [],

tensorpack/utils/viz.py

@@ -182,7 +182,7 @@ def build_patch_list(patch_list,
 def dump_dataflow_images(df, index=0, batched=True,
                          number=1000, output_dir=None,
                          scale=1, resize=None, viz=None,
-                         flipRGB=False, exit_after=True):
+                         flipRGB=False):
     """
     Dump or visualize images of a :class:`DataFlow`.
 
@@ -199,7 +199,6 @@ def dump_dataflow_images(df, index=0, batched=True,
             with :func:`build_patch_list` for visualization. No visualization will happen by
             default.
         flipRGB (bool): apply a RGB<->BGR conversion or not.
-        exit_after (bool): ``sys.exit()`` after this function.
     """
     if output_dir:
         mkdir_p(output_dir)
@@ -221,9 +220,6 @@ def dump_dataflow_images(df, index=0, batched=True,
             for img in imgbatch:
                 cnt += 1
                 if cnt == number:
-                    if exit_after:
-                        sys.exit()
-                    else:
                     return
                 if scale != 1:
                     img = img * scale