clean-up many old deprecations

docs/conf.py

 # -*- coding: utf-8 -*-
-#
+# flake8: noqa
 # tensorpack documentation build configuration file, created by
 # sphinx-quickstart on Sun Mar 27 01:41:24 2016.
 #
@@ -92,8 +92,8 @@ master_doc = 'index'
 
 # General information about the project.
 project = u'tensorpack'
-copyright = u'2015 - 2017, Yuxin Wu'
-author = u'Yuxin Wu'
+copyright = u'2015 - 2018, Yuxin Wu, et al.'
+author = u'Yuxin Wu, et al.'
 
 # The version info for the project you're documenting, acts as replacement for
 # |version| and |release|, also used in various other places throughout the
@@ -365,12 +365,10 @@ def autodoc_skip_member(app, what, name, obj, skip, options):
     # Hide some names that are deprecated or not intended to be used
     if name in [
         # deprecated stuff:
-        'GaussianDeform',
-        'set_tower_func',
         'TryResumeTraining',
+        'QueueInputTrainer',
 
         # renamed stuff:
-        'dump_chkpt_vars',
         'DumpTensor',
         'DumpParamAsImage',
         'StagingInputWrapper',
@@ -378,9 +376,9 @@ def autodoc_skip_member(app, what, name, obj, skip, options):
         'get_nr_gpu',
 
         # deprecated or renamed symbolic code
-        'Deconv2D', 'LeakyReLU',
-        'saliency_map', 'get_scalar_var', 'psnr',
-        'prediction_incorrect', 'huber_loss', 'SoftMax'
+        'Deconv2D',
+        'get_scalar_var', 'psnr',
+        'prediction_incorrect', 'huber_loss',
 
         # internal only
         'apply_default_prefetch',

examples/DisturbLabel/README.md

@@ -3,7 +3,8 @@
 
 I ran into the paper [DisturbLabel: Regularizing CNN on the Loss Layer](https://arxiv.org/abs/1605.00055) on CVPR16,
 which basically said that noisy data gives you better performance.
-As many, I didn't believe the method and the results.
+As many, I didn't believe the method and the results. This code exists to
+disprove the results in the paper.
 
 This is a simple mnist training script with DisturbLabel. It uses the simple architecture in the paper, and
 hyperparameters in my original [mnist example](../mnist-convnet.py).
@@ -21,6 +22,7 @@ The method didn't work for slightly harder problems such as SVHN:
 ![svhn](svhn.png)
 
 The SVHN experiements used the model & hyperparemeters as my original [svhn example](../svhn-digit-convnet.py).
-Experiements were all repeated 10 times to get the error bar.
+Experiements were all repeated 10 times to get the error bar. It apparently does not work.
 
-And I don't believe it will work for ImageNet either. And that's a CVPR paper..
+It will not work for ImageNet either. There is indeed a terribly weak
+ImageNet experiment in this paper, and that's a CVPR paper.

examples/DisturbLabel/disturb.py

@@ -11,7 +11,8 @@ class DisturbLabel(ProxyDataFlow, RNGDataFlow):
         self.prob = prob
 
     def reset_state(self):
-        super(DisturbLabel, self).reset_state()
+        RNGDataFlow.reset_state(self)
+        ProxyDataFlow.reset_state(self)
 
     def get_data(self):
         for dp in self.ds.get_data():

examples/DisturbLabel/mnist-disturb.py

 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 # File: mnist-disturb.py
-# Author: Yuxin Wu
 
 import os
 import argparse
 
 
 from tensorpack import *
+from tensorpack.utils import logger
 from tensorpack.dataflow import dataset
 import tensorflow as tf
 from disturb import DisturbLabel
 
 import imp
 mnist_example = imp.load_source('mnist_example',
-                                os.path.join(os.path.dirname(__file__), '..', 'mnist-convnet.py'))
+                                os.path.join(os.path.dirname(__file__), '..', 'basics', 'mnist-convnet.py'))
 get_config = mnist_example.get_config
 
 
@@ -25,7 +25,6 @@ def get_data():
 
 
 mnist_example.get_data = get_data
-IMAGE_SIZE = 28
 
 
 class Model(mnist_example.Model):
@@ -41,7 +40,7 @@ class Model(mnist_example.Model):
                   .FullyConnected('fc1', out_dim=10, activation=tf.identity)())
         tf.nn.softmax(logits, name='prob')
 
-        wrong = symbolic_functions.prediction_incorrect(logits, label)
+        wrong = tf.cast(tf.logical_not(tf.nn.in_top_k(logits, label, 1)), tf.float32, name='incorrect_vector')
         add_moving_summary(tf.reduce_mean(wrong, name='train_error'))
 
         cost = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=label)
@@ -60,5 +59,6 @@ if __name__ == '__main__':
     if args.gpu:
         os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
 
+    logger.auto_set_dir()
     config = get_config()
     launch_train_with_config(config, SimpleTrainer())

examples/DisturbLabel/svhn-disturb.py

 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 # File: svhn-disturb.py
-# Author: Yuxin Wu
 
 import argparse
 import os
@@ -9,14 +8,15 @@ import imp
 
 
 from tensorpack import *
+from tensorpack.utils import logger
 from tensorpack.dataflow import dataset
 
 from disturb import DisturbLabel
 
 svhn_example = imp.load_source('svhn_example',
-                               os.path.join(os.path.dirname(__file__), '..', 'svhn-digit-convnet.py'))
+                               os.path.join(os.path.dirname(__file__), '..',
+                                            'basics', 'svhn-digit-convnet.py'))
 Model = svhn_example.Model
-get_config = svhn_example.get_config
 
 
 def get_data():
@@ -41,19 +41,21 @@ def get_data():
     return data_train, data_test
 
 
-svhn_example.get_data = get_data
-
 if __name__ == '__main__':
     parser = argparse.ArgumentParser()
-    parser.add_argument('--gpu', help='a gpu to use')
-    parser.add_argument('--prob', help='disturb prob',
-                        type=float, required=True)
+    parser.add_argument('--prob', help='disturb prob', type=float, required=True)
     args = parser.parse_args()
 
-    if args.gpu:
-        os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
-    else:
-        os.environ['CUDA_VISIBLE_DEVICES'] = '0'
-
-    config = get_config(args.prob)
+    logger.auto_set_dir()
+    data_train, data_test = get_data()
+    config = TrainConfig(
+        model=Model(),
+        data=QueueInput(data_train),
+        callbacks=[
+            ModelSaver(),
+            InferenceRunner(data_test,
+                            ScalarStats(['cost', 'accuracy']))
+        ],
+        max_epoch=350,
+    )
     launch_train_with_config(config, SimpleTrainer())

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

@@ -6,7 +6,6 @@
 import argparse
 
 from tensorpack import *
-from tensorpack.tfutils.symbolic_functions import prediction_incorrect
 from tensorpack.tfutils.summary import add_moving_summary, add_param_summary
 from tensorpack.dataflow import dataset
 from tensorpack.tfutils.varreplace import remap_variables
@@ -109,7 +108,7 @@ class Model(ModelDesc):
         tf.nn.softmax(logits, name='output')
 
         # compute the number of failed samples
-        wrong = prediction_incorrect(logits, label)
+        wrong = tf.cast(tf.logical_not(tf.nn.in_top_k(logits, label, 1)), tf.float32, name='wrong_tensor')
         # monitor training error
         add_moving_summary(tf.reduce_mean(wrong, name='train_error'))
 
@@ -163,7 +162,7 @@ def get_config():
         callbacks=[
             ModelSaver(),
             InferenceRunner(data_test,
-                            [ScalarStats('cost'), ClassificationError()])
+                            [ScalarStats('cost'), ClassificationError('wrong_tensor')])
         ],
         model=Model(),
         max_epoch=200,

examples/ImageNetModels/inception-bn.py

@@ -9,7 +9,6 @@ import tensorflow as tf
 
 
 from tensorpack import *
-from tensorpack.tfutils.symbolic_functions import prediction_incorrect
 from tensorpack.tfutils.summary import add_moving_summary
 from tensorpack.dataflow import dataset
 from tensorpack.utils.gpu import get_nr_gpu
@@ -99,6 +98,9 @@ class Model(ModelDesc):
         cost = tf.add_n([loss3, 0.3 * loss2, 0.3 * loss1], name='weighted_cost')
         add_moving_summary([cost, loss1, loss2, loss3])
 
+        def prediction_incorrect(logits, label, topk, name):
+            return tf.cast(tf.logical_not(tf.nn.in_top_k(logits, label, topk)), tf.float32, name=name)
+
         wrong = prediction_incorrect(logits, label, 1, name='wrong-top1')
         add_moving_summary(tf.reduce_mean(wrong, name='train_error_top1'))
 

tensorpack/dataflow/imgaug/deform.py

@@ -6,7 +6,11 @@ from .base import ImageAugmentor
 from ...utils import logger
 import numpy as np
 
-__all__ = ['GaussianDeform']
+__all__ = []
+
+# Code was temporarily kept here for a future reference in case someone needs it
+# But it was already deprecated,
+# because this augmentation is not a general one that people will often find helpful.
 
 
 class GaussianMap(object):

tensorpack/models/fc.py

@@ -3,14 +3,24 @@
 
 
 import tensorflow as tf
+import numpy as np
 
 from .common import layer_register, VariableHolder
 from .tflayer import convert_to_tflayer_args, rename_get_variable
-from ..tfutils import symbolic_functions as symbf
 
 __all__ = ['FullyConnected']
 
 
+def batch_flatten(x):
+    """
+    Flatten the tensor except the first dimension.
+    """
+    shape = x.get_shape().as_list()[1:]
+    if None not in shape:
+        return tf.reshape(x, [-1, int(np.prod(shape))])
+    return tf.reshape(x, tf.stack([tf.shape(x)[0], -1]))
+
+
 @layer_register(log_shape=True)
 @convert_to_tflayer_args(
     args_names=['units'],
@@ -36,7 +46,7 @@ def FullyConnected(
     * ``b``: bias
     """
 
-    inputs = symbf.batch_flatten(inputs)
+    inputs = batch_flatten(inputs)
     with rename_get_variable({'kernel': 'W', 'bias': 'b'}):
         layer = tf.layers.Dense(
             units=units,

tensorpack/models/nonlin.py

@@ -6,9 +6,8 @@ import tensorflow as tf
 
 from .common import layer_register, VariableHolder
 from .batch_norm import BatchNorm
-from ..utils.develop import log_deprecated
 
-__all__ = ['Maxout', 'PReLU', 'LeakyReLU', 'BNReLU']
+__all__ = ['Maxout', 'PReLU', 'BNReLU']
 
 
 @layer_register(use_scope=None)
@@ -60,21 +59,6 @@ def PReLU(x, init=0.001, name='output'):
     return ret
 
 
-@layer_register(use_scope=None)
-def LeakyReLU(x, alpha, name='output'):
-    """
-    Leaky ReLU as in paper `Rectifier Nonlinearities Improve Neural Network Acoustic
-    Models
-    <http://ai.stanford.edu/~amaas/papers/relu_hybrid_icml2013_final.pdf>`_.
-
-    Args:
-        x (tf.Tensor): input
-        alpha (float): the slope.
-    """
-    log_deprecated("LeakyReLU", "Use tf.nn.leaky_relu in TF 1.4 instead!", "2018-03-30")
-    return tf.maximum(x, alpha * x, name=name)
-
-
 @layer_register(use_scope=None)
 def BNReLU(x, name=None):
     """

tensorpack/models/softmax.py

-# -*- coding: utf-8 -*- File: softmax.py
-
-
-import tensorflow as tf
-from .common import layer_register
-from ..utils.develop import log_deprecated
-
-__all__ = ['SoftMax']
-
-
-@layer_register(use_scope=None)
-def SoftMax(x, use_temperature=False, temperature_init=1.0):
-    """
-    A SoftMax layer (w/o linear projection) with optional temperature, as
-    defined in the paper `Distilling the Knowledge in a Neural Network
-    <https://arxiv.org/abs/1503.02531>`_.
-
-    Args:
-        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.
-    """
-    log_deprecated("models.SoftMax", "Please implement it by yourself!", "2018-05-01")
-    if use_temperature:
-        t = tf.get_variable('invtemp', [],
-                            initializer=tf.constant_initializer(1.0 / float(temperature_init)))
-        x = x * t
-    return tf.nn.softmax(x, name='output')

tensorpack/tfutils/sessinit.py

@@ -262,7 +262,7 @@ def get_model_loader(filename):
         return SaverRestore(filename)
 
 
-@deprecated("Write the logic yourself or use AutoResumeTrainConfig!", "2018-06-01")
+@deprecated("It's better to write the logic yourself or use AutoResumeTrainConfig!", "2018-07-01")
 def TryResumeTraining():
     """
     Try loading latest checkpoint from ``logger.get_logger_dir()``, only if there is one.

tensorpack/tfutils/symbolic_functions.py

@@ -7,15 +7,15 @@ import numpy as np
 
 from ..utils.develop import deprecated
 
-# __all__ = ['get_scalar_var']
+__all__ = ['get_scalar_var', 'prediction_incorrect', 'flatten', 'batch_flatten', 'print_stat', 'rms', 'huber_loss']
 
 
 # this function exists for backwards-compatibility
 def prediction_incorrect(logits, label, topk=1, name='incorrect_vector'):
-    return tf.cast(tf.logical_not(tf.nn.in_top_k(logits, label, topk)),
-                   tf.float32, name=name)
+    return tf.cast(tf.logical_not(tf.nn.in_top_k(logits, label, topk)), tf.float32, name=name)
 
 
+@deprecated("Please implement it yourself!", "2018-08-01")
 def flatten(x):
     """
     Flatten the tensor.
@@ -23,6 +23,7 @@ def flatten(x):
     return tf.reshape(x, [-1])
 
 
+@deprecated("Please implement it yourself!", "2018-08-01")
 def batch_flatten(x):
     """
     Flatten the tensor except the first dimension.
@@ -46,6 +47,8 @@ def print_stat(x, message=None):
                     message=message, name='print_' + x.op.name)
 
 
+# after deprecated, keep it for internal use only
+# @deprecated("Please implement it yourself!", "2018-08-01")
 def rms(x, name=None):
     """
     Returns:
@@ -58,7 +61,7 @@ def rms(x, name=None):
     return tf.sqrt(tf.reduce_mean(tf.square(x)), name=name)
 
 
-@deprecated("Please use tf.losses.huber_loss instead!")
+@deprecated("Please use tf.losses.huber_loss instead!", "2018-08-01")
 def huber_loss(x, delta=1, name='huber_loss'):
     r"""
     Huber loss of x.
@@ -88,6 +91,7 @@ def huber_loss(x, delta=1, name='huber_loss'):
 
 # TODO deprecate this in the future
 # doesn't hurt to keep it here for now
+@deprecated("Simply use tf.get_variable instead!", "2018-08-01")
 def get_scalar_var(name, init_value, summary=False, trainable=False):
     """
     Get a scalar float variable with certain initial value.
@@ -142,51 +146,3 @@ def psnr(prediction, ground_truth, maxp=None, name='psnr'):
         psnr = tf.add(tf.multiply(20., log10(maxp)), psnr, name=name)
 
     return psnr
-
-
-@deprecated("Please implement it by yourself.", "2018-04-28")
-def saliency_map(output, input, name="saliency_map"):
-    """
-    Produce a saliency map as described in the paper:
-    `Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps
-    <https://arxiv.org/abs/1312.6034>`_.
-    The saliency map is the gradient of the max element in output w.r.t input.
-
-    Returns:
-        tf.Tensor: the saliency map. Has the same shape as input.
-    """
-    max_outp = tf.reduce_max(output, 1)
-    saliency_op = tf.gradients(max_outp, input)[:][0]
-    saliency_op = tf.identity(saliency_op, name=name)
-    return saliency_op
-
-
-@deprecated("Please implement it by yourself.", "2018-04-28")
-def shapeless_placeholder(x, axis, name):
-    """
-    Make the static shape of a tensor less specific.
-
-    If you want to feed to a tensor, the shape of the feed value must match
-    the tensor's static shape. This function creates a placeholder which
-    defaults to x if not fed, but has a less specific static shape than x.
-    See also `tensorflow#5680
-    <https://github.com/tensorflow/tensorflow/issues/5680>`_.
-
-    Args:
-        x: a tensor
-        axis(int or list of ints): these axes of ``x.get_shape()`` will become
-            None in the output.
-        name(str): name of the output tensor
-
-    Returns:
-        a tensor equal to x, but shape information is partially cleared.
-    """
-    shp = x.get_shape().as_list()
-    if not isinstance(axis, list):
-        axis = [axis]
-    for a in axis:
-        if shp[a] is None:
-            raise ValueError("Axis {} of shape {} is already unknown!".format(a, shp))
-        shp[a] = None
-    x = tf.placeholder_with_default(x, shape=shp, name=name)
-    return x

tensorpack/tfutils/varmanip.py

@@ -11,7 +11,7 @@ from ..utils.develop import deprecated
 from ..utils import logger
 from .common import get_op_tensor_name
 
-__all__ = ['SessionUpdate', 'dump_session_params', 'dump_chkpt_vars',
+__all__ = ['SessionUpdate', 'dump_session_params',
            'load_chkpt_vars', 'save_chkpt_vars', 'get_checkpoint_path']
 
 

tensorpack/train/tower.py

@@ -6,7 +6,6 @@ import six
 from abc import abstractmethod, ABCMeta
 
 from ..utils.argtools import call_only_once, memoized
-from ..utils.develop import deprecated
 from ..graph_builder.predict import SimplePredictBuilder
 from ..input_source import PlaceholderInput
 from ..predict.base import OnlinePredictor
@@ -37,10 +36,6 @@ class TowerTrainer(Trainer):
         assert isinstance(tower_func, TowerFuncWrapper), tower_func
         self._tower_func = tower_func
 
-    @deprecated("Just use tower_func = xxx instead!", "2018-06-01")
-    def set_tower_func(self, tower_func):
-        self._set_tower_func(tower_func)
-
     @property
     def tower_func(self):
         """