bug fix in summary. fix TF comptability break. update PTB readme

README.md

@@ -58,7 +58,7 @@ The components are designed to be independent. You can use Model or DataFlow in
 + other requirements:
 ```
 pip install --user -r requirements.txt
-pip install --user -r opt-requirements.txt # (some optional dependencies, you can install later if needed)
+pip install --user -r opt-requirements.txt # (some optional dependencies required by certain submodule, you can install later if needed)
 ```
 + Enable `import tensorpack`:
 ```

examples/DoReFa-Net/alexnet-dorefa.py

@@ -35,7 +35,8 @@ Accuracy:
     BATCH_SIZE * NUM_GPU. With a different number of GPUs in use, things might
     be a bit different, especially for learning rate.
 
-    With (W,A,G)=(32,32,32), 43% error.
+    With (W,A,G)=(32,32,32) -- full precision baseline, 43% error.
+    With (W,A,G)=(1,32,32) -- BWN, 46% error.
     With (W,A,G)=(1,2,6), 51% error.
     With (W,A,G)=(1,2,4), 63% error.
 

examples/PennTreebank/READER.md → examples/PennTreebank/README.md

@@ -2,7 +2,18 @@
 # LSTM language modeling on Penn Treebank dataset
 
 This example is mainly to demonstrate:
+
 1. How to train an RNN with persistent state between iterations.
 2. How to use a TF reader pipeline instead of a DataFlow for training & inference.
 
-More information to come.
+It trains an language model on PTB dataset, basically an equivalent of the PTB example
+in [tensorflow/models](https://github.com/tensorflow/models/tree/master/tutorials/rnn/ptb).
+It has the same performance & speed as the original example as well.
+Note that the data pipeline is completely copied from the tensorflow example.
+
+To Train:
+```
+./PTB-LSTM.py
+```
+
+

tensorpack/models/batch_norm.py

@@ -84,7 +84,7 @@ def BatchNormV1(x, use_local_stat=None, decay=0.9, epsilon=1e-5):
 
     if use_local_stat:
         batch = tf.cast(tf.shape(x)[0], tf.float32)
-        mul = tf.select(tf.equal(batch, 1.0), 1.0, batch / (batch - 1))
+        mul = tf.where(tf.equal(batch, 1.0), 1.0, batch / (batch - 1))
         batch_var = batch_var * mul  # use unbiased variance estimator in training
 
         with tf.control_dependencies([ema_apply_op] if ctx.is_training else []):

tensorpack/tfutils/summary.py

@@ -62,7 +62,7 @@ def add_param_summary(*summary_lists):
     ctx = get_current_tower_context()
     if ctx is not None and not ctx.is_main_training_tower:
         return
-    if len(summary_lists) == 0 and isinstance(summary_lists[0], list):
+    if len(summary_lists) == 1 and isinstance(summary_lists[0], list):
         logger.warn("[Deprecated] Use positional args to call add_param_summary() instead of a list.")
         summary_lists = summary_lists[0]
 

tensorpack/tfutils/symbolic_functions.py

@@ -121,7 +121,7 @@ def huber_loss(x, delta=1, name='huber_loss'):
     """
     sqrcost = tf.square(x)
     abscost = tf.abs(x)
-    return tf.select(abscost < delta,
+    return tf.where(abscost < delta,
                     sqrcost * 0.5,
                     abscost * delta - 0.5 * delta ** 2,
                     name=name)