chkpt manip

README.md

@@ -24,10 +24,11 @@ You need to abstract your training task into three components:
 	+ Use Python to easily handle your own data format, yet still keep a good training speed thanks to multiprocess prefetch & TF Queue prefetch.
 	For example, InceptionV3 can run in the same speed as the official code which reads data using TF operators.
 
-3. The callbacks, including everything you want to do apart from the training iterations. Such as:
+3. Callbacks, including everything you want to do apart from the training iterations. Such as:
 	+ Change hyperparameters during training
 	+ Print some variables of interest
 	+ Run inference on a test dataset
+	+ Run some operations once a while
 
 With the above components defined, tensorpack trainer will run the training iterations for you.
 Multi-GPU training is ready to use by simply switching the trainer.

examples/DoReFa-Net/alexnet-dorefa.py

@@ -33,7 +33,9 @@ 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), 43.3% error.
+    With (W,A,G)=(32,32,32), 43% error.
+    With (W,A,G)=(1,2,6), 51% error.
+    With (W,A,G)=(1,2,4), 63% error.
 
 Speed:
     About 3.5 iteration/s on 4 Tesla M40. (Each epoch is set to 10000 iterations)

scripts/checkpoint-manipulate.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# File: checkpoint-manipulate.py
+# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
+
+
+from tensorpack.tfutils.varmanip import dump_chkpt_vars
+import tensorflow as tf
+import sys
+
+model_path = sys.argv[1]
+reader = tf.train.NewCheckpointReader(model_path)
+var_names = reader.get_variable_to_shape_map().keys()
+result = {}
+for n in var_names:
+    result[n] = reader.get_tensor(n)
+import IPython as IP; IP.embed(config=IP.terminal.ipapp.load_default_config())