Add comments to utils

codebase/utils.py

@@ -6,37 +6,60 @@
 import numpy as np
 
 def u2t(x):
+ """Convert uint8 to [-1, 1] float
+ """
  return x.astype('float32') / 255 * 2 - 1
 
 def s2t(x):
+ """Convert [0, 1] float to [-1, 1] float
+ """
  return x * 2 - 1
 
 def delete_existing(path):
+ """Delete directory if it exists
+
+ Used for automatically rewrites existing log directories
+ """
  if args.run < 999:
  assert not os.path.exists(path), "Cannot overwrite {:s}".format(path)
 
  else:
  if os.path.exists(path):
  shutil.rmtree(path)
 
-def save_accuracy(M, fn_acc_key, tag, dataloader,
+def save_accuracy(M, fn_acc_key, tag, data,
  train_writer=None, step=None, print_list=None,
  full=True):
+ """Log the accuracy of model on data to tf.summary.FileWriter
+
+ M - (TensorDict) the model
+ fn_acc_key - (str) key to query the relevant function in M
+ tag - (str) summary tag for FileWriter
+ data - (Data) data object with images/labels attributes
+ train_writer - (FileWriter)
+ step - (int) global step in file writer
+ print_list - (list) list of vals to print to stdout
+ full - (bool) use full dataset v. first 1000 samples
+ """
  fn_acc = getattr(M, fn_acc_key, None)
  if fn_acc:
- acc, summary = exact_accuracy(fn_acc, tag, dataloader, full)
+ acc, summary = compute_accuracy(fn_acc, tag, data, full)
  train_writer.add_summary(summary, step + 1)
  print_list += [os.path.basename(tag), acc]
 
-def exact_accuracy(fn_acc, tag, dataloader, full=True):
- # Fixed shuffling scheme
- state = np.random.get_state()
- np.random.seed(0)
- shuffle = np.random.permutation(len(dataloader.images))
- np.random.set_state(state)
+def compute_accuracy(fn_acc, tag, data, full=True):
+ """Compute accuracy w.r.t. data
+
+ fn_acc - (fn) Takes (x, y) as input and returns accuracy
+ tag - (str) summary tag for FileWriter
+ data - (Data) data object with images/labels attributes
+ full - (bool) use full dataset v. first 1000 samples
+ """
+ with tb.nputils.FixedSeed(0):
+ shuffle = np.random.permutation(len(data.images))
 
- xs = dataloader.images[shuffle]
- ys = dataloader.labels[shuffle] if dataloader.labels is not None else None
+ xs = data.images[shuffle]
+ ys = data.labels[shuffle] if data.labels is not None else None
 
  if not full:
  xs = xs[:1000]
@@ -47,9 +70,9 @@ def exact_accuracy(fn_acc, tag, dataloader, full=True):
  bs = 200
 
  for i in xrange(0, n, bs):
- x = u2t(xs[i:i+bs]) if dataloader.cast else xs[i:i+bs]
- y = ys[i:i+bs] if ys is not None else dataloader.labeler(x)
- acc += fn_acc(x, y) * len(x) / n
+ x = data.preprocess(xs[i:i+bs])
+ y = ys[i:i+bs] if ys is not None else data.labeler(x)
+ acc += fn_acc(x, y) / n * len(x)
 
  summary = tf.Summary.Value(tag=tag, simple_value=acc)
  summary = tf.Summary(value=[summary])