Merge pull request #692 from brettkoonce/minor_sp

tensor2tensor: minor spelling tweaks

Author: lukaszkaiser

tensor2tensor/bin/t2t_bleu.py

@@ -74,7 +74,7 @@
 flags.DEFINE_string("translation", None,
                     "Path to the MT system translation file")
 flags.DEFINE_string("translations_dir", None,
-                    "Directory with translated files to be evaulated.")
+                    "Directory with translated files to be evaluated.")
 flags.DEFINE_string("event_dir", None, "Where to store the event file.")
 
 flags.DEFINE_string("bleu_variant", "both",

tensor2tensor/data_generators/algorithmic_math.py

@@ -181,7 +181,7 @@ def algebra_inverse_solve(left, right, var, solve_ops):
           right- Expression on the right side of the op.
           to_tree- The tree on the other side of the equal sign. The canceled
               out expression will be moved here.
-          new_from_tree- The resuling from_tree after the algebraic
+          new_from_tree- The resulting from_tree after the algebraic
               manipulation.
           new_to_tree- The resulting to_tree after the algebraic manipulation.
 
@@ -355,7 +355,7 @@ def generate_calculus_integrate_sample(vlist, ops, min_depth, max_depth,
 # functions: Dict of special function names. Maps human readable string names to
 #     single char names used in flist.
 # ops: Dict mapping op symbols (chars) to ExprOp instances.
-# solve_ops: Encodes rules for how to algebraicly cancel out each operation. See
+# solve_ops: Encodes rules for how to algebraically cancel out each operation. See
 #     doc-string for `algebra_inverse_solve`.
 # int_encoder: Function that maps a string to a list of tokens. Use this to
 #     encode an expression to feed into a model.
@@ -377,7 +377,7 @@ def math_dataset_init(alphabet_size=26, digits=None, functions=None):
 
   Args:
     alphabet_size: How many possible variables there are. Max 52.
-    digits: How many numerical digits to encode as tokens, "0" throuh
+    digits: How many numerical digits to encode as tokens, "0" through
         str(digits-1), or None to encode no digits.
     functions: Defines special functions. A dict mapping human readable string
         names, like "log", "exp", "sin", "cos", etc., to single chars. Each

tensor2tensor/data_generators/speech_recognition.py

@@ -13,7 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-"""Common classes for automatic speech recogntion (ASR) datasets.
+"""Common classes for automatic speech recognition (ASR) datasets.
 
 The audio import uses sox to generate normalized waveforms, please install
 it as appropriate (e.g. using apt-get or yum).

tensor2tensor/data_generators/text_encoder.py

@@ -348,7 +348,7 @@ def store_to_file(self, filename):
 def _escape_token(token, alphabet):
   """Escape away underscores and OOV characters and append '_'.
 
-  This allows the token to be experessed as the concatenation of a list
+  This allows the token to be expressed as the concatenation of a list
   of subtokens from the vocabulary. The underscore acts as a sentinel
   which allows us to invertibly concatenate multiple such lists.
 

tensor2tensor/data_generators/translate_enzh.py

@@ -172,7 +172,7 @@ class TranslateEnzhWmt32k(translate.TranslateProblem):
 
   CWMT:
     - http://nlp.nju.edu.cn/cwmt-wmt/
-    - Website contrains instructions for FTP server access.
+    - Website contains instructions for FTP server access.
     - You'll need to download CASIA, CASICT, DATUM2015, DATUM2017,
         NEU datasets
 

tensor2tensor/data_generators/wiki.py

@@ -190,7 +190,7 @@ def scramble_fraction(self):
 
 @registry.register_problem
 class LanguagemodelWikiScrambleL1k(LanguagemodelWikiScramble):
-  """Sequence length 1024, 50% scrambed."""
+  """Sequence length 1024, 50% scrambled."""
 
   @property
   def sequence_length(self):

tensor2tensor/layers/common_attention.py

@@ -83,7 +83,7 @@ def register_layer(
       default_args (list): The default parameters to add to the function.
       default_kwargs (dict): The default parameters to add to the function.
         Those arguments can be overwritten when calling the function.
-      use_dp (bool): Wrap the function call within a dataparalellism object if
+      use_dp (bool): Wrap the function call within a dataparallelism object if
         dp is available. Some layers (like MOE) must be called without dp.
       recompute_grad (bool): If True, recompute the function during the
         backward pass to save memory
@@ -1378,7 +1378,7 @@ def _relative_attention_inner(x, y, z, transpose):
     x: Tensor with shape [batch_size, heads, length, length or depth].
     y: Tensor with shape [batch_size, heads, length, depth].
     z: Tensor with shape [length, length, depth].
-    transpose: Whether to tranpose inner matrices of y and z. Should be true if
+    transpose: Whether to transpose inner matrices of y and z. Should be true if
         last dimension of x is depth, not length.
 
   Returns:
@@ -1422,7 +1422,7 @@ def dot_product_attention_relative(q,
     k: a Tensor with shape [batch, heads, length, depth].
     v: a Tensor with shape [batch, heads, length, depth].
     bias: bias Tensor.
-    max_relative_position: an integer specifying the maxmimum distance between
+    max_relative_position: an integer specifying the maximum distance between
         inputs that unique position embeddings should be learned for.
     dropout_rate: a floating point number.
     image_shapes: optional tuple of integer scalars.
@@ -2141,7 +2141,7 @@ def gather_indices_2d(x, block_shape, block_stride):
 
 
 def make_2d_block_raster_mask(query_shape, memory_flange):
-  """creates a mask for 2d block raster scany.
+  """creates a mask for 2d block raster scan.
 
   The query mask can look to the left, top left, top, and top right, but
   not to the right. Inside the query, we have the standard raster scan
@@ -2661,7 +2661,7 @@ def ffn_self_attention_layer(x,
   We use self-attention to do feedforward computations. We apply this function
   positionwise where for each position, we linearly transform the output to have
   depth filter_depth, and break up the result depth-wise into num_parts
-  contiguous parts.  The parts self-attentd, we concatenate the results
+  contiguous parts.  The parts self-attend, we concatenate the results
   depth-wise, and we linearly transform to a depth of output_depth. The
   goal is to get multiplicative interactions between components of a
   representation.
@@ -2764,7 +2764,7 @@ def parameter_attention(x,
         x, total_key_depth, use_bias=False, name="q_transform")
     if dropout_rate:
       # This is a cheaper form of attention dropout where we use to use
-      # the same dropout decisions across batch elemets and query positions,
+      # the same dropout decisions across batch elements and query positions,
       # but different decisions across heads and memory positions.
       v = tf.nn.dropout(
           v, 1.0 - dropout_rate, noise_shape=[num_heads, memory_rows, 1])

tensor2tensor/layers/common_hparams.py

@@ -102,13 +102,13 @@ def basic_params1():
       moe_loss_coef=1e-2,
       # Sequences of operations to perform on layer input and layer output.
       # Used by common_layers.layer_preprocess, common_layers.layer_postprocess
-      # Each character repsesnts an operation:
+      # Each character represents an operation:
       # none: no preprocessing
       #    d: apply dropout
       #    n: apply normalization (see norm_type and norm_epsilon)
       #    a: add layer input (residual connection - only during postprocess)
       # The special string "none" is used instead of the empty string
-      # to indicate no pre/postprocesisng, since the empty string causes
+      # to indicate no pre/postprocessing, since the empty string causes
       # trouble for hyperparameter tuning.
       # TODO(noam): The current settings ("", "dan") are the published version
       # of the transformer.  ("n", "da") seems better for harder-to-learn
@@ -174,13 +174,13 @@ def basic_params1():
       # The maximum length of "input" sequence.
       # Sequences longer than this value will be truncated. 0 or negative values
       # mean there is no maximum or truncation.
-      # You can change this behavior by overridding preprocess_example() method
+      # You can change this behavior by overriding preprocess_example() method
       # in your problem class.
       max_input_seq_length=0,
       # The maximum length of "target" sequence.
       # Sequences longer than this value will be truncated. 0 or negative values
       # mean there is no maximum or truncation.
-      # You can change this behavior by overridding preprocess_example() method
+      # You can change this behavior by overriding preprocess_example() method
       # in your problem class.
       max_target_seq_length=0,
       # if nonzero, we split the target sequences on example read.

tensor2tensor/layers/common_layers.py

@@ -1201,7 +1201,7 @@ def add_timing_signal(x, min_timescale=1, max_timescale=1e4, num_timescales=16):
   and the target of the attention.
 
   The use of relative position is possible because sin(x+y) and cos(x+y) can be
-  experessed in terms of y, sin(x) and cos(x).
+  expressed in terms of y, sin(x) and cos(x).
 
   In particular, we use a geometric sequence of timescales starting with
   min_timescale and ending with max_timescale.  For each timescale, we

tensor2tensor/models/research/attention_lm.py

@@ -70,7 +70,7 @@ def attention_lm_prepare_decoder(targets, hparams):
   Returns:
     decoder_input: a Tensor, bottom of decoder stack
     decoder_self_attention_bias: a Tensor, containing large negative values
-    to implement masked attention and possibly baises for diagonal alignments
+    to implement masked attention and possibly biases for diagonal alignments
   """
   if hparams.prepend_mode == "prepend_inputs_full_attention":
     decoder_self_attention_bias = (