Attention visualization

neuralmonkey/decoders/decoder.py

@@ -113,7 +113,10 @@ def __init__(self, encoders, vocabulary, data_id, **kwargs):
 
         self.runtime_rnn_outputs, _ = attention_decoder(
             runtime_inputs, state, attention_objects, cell,
-            attention_maxout_size, loop_function=loop_function)
+            attention_maxout_size, loop_function=loop_function,
+            summary_collections=["summary_val_img"])
+
+        self.summary_val_img = tf.merge_summary(tf.get_collection("summary_val_img"))
 
         _, train_logits = self._decode(self.train_rnn_outputs)
         self.decoded, runtime_logits = self._decode(self.runtime_rnn_outputs)

neuralmonkey/decoding_function.py

@@ -14,7 +14,8 @@
 # pylint: disable=too-many-arguments
 # Great functions require great number of parameters
 def attention_decoder(decoder_inputs, initial_state, attention_objects,
-                      cell, maxout_size, loop_function=None, scope=None):
+                      cell, maxout_size, loop_function=None, scope=None,
+                      summary_collections=None):
     outputs = []
     states = []
 
@@ -47,6 +48,12 @@ def attention_decoder(decoder_inputs, initial_state, attention_objects,
             outputs.append(output)
             states.append(state)
 
+        if summary_collections:
+            for i, a in enumerate(attention_objects):
+                alignments = tf.expand_dims(tf.transpose(tf.pack(a.attentions_in_time), perm=[1, 2, 0]), -1)
+                tf.image_summary("attention_{}".format(i), alignments,
+                                 collections=summary_collections, max_images=128)
+
     return outputs, states
 
 

neuralmonkey/learning_utils.py

@@ -223,9 +223,10 @@ def training_loop(sess, saver,
 
                 if step % validation_period == validation_period - 1:
                     decoded_val_sentences, decoded_raw_val_sentences, \
-                        val_evaluation = run_on_dataset(
+                        val_evaluation, val_images = run_on_dataset(
                             sess, runner, all_coders, decoder, val_dataset,
-                            evaluators, postprocess, write_out=False)
+                            evaluators, postprocess, write_out=False,
+                            extra_fetches=decoder.summary_val_img)
 
                     this_score = val_evaluation[evaluators[-1].name]
 
@@ -309,6 +310,8 @@ def argworst(scores, minimize):
 
                     log_print("")
 
+                    tb_writer.add_summary(val_images[0], step)
+
     except KeyboardInterrupt:
         log("Training interrupted by user.")
 
@@ -329,7 +332,8 @@ def argworst(scores, minimize):
 
 
 def run_on_dataset(sess, runner, all_coders, decoder, dataset,
-                   evaluators, postprocess, write_out=False):
+                   evaluators, postprocess, write_out=False,
+                   extra_fetches=None):
     """
     Applies the model on a dataset and optionally writes outpus into a file.
 
@@ -353,13 +357,17 @@ def run_on_dataset(sess, runner, all_coders, decoder, dataset,
         write_out: Flag whether the outputs should be printed to a file defined
             in the dataset object.
 
+        extra_fetches: Extra tensors to evaluate for each batch.
+
     Returns:
 
         Tuple of resulting sentences/numpy arrays, and evaluation results if
             they are available which are dictionary function -> value.
 
     """
-    result_raw, opt_loss, dec_loss = runner(sess, dataset, all_coders)
+    result_raw, opt_loss, dec_loss, evaluated_fetches = \
+        runner(sess, dataset, all_coders, extra_fetches)
+
     if postprocess is not None:
         result = postprocess(result_raw)
     else:
@@ -387,7 +395,10 @@ def run_on_dataset(sess, runner, all_coders, decoder, dataset,
         for func in evaluators:
             evaluation[func.name] = func(result, test_targets)
 
-    return result, result_raw, evaluation
+    if extra_fetches is not None:
+        return result, result_raw, evaluation, evaluated_fetches
+    else:
+        return result, result_raw, evaluation
 
 
 def process_evaluation(evaluators, tb_writer, eval_result,

neuralmonkey/runners/runner.py

@@ -10,9 +10,13 @@ def __init__(self, decoder, batch_size):
         self.batch_size = batch_size
         self.vocabulary = decoder.vocabulary
 
-    def __call__(self, sess, dataset, coders):
+    def __call__(self, sess, dataset, coders, extra_fetches=None):
         batched_dataset = dataset.batch_dataset(self.batch_size)
         decoded_sentences = []
+        evaluated_fetches = []
+
+        if extra_fetches is None:
+            extra_fetches = []
 
         loss_with_gt_ins = 0.0
         loss_with_decoded_ins = 0.0
@@ -29,14 +33,16 @@ def __call__(self, sess, dataset, coders):
             else:
                 losses = [tf.zeros([]), tf.zeros([])]
 
-            computation = sess.run(losses + self.decoder.decoded,
-                                   feed_dict=batch_feed_dict)
-            loss_with_gt_ins += computation[0]
-            loss_with_decoded_ins += computation[1]
+            (loss_with_gt_ins, loss_with_decoded_ins), \
+                    decoded, fetches_batch = \
+                    sess.run((losses, self.decoder.decoded, extra_fetches),
+                             feed_dict=batch_feed_dict)
             decoded_sentences_batch = \
-                    self.vocabulary.vectors_to_sentences(computation[len(losses):])
+                    self.vocabulary.vectors_to_sentences(decoded)
             decoded_sentences += decoded_sentences_batch
+            evaluated_fetches += [fetches_batch]
 
         return decoded_sentences, \
                loss_with_gt_ins / batch_count, \
-               loss_with_decoded_ins / batch_count
+               loss_with_decoded_ins / batch_count, \
+               evaluated_fetches

tests/small.ini

@@ -29,7 +29,7 @@ name=translation
 output=tests/tmp-test-output
 overwrite_output_dir=True
 batch_size=16
-epochs=2
+epochs=5
 encoders=[<encoder>]
 decoder=<decoder>
 train_dataset=<train_data>