Enable and refactor image summaries (resolves #108)

neuralmonkey/decoders/decoder.py

@@ -116,14 +116,7 @@ def __init__(self, encoders, vocabulary, data_id, name, **kwargs):
 
         self.runtime_rnn_outputs, self.runtime_rnn_states, runtime_logits = \
             self._attention_decoder(
-                runtime_inputs, state, runtime_mode=True,
-                summary_collections=["summary_val_plots"])
-
-        val_plots_collection = tf.get_collection("summary_val_plots")
-        self.summary_val_plots = (
-            tf.merge_summary(val_plots_collection)
-            if val_plots_collection else None
-        )
+                runtime_inputs, state, runtime_mode=True)
 
         self.train_logprobs = [tf.nn.log_softmax(l) for l in train_logits]
         self.decoded = [tf.argmax(l[:, 1:], 1) + 1 for l in runtime_logits]
@@ -318,7 +311,7 @@ def _logit_function(self, rnn_output):
     #pylint: disable=too-many-arguments
     # TODO reduce the number of arguments
     def _attention_decoder(self, inputs, initial_state, runtime_mode=False,
-                           summary_collections=None, scope="attention_decoder"):
+                           scope="attention_decoder"):
         """Run the decoder RNN.
 
         Arguments:
@@ -327,8 +320,6 @@ def _attention_decoder(self, inputs, initial_state, runtime_mode=False,
             initial_state: The initial state of the decoder.
             runtime_mode: Boolean flag whether the decoder is running in
                           runtime mode (with loop function).
-            summary_collections: The list of summary collections to which
-                                 the alignments are logged.
             scope: The variable scope to use with this function.
         """
         cell = self._get_rnn_cell()
@@ -371,14 +362,14 @@ def _attention_decoder(self, inputs, initial_state, runtime_mode=False,
                 rnn_outputs.append(output)
                 rnn_states.append(state)
 
-            if summary_collections:
+            if runtime_mode:
                 for i, a in enumerate(att_objects):
                     attentions = a.attentions_in_time[-len(inputs):]
                     alignments = tf.expand_dims(tf.transpose(
                         tf.pack(attentions), perm=[1, 2, 0]), -1)
 
                     tf.image_summary("attention_{}".format(i), alignments,
-                                     collections=summary_collections,
+                                     collections=["summary_val_plots"],
                                      max_images=256)
 
         return rnn_outputs, rnn_states, output_logits

neuralmonkey/runners/runner.py

@@ -18,6 +18,9 @@ def __init__(self,
         super(GreedyRunner, self).__init__(output_series, decoder)
         self._postprocess = postprocess
 
+        self.image_summaries = tf.merge_summary(
+                tf.get_collection("summary_val_plots"))
+
     def get_executable(self, train=False, summaries=True):
         if train:
             fecthes = {"train_xent": self._decoder.train_loss,
@@ -26,6 +29,10 @@ def get_executable(self, train=False, summaries=True):
             fecthes = {"train_xent": tf.zeros([]),
                        "runtime_xent": tf.zeros([])}
         fecthes["decoded_logprobs"] = self._decoder.runtime_logprobs
+
+        if summaries:
+            fecthes['image_summaries'] = self.image_summaries
+
         return GreedyRunExecutable(self.all_coders, fecthes,
                                    self._decoder.vocabulary,
                                    self._postprocess)
@@ -69,10 +76,13 @@ def collect_results(self, results: List[Dict]) -> None:
         if self._postprocess is not None:
             decoded_tokens = [self._postprocess(seq) for seq in decoded_tokens]
 
+
+        image_summaries = results[0].get('image_summaries')
+
         self.result = ExecutionResult(
             outputs=decoded_tokens,
             losses=[train_loss, runtime_loss],
             scalar_summaries=None,
             histogram_summaries=None,
-            image_summaries=None
+            image_summaries=image_summaries
         )