vqa_mde y/n fix

configs/vqa_med_2019/c2_classification/c2_word_answer_onehot_bow.yml

@@ -5,6 +5,8 @@ default_configs: vqa_med_2019/c2_classification/default_c2_classification.yml
 training:
   problem:
     batch_size: 128
+  terminal_conditions:
+    episode_limit: 1000
 
 # Validation parameters:
 validation:
@@ -49,7 +51,7 @@ pipeline:
   # Model.
   classifier:
     type: FeedForwardNetwork 
-    hidden_sizes: [500, 500]
+    hidden_sizes: [50]
     dropout_rate: 0.5
     priority: 3
     streams:

configs/vqa_med_2019/c4_classification/c4_word_answer_onehot_bow.yml

@@ -4,8 +4,7 @@ default_configs: vqa_med_2019/c4_classification/default_c4_classification.yml
 # Training parameters:
 training:
   problem:
-    batch_size: 128
-
+    batch_size: 128    
 # Validation parameters:
 validation:
   problem:

ptp/components/problems/image_text_to_class/vqa_med_2019.py

@@ -448,7 +448,7 @@ def __getitem__(self, index):
         image_transformations_list = []
         # Optional.
         if 'random_affine' in self.image_preprocessing:
-            rotate = (-45, 135)
+            rotate = (-45, 80)
             translate = (0.05, 0.25)
             scale = (0.5, 2)
             image_transformations_list.append(transforms.RandomAffine(rotate, translate, scale))
@@ -511,9 +511,10 @@ def predict_yes_no(self, qtext):
         Determines whether this is binary (yes/no) type of question.
         """
         yes_no_starters = ['is','was','are','does']
-        tokens = qtext.split(' ')
-        first_token = tokens[0]
-        if first_token in yes_no_starters and ('or' not in tokens):
+        if 'tokenize' not in self.question_preprocessing:
+            qtext = qtext.split(' ')
+        first_token = qtext[0]
+        if first_token in yes_no_starters and ('or' not in qtext):
             return True
         return False