added preprocess for imagenet, combined name of model and init weights

src/predict_model.py

@@ -35,7 +35,7 @@ def call(self, inputs, training=False):
     parser.add_argument('-d', '--data_info', help='path to dataframe of filepaths')
     parser.add_argument('-m', '--model_dir', help='input directory')
     parser.add_argument('-o', '--output_dir', help='output directory')
-    parser.add_argument('-p', '--parameters', help='list of training parameters')
+    parser.add_argument('-p', '--parameters', help='list of prediction parameters')
 
     args = parser.parse_args()
     data_parameters = DataAugmentationParams(**json.loads(args.parameters))
@@ -44,20 +44,28 @@ def call(self, inputs, training=False):
 
     # Load trained model and parameters
     loaded_model = load_model(args.model_dir+'/model.keras')
-    target_size = model_list[loaded_model._name]
+    model_name, *weights = loaded_model._name.split('_')
+    target_size = model_list[model_name]
     custom_model = CustomModel(loaded_model)    # Modify trained model to return prob and f_vec
 
     # Prepare data generators and create a tf.data pipeline of augmented images
-    test_dataset, datasets_uris, data_type = get_dataset(args.data_info, shuffle=False)
+    predict_dataset, datasets_uris, data_type = get_dataset(args.data_info, shuffle=False)
+    predict_generator = predict_dataset.map(lambda x: data_preprocessing(x, (target_size,target_size), data_type, data_parameters.log))
+
+    # Preprocess input according to the model if weights are set to imagenet
+    if weights == ['imagenet']:
+        preprocess_input_description = f"tf.keras.applications.{model_name}.preprocess_input"
+        preprocess_input = compile(preprocess_input_description, "<string>", 'eval')
+        predict_generator = predict_dataset.batch(batch_size).map(lambda x: (preprocess_input(x)))
+    else:
+        predict_generator = predict_generator.batch(batch_size)
 
-    test_generator = test_dataset.map(lambda x: data_preprocessing(x, (target_size,target_size), data_type, data_parameters.log))
-    test_generator = test_generator.batch(batch_size)
     with open(args.model_dir+'/class_info.json', 'r') as json_file:
         classes = json.load(json_file)
     class_num = len(classes)
 
     # Start prediction process    
-    prob, f_vec = custom_model.predict(test_generator,
+    prob, f_vec = custom_model.predict(predict_generator,
                                        verbose=0,
                                        callbacks=[PredictionCustomCallback(datasets_uris, classes)])
 

src/train_model.py

@@ -70,8 +70,16 @@
     train_generator = train_dataset.map(lambda x, y: (data_preprocessing(x, (target_size,target_size), data_type, data_parameters.log), y))
     val_generator = val_dataset.map(lambda x, y: (data_preprocessing(x, (target_size,target_size), data_type, data_parameters.log), y))
 
-    train_generator = train_generator.batch(batch_size).map(lambda x, y: (data_augmentation(x), y))
-    val_generator = val_generator.batch(batch_size).map(lambda x, y: (data_augmentation(x), y))
+    # Preprocess input according to the model if weights are set to imagenet
+    if weights == 'imagenet':
+        preprocess_input_description = f"tf.keras.applications.{nn_model}.preprocess_input"
+        preprocess_input = compile(preprocess_input_description, "<string>", 'eval')
+        train_generator = train_generator.batch(batch_size).map(lambda x, y: (preprocess_input(data_augmentation(x)), y))
+        val_generator = val_generator.batch(batch_size).map(lambda x, y: (preprocess_input(data_augmentation(x)), y))
+    else:
+        train_generator = train_generator.batch(batch_size).map(lambda x, y: (data_augmentation(x), y))
+        val_generator = val_generator.batch(batch_size).map(lambda x, y: (data_augmentation(x), y))
+
     class_num = len(classes)
 
     # Define optimizer
@@ -93,7 +101,7 @@
         predictions = layers.Dense(class_num, activation='softmax', name="predictions")(x)
         model = tf.keras.models.Model(inputs=base_model.input,
                                       outputs=predictions,
-                                      name=base_model._name)
+                                      name=f"{base_model._name}_{weights}")
     else:
         model_description = f"tf.keras.applications.{nn_model}(include_top=True, weights=None, \
                              input_tensor=None, classes={class_num})"