- support of disabling layers for training

mini_imagenet_dataloader.py

@@ -19,7 +19,7 @@
 import numpy as np
 from tqdm import trange
 import imageio
-
+import ast
 
 import argparse
 parser = argparse.ArgumentParser(description='train recurrent net.')
@@ -33,6 +33,9 @@
 parser.add_argument('--use_independent_base', dest='use_independent_base', action='store_true')
 parser.add_argument('--train_indep_and_dependent', dest='train_indep_and_dependent', action='store_true')
 parser.add_argument('--tensorboard_log_dir', dest='tensorboard_log_dir',  type=str, default='./logs')
+parser.add_argument('--enable_only_layers_of_list', dest='enable_only_layers_of_list',  type=str, default=None)
+parser.add_argument('--episode_test_sample_num', dest='episode_test_sample_num',  type=int, default=15)
+
 args = parser.parse_args()
 
 class MiniImageNetDataLoader(object):
@@ -249,7 +252,7 @@ def get_batch(self, phase='train', idx=0):
         return this_inputa, this_labela, this_inputb, this_labelb
 
 cathegories = 5
-dataloader = MiniImageNetDataLoader(shot_num=5 * 2, way_num=cathegories, episode_test_sample_num=15) #twice shot_num is because one might be uses as the base for the samples
+dataloader = MiniImageNetDataLoader(shot_num=5 * 2, way_num=cathegories, episode_test_sample_num=args.episode_test_sample_num) #twice shot_num is because one might be uses as the base for the samples
 
 dataloader.generate_data_list(phase='train')
 dataloader.generate_data_list(phase='val')
@@ -452,9 +455,6 @@ def call(x):
     lambda_model = load_model(args.pretrained_name, custom_objects = { "keras": tensorflow.keras , "args":args})
     print("loaded model",lambda_model)
 
-#after loading to set learning rate
-lambda_model.compile(loss='categorical_crossentropy', optimizer=op.SGD(args.lr), metrics=['categorical_accuracy'])
-print(lambda_model.summary(line_length=180, positions = [.33, .55, .67, 1.]))
 
 
 # models in models forget the layer name, therefore one must use the automatically given layer name and iterate throught the models by hand
@@ -477,16 +477,34 @@ def all_layers(model):
         l2=l.layer
     print(l2.name,l2.trainable, len(l2.get_weights()))
 
-for l in all_layers(lambda_model):
-    l2=l
+lambda_model_layers = all_layers(lambda_model)
+for l in range(len(lambda_model_layers)):
+    l2=lambda_model_layers[l]
     p='normal'
-    if isinstance(l,TimeDistributed):
-        l2=l.layer
+    if isinstance(l2,TimeDistributed):
+        l2=l2.layer
         p='timedi'
-#    if (l2.name == 'dense_1'):
-#        l2.trainable = False
-    print(p,l2.name,l2.trainable, len(l2.get_weights()))
-
+    if args.enable_only_layers_of_list is not None:
+        l2.trainable = False
+    print(l, p,l2.name,l2.trainable, len(l2.get_weights()))
+
+if args.enable_only_layers_of_list is not None:
+    print('\nenable some layers for training')
+
+    for i in ast.literal_eval(args.enable_only_layers_of_list):
+        lambda_model_layers[i].trainable = True
+
+    for l in range(len(lambda_model_layers)):
+        l2=lambda_model_layers[l]
+        p='normal'
+        if isinstance(l2,TimeDistributed):
+            l2=l2.layer
+            p='timedi'
+        print(l, p,l2.name,l2.trainable, len(l2.get_weights()))
+
+#after loading to set learning rate
+lambda_model.compile(loss='categorical_crossentropy', optimizer=op.SGD(args.lr), metrics=['categorical_accuracy'])
+print(lambda_model.summary(line_length=180, positions = [.33, .55, .67, 1.]))
 #lambda_model.get_layer("dense_1").trainable = False
 
 # testing with additional batch axis ?!