- fixes

mini_imagenet_dataloader.py

@@ -193,6 +193,22 @@ def one_hot(self, inp):
             out[idx, inp[idx]] = 1
         return out
 
+    def idx_to_big(self, phase, idx):
+        if phase=='train':
+            all_filenames = self.train_filenames
+#            labels = self.train_labels 
+        elif phase=='val':
+            all_filenames = self.val_filenames
+#            labels = self.val_labels 
+        elif phase=='test':
+            all_filenames = self.test_filenames
+#            labels = self.test_labels
+        else:
+            print('Please select vaild phase')
+
+        one_episode_sample_num = self.num_samples_per_class*self.shot_num
+        return ((idx+1)*one_episode_sample_num >= len(all_filenames))
+
     def get_batch(self, phase='train', idx=0):
         if phase=='train':
             all_filenames = self.train_filenames
@@ -286,6 +302,12 @@ def generate_add_samples(self, phase = 'train'):
             if phase == 'train':
                 if args.enable_idx_increase:
                     self.idx += 1 # only train phase allowed to change
+                    if dataloader.idx_to_big(args.dataset, self.idx):
+                        self.idx=0
+                        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+                        print("all data used, starting from beginning")
+                        print("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!")
+
                 #print(episode_train_img.shape[0])
                 #assert(episode_train_img.shape[0] == 25)
                 for i in range(episode_train_img.shape[0]):
@@ -417,6 +439,39 @@ def call(x):
 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
+# here we can try setting the layer not trainable
+def all_layers(model):
+    layers = []
+    for l in model.layers:
+        #print(l.name, l.trainable, isinstance(l,Model))
+        if isinstance(l, Model):
+            a = all_layers(l)
+            #print(a)
+            layers.extend(a)
+        else:
+            layers.append(l)
+    return layers
+
+for l in all_layers(siamese_net):
+    l2=l
+    if isinstance(l,TimeDistributed):
+        l2=l.layer
+    print(l2.name,l2.trainable, len(l2.get_weights()))
+
+for l in all_layers(lambda_model):
+    l2=l
+    p='normal'
+    if isinstance(l,TimeDistributed):
+        l2=l.layer
+        p='timedi'
+#    if (l2.name == 'dense_1'):
+#        l2.trainable = False
+    print(p,l2.name,l2.trainable, len(l2.get_weights()))
+
+#lambda_model.get_layer("dense_1").trainable = False
+
 # testing with additional batch axis ?!
 i=1
 test_lambda = lambda_model([K.expand_dims(K.variable(base_train_img[0:0+1]),axis=0),K.expand_dims(K.variable(base_train_img), axis=0), K.expand_dims(K.variable(base_train_label), axis=0)])
@@ -428,7 +483,8 @@ def call(x):
 checkpointer = ModelCheckpoint(filepath='checkpoints/model-{epoch:02d}.hdf5', verbose=1)
 tensorboard = TensorBoard()
 lambda_model.fit_generator(keras_gen_train.generate_add_samples(), train_epoch_size, args.epochs, 
-                           validation_data=keras_gen_train.generate_add_samples('test'), validation_steps=test_epoch_size, callbacks = [tensorboard]) 
+                           validation_data=keras_gen_train.generate_add_samples('test'), validation_steps=test_epoch_size, callbacks = [tensorboard], workers = 0) 
+#workers = 0 is a work around to correct the number of calls to the validation_data generator
 lambda_model.save(args.final_name+'.hdf5')