Ajk4 · August 25, 2017 10:39 · Aug 25, 2017 · Aug 25, 2017 · Aug 25, 2017 · Aug 25, 2017
diff --git a/main.py b/main.py
@@ -80,7 +80,7 @@ def change_learning_rate(new_lr):
     #    K.set_value(model.optimizer.lr, float(new_lr))
     #
 
-    # It will blow up, because set_values will try to create Assign Op using different graph
+    # It will blow up, because set_value will try to create Assign Op using different graph
     K.set_value(model.optimizer.lr, float(new_lr))
     print 'Changed learning rate to {}'.format(new_lr)
 

diff --git a/main.py b/main.py
@@ -79,6 +79,8 @@ def change_learning_rate(new_lr):
     # with graph.as_default():
     #    K.set_value(model.optimizer.lr, float(new_lr))
     #
+
+    # It will blow up, because set_values will try to create Assign Op using different graph
     K.set_value(model.optimizer.lr, float(new_lr))
     print 'Changed learning rate to {}'.format(new_lr)
 

diff --git a/main.py b/main.py
@@ -77,6 +77,8 @@ def change_learning_rate(new_lr):
     # set it here as default:
     #
     # with graph.as_default():
+    #    K.set_value(model.optimizer.lr, float(new_lr))
+    #
     K.set_value(model.optimizer.lr, float(new_lr))
     print 'Changed learning rate to {}'.format(new_lr)
 

diff --git a/main.py b/main.py
@@ -38,8 +38,6 @@
 
 letters = "ABCDEFGHIJ"
 
-graph = tf.get_default_graph()
-
 def array_2d_to_image(array, autorescale=True):
     assert array.min() >= 0
     assert len(array.shape) == 2
@@ -70,6 +68,8 @@ def array_2d_to_image(array, autorescale=True):
               optimizer='adam',
               metrics=['accuracy'])
 
+graph = tf.get_default_graph()
+
 def change_learning_rate(new_lr):
     time.sleep(5)
 

diff --git a/main.py b/main.py
@@ -0,0 +1,98 @@
+from scipy import io
+import numpy as np
+from keras.utils import np_utils
+from sklearn.model_selection import train_test_split
+from PIL import Image
+from sklearn.metrics import confusion_matrix
+import tensorflow as tf
+from threading import Thread
+
+from keras.models import Sequential
+from keras.layers import Dense, Activation, Flatten
+from keras.layers import Conv2D, MaxPool2D, Dropout
+from keras.callbacks import Callback, TensorBoard
+import keras.backend as K
+import time
+
+
+X = np.zeros((1000, 28, 28))
+Y = np.zeros((1000))
+
+resolution = 28  # 28x28 images, grayscale
+classes = 10     # 10 letters: ABCDEFGHIJ
+
+# transforming data for TensorFlow backend
+X = np.transpose(X, (2, 0, 1))
+X = X.reshape((-1, resolution, resolution, 1))
+X = X.astype('float32') / 255.
+
+# 3 -> [0., 0., 0., 1., 0., 0., 0., 0., 0., 0.]
+y = Y.astype('int32')
+Y = np_utils.to_categorical(y, classes)
+
+# splitting data into training and test sets
+X_train, X_test, Y_train, Y_test = train_test_split(
+    X, Y,
+    test_size=0.20,
+    random_state=137)
+
+letters = "ABCDEFGHIJ"
+
+graph = tf.get_default_graph()
+
+def array_2d_to_image(array, autorescale=True):
+    assert array.min() >= 0
+    assert len(array.shape) == 2
+    if array.max() <= 1 and autorescale:
+        array = 255 * array
+    array = array.astype('uint8')
+    return Image.fromarray(array)
+
+
+
+
+model = Sequential()
+
+model.add(Conv2D(16, (3, 3), activation='relu',
+                 input_shape=(resolution, resolution, 1), name='block1_conv1'))
+model.add(Conv2D(16, (3, 3), activation='relu', name='block1_conv2')) 
+model.add(MaxPool2D(name='block1_maxpool'))
+
+model.add(Conv2D(32, (3, 3), activation='relu', name='block2_conv1')) 
+model.add(Conv2D(32, (3, 3), activation='relu', name='block2_conv2')) 
+model.add(MaxPool2D(name='block2_maxpool'))
+
+model.add(Flatten(name='flatten'))
+model.add(Dropout(0.5, name='dropout'))
+model.add(Dense(classes, activation='softmax', name='output'))
+
+model.compile(loss='categorical_crossentropy',
+              optimizer='adam',
+              metrics=['accuracy'])
+
+def change_learning_rate(new_lr):
+    time.sleep(5)
+
+    # Need to explicitely store reference to graph from main thread and
+    # set it here as default:
+    #
+    # with graph.as_default():
+    K.set_value(model.optimizer.lr, float(new_lr))
+    print 'Changed learning rate to {}'.format(new_lr)
+
+
+# K.set_value lazily creates TF AssignOp and saves it for later. So if I would do this
+#
+# K.set_value(model.optimizer.lr, K.get_value(model.optimizer.lr))
+#
+# It would work, because other thread would use already created operation instead of trying to create new operation.
+
+thread = Thread(target=change_learning_rate, args=(10,))
+thread.start()
+
+model.fit(X_train, Y_train,
+          epochs=1000,
+          batch_size=32,
+          validation_data=(X_test, Y_test),
+          verbose=2)
+