Adding Explicit Zero Padding for Conv Ops - Keras/CuDNN inconsistency

llvm/projects/keras/src/mobilenet_cifar10.py

+
 import sys
 import os
 os.environ['CUDA_VISIBLE_DEVICES'] = '1'
@@ -11,7 +12,7 @@ from keras.preprocessing.image import ImageDataGenerator
 from keras.models import Model
 from keras import optimizers
 import keras.backend as K
-#from frontend.approxhpvm_translator import translate_to_approxhpvm
+from frontend.approxhpvm_translator import translate_to_approxhpvm
 
 
 K.set_image_data_format('channels_first')
@@ -22,13 +23,14 @@ test_labels = y_test
 print ("X_train.shape = ", X_train.shape)
 print ("X_test.shape = ", X_test.shape)
 
-#sys.exit(0)
 
 X_train = X_train.astype('float32')
 X_test = X_test.astype('float32')
 
-mean = np.mean(X_train, axis=(0, 1, 2), keepdims=True)
-std = np.std(X_train, axis=(0, 1, 2), keepdims=True)
+
+mean = np.mean(X_train, axis=(0, 2, 3), keepdims=True)
+std = np.std(X_train, axis=(0, 2, 3), keepdims=True)
+
 X_train = (X_train - mean) / (std + 1e-9)
 X_test = (X_test - mean) / (std + 1e-9)
 
@@ -40,8 +42,8 @@ def get_mobilenet(alpha=1, depth_multiplier=1):
     model = Sequential()
     
     def _conv_block(filters, alpha, kernel=(3, 3), strides=(1, 1)):
-        channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
-        filters = int(filters * alpha)
+        channel_axis = 1
+        
         model.add(Conv2D(filters, kernel,
                           padding='same',
                           use_bias=False,
@@ -51,23 +53,23 @@ def get_mobilenet(alpha=1, depth_multiplier=1):
         model.add(Activation('relu'))
     
     def _depthwise_conv_block(pointwise_conv_filters, alpha, depth_multiplier=1, strides=(1, 1)):
-        channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
-        pointwise_conv_filters = int(pointwise_conv_filters * alpha)
+        channel_axis = 1 
+
+        model.add(ZeroPadding2D(padding = ((1,1), (1,1) )))
 
         model.add(DepthwiseConv2D((3, 3),
-                                   padding='same',
-                                   depth_multiplier=depth_multiplier,
+                                   padding='valid',
+                                   #depth_multiplier=depth_multiplier,
                                    strides=strides,
                                    use_bias=False))    
         model.add(BatchNormalization(axis=channel_axis))
-        #model.add(ReLU())
+        
         model.add(Activation('relu'))
         model.add(Conv2D(pointwise_conv_filters, (1, 1),
                           padding='same',
                           use_bias=False,
                           strides=(1, 1)))
         model.add(BatchNormalization(axis=channel_axis))
-        #model.add(ReLU())
         model.add(Activation('relu'))
 
 
@@ -149,10 +151,11 @@ model.fit_generator(
     steps_per_epoch=int(np.ceil(50000 / 128)),
     validation_data=(X_test, y_test),
     #epochs=300,
-    epochs=30,
+    epochs=50,
     callbacks=callbacks
 )
 
+model.summary()
 
-#translate_to_approxhpvm(model, "data/mobilenet_hpvm/", X_test, test_labels, 10)
+translate_to_approxhpvm(model, "data/mobilenet_hpvm/", X_test, test_labels, 10)