Merge branch 'master' of https://github.com/ace19-dev/image-retrieval-tf

README.md

@@ -2,9 +2,9 @@
 
 
 ## TODO
-- [On progress] Multi GPU
-- TTA
-- [On progress] Balanced Batch
+- [testing...] Multi GPU
+- [testing...] TTA
+- [developing...] Balanced Batch
 
 ## References
 - https://github.com/ildoonet/tf-mobilenet-v2

train.py

@@ -121,6 +121,8 @@
 TRAIN_DATA_SIZE = 263+384+285+606+215+457+648+219+516+233+490+393   # 4709
 VALIDATE_DATA_SIZE = 46+68+50+107+38+81+114+38+91+41+86+70     # 830
 
+TEN_CROP = 10
+
 
 def show_batch_data(filenames, batch_x, batch_y, additional_path=None):
     default_path = '/home/ace19/Pictures/'
@@ -218,11 +220,13 @@ def main(unused_argv):
                 #                                                      is_training=is_training,
                 #                                                      keep_prob=keep_prob,
                 #                                                      attention_module='se_block')
-
+                logit = tf.cond(is_training,
+                                lambda: tf.identity(logit),
+                                lambda: tf.reduce_mean(tf.reshape(logit, [FLAGS.val_batch_size, TEN_CROP, -1]), axis=1))
                 logits.append(logit)
 
                 l = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=ground_truth,
-                                                                      logits=logit)
+                                                                   logits=logit)
                 losses.append(l)
                 loss_w_reg = tf.reduce_sum(l) + tf.add_n(slim.losses.get_regularization_losses(scope=scope_name))
 
@@ -401,16 +405,21 @@ def main(unused_argv):
                 sess.run(val_iterator.initializer, feed_dict={tfrecord_filenames: validate_record_filenames})
                 for step in range(val_batches):
                     filenames, validation_batch_xs, validation_batch_ys = sess.run(val_next_batch)
-                    # show_batch_data(filenames, validation_batch_xs, validation_batch_ys)
+                    # TTA
+                    batch_size, n_crops, c, h, w = validation_batch_xs.shape
+                    # fuse batch size and ncrops
+                    tencrop_val_batch_xs = np.reshape(validation_batch_xs, (-1, c, h, w))
+                    # show_batch_data(filenames, tencrop_val_batch_xs, validation_batch_ys)
 
-                    # augmented_val_batch_xs = aug_utils.aug(validation_batch_xs)
+                    # augmented_val_batch_xs = aug_utils.aug(tencrop_val_batch_xs)
                     # show_batch_data(filenames, augmented_val_batch_xs,
                     #                 validation_batch_ys, 'aug')
 
+                    # TODO: Verify TTA(TenCrop)
                     val_summary, val_loss, val_top1_acc, _confusion_matrix = sess.run(
                         [summary_op, loss, top1_acc, confusion_matrix],
                         feed_dict={
-                            X: validation_batch_xs,
+                            X: tencrop_val_batch_xs,
                             ground_truth: validation_batch_ys,
                             is_training: False,
                             keep_prob: 1.0

val_data.py

@@ -10,6 +10,8 @@
 MEAN=[0.485, 0.456, 0.406]
 STD=[0.229, 0.224, 0.225]
 
+RANDOM_CROP_SIZE = 200
+
 
 class Dataset(object):
     """
@@ -30,6 +32,7 @@ def __init__(self, tfrecord_path, batch_size, num_classes, num_epochs, data_size
         # of the self.dataset.
         self.dataset = self.dataset.map(self.decode, num_parallel_calls=8)
         # self.dataset = self.dataset.map(self.augment, num_parallel_calls=8)
+        self.dataset = self.dataset.map(self.tencrop, num_parallel_calls=8)
         self.dataset = self.dataset.map(self.normalize, num_parallel_calls=8)
 
         # Prefetches a batch at a time to smooth out the time taken to load input
@@ -80,6 +83,20 @@ def augment(self, filename, image, label):
         return filename, image, label
 
 
+    def tencrop(self, filename, image, label):
+        """Placeholder for TenCrop
+        horizontal flipping is used by default
+        """
+        images = []
+        for i in range(5):
+            img = tf.random_crop(image, [RANDOM_CROP_SIZE, RANDOM_CROP_SIZE, 3])
+            img = tf.image.resize(img, [self.resize_h, self.resize_w])
+            images.append(img)
+            images.append(tf.image.flip_left_right(img))
+
+        return filename, tf.stack(images), label
+
+
     def normalize(self, filename, image, label):
         # """Convert `image` from [0, 255] -> [-0.5, 0.5] floats."""
         # image = tf.cast(image, tf.float32) * (1. / 255) - 0.5