update

retrieval/retrieve.py

@@ -28,7 +28,7 @@
                     '/home/ace19/dl_results/image_retrieve/_result',
                     'Where the dataset reside.')
 
-flags.DEFINE_string('pre_trained_checkpoint',
+flags.DEFINE_string('checkpoint_dir',
                     '../tfmodels/best.ckpt-5',
                     'Directory where to read training checkpoints.')
 flags.DEFINE_string('checkpoint_exclude_scopes',
@@ -43,13 +43,6 @@
 flags.DEFINE_string('model_name',
                     'resnet_v2_50',
                     'The name of the architecture to train.')
-# flags.DEFINE_string('extra_model_name',
-#                     'fc1',
-#                     # None,
-#                     'The name of the architecture to extra train.')
-# flags.DEFINE_string('checkpoint_model_scope2',
-#                     'tower0/fc1',
-#                     'Model scope in the checkpoint. None if the same as the trained model.')
 
 flags.DEFINE_integer('batch_size', 64, 'batch size')
 flags.DEFINE_integer('height', 224, 'height')
@@ -211,15 +204,14 @@ def main(unused_argv):
     with tf.compat.v1.Session(config=sess_config) as sess:
         sess.run(tf.global_variables_initializer())
 
-        # TODO: supports multi gpu - add scope ('tower%d' % gpu_idx)
-        if FLAGS.pre_trained_checkpoint:
+        if FLAGS.checkpoint_dir:
             train_utils.custom_restore_fn(FLAGS)
 
-        # if FLAGS.pre_trained_checkpoint:
-        #     if tf.gfile.IsDirectory(FLAGS.pre_trained_checkpoint):
-        #         checkpoint_path = tf.train.latest_checkpoint(FLAGS.pre_trained_checkpoint)
+        # if FLAGS.checkpoint_dir:
+        #     if tf.gfile.IsDirectory(FLAGS.checkpoint_dir):
+        #         checkpoint_path = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
         #     else:
-        #         checkpoint_path = FLAGS.pre_trained_checkpoint
+        #         checkpoint_path = FLAGS.checkpoint_dir
         #     saver.restore(sess, checkpoint_path)
 
         # global_step = checkpoint_path.split('/')[-1].split('-')[-1]

train.py

@@ -210,12 +210,12 @@ def main(unused_argv):
                                                      keep_prob=keep_prob,
                                                      attention_module='se_block')
 
-                # Print name and shape of parameter nodes  (values not yet initialized)
-                tf.compat.v1.logging.info("++++++++++++++++++++++++++++++++++")
-                tf.compat.v1.logging.info("Parameters")
-                tf.compat.v1.logging.info("++++++++++++++++++++++++++++++++++")
-                for v in slim.get_model_variables():
-                    tf.compat.v1.logging.info('name = %s, shape = %s' % (v.name, v.get_shape()))
+                # # Print name and shape of parameter nodes  (values not yet initialized)
+                # tf.compat.v1.logging.info("++++++++++++++++++++++++++++++++++")
+                # tf.compat.v1.logging.info("Parameters")
+                # tf.compat.v1.logging.info("++++++++++++++++++++++++++++++++++")
+                # for v in slim.get_model_variables():
+                #     tf.compat.v1.logging.info('name = %s, shape = %s' % (v.name, v.get_shape()))
 
                 # # TTA
                 # logit = tf.cond(is_training,
@@ -333,7 +333,7 @@ def main(unused_argv):
             train_writer = tf.compat.v1.summary.FileWriter(FLAGS.summaries_dir, graph)
             validation_writer = tf.compat.v1.summary.FileWriter(FLAGS.summaries_dir + '/validation', graph)
 
-            # TODO: supports multi gpu - add scope ('tower%d' % gpu_idx)
+            # TODO: supports multi gpu -> add scope ('tower%d' % gpu_idx)
             if FLAGS.pre_trained_checkpoint:
                 train_utils.restore_fn(FLAGS)
 

train_data.py

@@ -94,14 +94,14 @@ def distort_image(self, filename, image, label):
         color_ordering = random.randint(0, 1)
         if color_ordering == 0:
             image = tf.image.random_brightness(image, max_delta=32. / 255.)
-            image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
-            image = tf.image.random_hue(image, max_delta=0.2)
-            image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
+            image = tf.image.random_saturation(image, lower=0.2, upper=1.2)
+            image = tf.image.random_hue(image, max_delta=0.1)
+            image = tf.image.random_contrast(image, lower=0.2, upper=1.2)
         elif color_ordering == 1:
             image = tf.image.random_brightness(image, max_delta=32. / 255.)
-            image = tf.image.random_contrast(image, lower=0.5, upper=1.5)
-            image = tf.image.random_saturation(image, lower=0.5, upper=1.5)
-            image = tf.image.random_hue(image, max_delta=0.2)
+            image = tf.image.random_contrast(image, lower=0.2, upper=1.2)
+            image = tf.image.random_saturation(image, lower=0.2, upper=1.2)
+            image = tf.image.random_hue(image, max_delta=0.1)
 
         # The random_* ops do not necessarily clamp.
         image = tf.clip_by_value(image, 0.0, 1.0)

utils/train_utils.py

@@ -342,12 +342,6 @@ def restore_fn(flags):
                                  flags.checkpoint_model_scope): var
              for var in variables_to_restore}
 
-    # # modify
-    # if flags.extra_model_name is not None:
-    #     for key, var in variables_to_restore.items():
-    #         if key.split('/')[0] == flags.extra_model_name:
-    #             key.replace(flags.extra_model_name, flags.checkpoint_model_scope2)
-
     slim.assign_from_checkpoint_fn(flags.pre_trained_checkpoint,
                                    variables_to_restore)
     tf.compat.v1.logging.info('Fine-tuning from %s. Ignoring missing vars: %s' %
@@ -383,7 +377,7 @@ def custom_restore_fn(flags):
                 excluded = True
                 break
         if not excluded:
-            variables_to_restore.append(var)
+            variables_to_restore.append('tower0/' + var)
 
     # Change model scope if necessary.
     if flags.checkpoint_model_scope is not None:
@@ -392,12 +386,6 @@ def custom_restore_fn(flags):
                                  flags.checkpoint_model_scope): var
              for var in variables_to_restore}
 
-    # # modify
-    # if flags.extra_model_name is not None:
-    #     for key, var in variables_to_restore.items():
-    #         if key.split('/')[0] == flags.extra_model_name:
-    #             key.replace(flags.extra_model_name, flags.checkpoint_model_scope2)
-
     slim.assign_from_checkpoint_fn(flags.pre_trained_checkpoint,
                                    variables_to_restore)
     tf.compat.v1.logging.info('Fine-tuning from %s. Ignoring missing vars: %s' %