Merge branch 'main' of https://github.com/awsaf49/gcvit-tf

CITATION.cff

@@ -0,0 +1,9 @@
+cff-version: 1.2.0
+message: "If you use this software, please cite it as below."
+authors:
+- family-names: "Rahman"
+  given-names: "Md Awsafur"
+title: "Tensorflow 2.0 Implementation of GCViT: Global Context Vision Transformer"
+version: 1.1.5
+date-released: 2020-07-20
+url: "https://awsaf49.github.io/gcvit-tf"

gcvit/layers/attention.py

@@ -33,7 +33,7 @@ def build(self, input_shape):
             dim * self.qkv_size, use_bias=self.qkv_bias, name="qkv"
         )
         self.relative_position_bias_table = self.add_weight(
-            "relative_position_bias_table",
+            name="relative_position_bias_table",
             shape=[
                 (2 * self.window_size[0] - 1) * (2 * self.window_size[1] - 1),
                 self.num_heads,
@@ -50,7 +50,7 @@ def build(self, input_shape):
             self.proj_dropout, name="proj_drop"
         )
         self.softmax = tf.keras.layers.Activation("softmax", name="softmax")
-        self.relative_position_index = self.get_relative_position_index()
+        # self.relative_position_index = self.get_relative_position_index()
         super().build(input_shape)
 
     def get_relative_position_index(self):
@@ -101,7 +101,7 @@ def call(self, inputs, **kwargs):
         attn = q @ tf.transpose(k, perm=[0, 1, 3, 2])
         relative_position_bias = tf.gather(
             self.relative_position_bias_table,
-            tf.reshape(self.relative_position_index, shape=[-1]),
+            tf.reshape(self.get_relative_position_index(), shape=[-1]),
         )
         relative_position_bias = tf.reshape(
             relative_position_bias,

gcvit/layers/block.py

@@ -65,14 +65,14 @@ def build(self, input_shape):
         )
         if self.layer_scale is not None:
             self.gamma1 = self.add_weight(
-                "gamma1",
+                name="gamma1",
                 shape=[C],
                 initializer=tf.keras.initializers.Constant(self.layer_scale),
                 trainable=True,
                 dtype=self.dtype,
             )
             self.gamma2 = self.add_weight(
-                "gamma2",
+                name="gamma2",
                 shape=[C],
                 initializer=tf.keras.initializers.Constant(self.layer_scale),
                 trainable=True,

gcvit/layers/feature.py

@@ -67,11 +67,11 @@ def build(self, input_shape):
         self.avg_pool = AdaptiveAveragePooling2D(1, name="avg_pool")
         self.fc = [
             tf.keras.layers.Dense(
-                int(inp * self.expansion), use_bias=False, name="fc/0"
+                int(inp * self.expansion), use_bias=False, name="fc_0"
             ),
-            tf.keras.layers.Activation("gelu", name="fc/1"),
-            tf.keras.layers.Dense(self.oup, use_bias=False, name="fc/2"),
-            tf.keras.layers.Activation("sigmoid", name="fc/3"),
+            tf.keras.layers.Activation("gelu", name="fc_1"),
+            tf.keras.layers.Dense(self.oup, use_bias=False, name="fc_2"),
+            tf.keras.layers.Activation("sigmoid", name="fc_3"),
         ]
         super().build(input_shape)
 
@@ -111,17 +111,17 @@ def build(self, input_shape):
                 strides=1,
                 padding="valid",
                 use_bias=False,
-                name="conv/0",
+                name="conv_0",
             ),
-            tf.keras.layers.Activation("gelu", name="conv/1"),
-            SE(name="conv/2"),
+            tf.keras.layers.Activation("gelu", name="conv_1"),
+            SE(name="conv_2"),
             tf.keras.layers.Conv2D(
                 dim,
                 kernel_size=1,
                 strides=1,
                 padding="valid",
                 use_bias=False,
-                name="conv/3",
+                name="conv_3",
             ),
         ]
         self.reduction = tf.keras.layers.Conv2D(
@@ -179,17 +179,17 @@ def build(self, input_shape):
                 strides=1,
                 padding="valid",
                 use_bias=False,
-                name="conv/0",
+                name="conv_0",
             ),
-            tf.keras.layers.Activation("gelu", name="conv/1"),
-            SE(name="conv/2"),
+            tf.keras.layers.Activation("gelu", name="conv_1"),
+            SE(name="conv_2"),
             tf.keras.layers.Conv2D(
                 dim,
                 kernel_size=1,
                 strides=1,
                 padding="valid",
                 use_bias=False,
-                name="conv/3",
+                name="conv_3",
             ),
         ]
         if not self.keep_dim:
@@ -237,7 +237,7 @@ def __init__(self, keep_dims=False, **kwargs):
 
     def build(self, input_shape):
         self.to_q_global = [
-            FeatExtract(keep_dim, name=f"to_q_global/{i}")
+            FeatExtract(keep_dim, name=f"to_q_global_{i}")
             for i, keep_dim in enumerate(self.keep_dims)
         ]
         super().build(input_shape)

gcvit/layers/level.py

@@ -59,7 +59,7 @@ def build(self, input_shape):
                 attn_drop=self.attn_drop,
                 path_drop=path_drop[i],
                 layer_scale=self.layer_scale,
-                name=f"blocks/{i}",
+                name=f"blocks_{i}",
             )
             for i in range(self.depth)
         ]

gcvit/models/gcvit.py

@@ -142,7 +142,7 @@ def __init__(
                 path_drop=path_drop,
                 layer_scale=layer_scale,
                 resize_query=resize_query,
-                name=f"levels/{i}",
+                name=f"levels_{i}",
             )
             self.levels.append(level)
         self.norm = tf.keras.layers.LayerNormalization(