update: namespaces

Author: innat

.github/workflows/format.yml

@@ -26,7 +26,9 @@ This is a unofficial `Keras` reimplementation of [VideoMAE: Masked Autoencoders
 # Install 
 
 ```bash
-pip install -U git+https://github.com/innat/VideoMAE.git
+git clone https://github.com/innat/VideoMAE.git
+cd VideoMAE
+pip install -e . 
 ```
 
 # Usage

README.md

@@ -1,4 +1,2 @@
 tensorflow>=2.12
 opencv-python>=4.1.2
-isort
-flake8

requirements.txt

@@ -10,7 +10,7 @@
     long_description_content_type="text/markdown",
     author="Mohammed Innat",
     author_email="[email protected]",
-    url="https://github.com/innat/DOLG-TensorFlow",
+    url="https://github.com/innat/VideoMAE",
     keywords=["deep learning", "image retrieval", "image recognition"],
     install_requires=[
         "opencv-python>=4.1.2",
@@ -25,6 +25,6 @@
         "Intended Audience :: Developers",
         "Topic :: Scientific/Engineering :: Artificial Intelligence",
         "License :: OSI Approved :: MIT License",
-        "Programming Language :: Python :: 3.7",
+        "Programming Language :: Python :: 3.8",
     ],
 )

setup.py

@@ -3,7 +3,7 @@
 from tensorflow.keras import layers
 from tensorflow.keras.initializers import Constant
 
-from videomae.layers import TFAttention, TFDropPath, TFMlp
+from layers import TFAttention, TFDropPath, TFMlp
 
 
 class TFBlock(keras.Model):

videomae/blocks/basic.py

@@ -4,8 +4,8 @@
 from tensorflow import keras
 from tensorflow.keras import layers
 
-from videomae.blocks import TFBlock
-from videomae.layers import TFAttention
+from blocks import TFBlock
+from layers import TFAttention
 
 
 class TFPretrainVisionTransformerDecoder(keras.Model):

videomae/blocks/vit_decoder.py

@@ -5,9 +5,9 @@
 from tensorflow.keras import layers
 from tensorflow.keras.initializers import TruncatedNormal
 
-from videomae.blocks import TFBlock
-from videomae.layers import TFAttention, TFPatchEmbed
-from videomae.utils import get_sinusoid_encoding_table_tf
+from blocks import TFBlock
+from layers import TFAttention, TFPatchEmbed
+from utils import get_sinusoid_encoding_table_tf
 
 
 class TFPretrainVisionTransformerEncoder(keras.Model):

videomae/blocks/vit_encoder.py

@@ -1,5 +1,4 @@
 from .attention import TFAttention
 from .drop_path import TFDropPath
-from .masking_generator import TubeMaskingGenerator
 from .mlp import TFMlp
 from .patch_embed import TFPatchEmbed

videomae/layers/__init__.py

@@ -1 +1,2 @@
 from .sinusoid_encoding_table import get_sinusoid_encoding_table_tf
+from .masking_generator import TubeMaskingGenerator

videomae/utils/__init__.py

@@ -1,27 +1,27 @@
-import numpy as np
-
-
-class TubeMaskingGenerator:
-    def __init__(self, input_size, mask_ratio):
-        self.frames, self.height, self.width = input_size
-        self.num_patches_per_frame = self.height * self.width
-        self.total_patches = self.frames * self.num_patches_per_frame
-        self.num_masks_per_frame = int(mask_ratio * self.num_patches_per_frame)
-        self.total_masks = self.frames * self.num_masks_per_frame
-
-    def __repr__(self):
-        repr_str = "Maks: total patches {}, mask patches {}".format(
-            self.total_patches, self.total_masks
-        )
-        return repr_str
-
-    def __call__(self):
-        mask_per_frame = np.hstack(
-            [
-                np.zeros(self.num_patches_per_frame - self.num_masks_per_frame),
-                np.ones(self.num_masks_per_frame),
-            ]
-        )
-        np.random.shuffle(mask_per_frame)
-        mask = np.tile(mask_per_frame, (self.frames, 1)).flatten()
-        return mask
+import numpy as np
+
+
+class TubeMaskingGenerator:
+    def __init__(self, input_size, mask_ratio):
+        self.frames, self.height, self.width = input_size
+        self.num_patches_per_frame = self.height * self.width
+        self.total_patches = self.frames * self.num_patches_per_frame
+        self.num_masks_per_frame = int(mask_ratio * self.num_patches_per_frame)
+        self.total_masks = self.frames * self.num_masks_per_frame
+
+    def __repr__(self):
+        repr_str = "Maks: total patches {}, mask patches {}".format(
+            self.total_patches, self.total_masks
+        )
+        return repr_str
+
+    def __call__(self):
+        mask_per_frame = np.hstack(
+            [
+                np.zeros(self.num_patches_per_frame - self.num_masks_per_frame),
+                np.ones(self.num_masks_per_frame),
+            ]
+        )
+        np.random.shuffle(mask_per_frame)
+        mask = np.tile(mask_per_frame, (self.frames, 1)).flatten()
+        return mask

videomae/layers/masking_generator.py videomae/utils/masking_generator.py

@@ -10,8 +10,8 @@
 from tensorflow.keras import layers
 from tensorflow.keras.initializers import RandomNormal
 
-from videomae.blocks import (TFPretrainVisionTransformerDecoder,
-                             TFPretrainVisionTransformerEncoder)
+from videomae.blocks import TFPretrainVisionTransformerDecoder
+from videomae.blocks import TFPretrainVisionTransformerEncoder
 from videomae.utils import get_sinusoid_encoding_table_tf
 
 from .model_configs import MODEL_CONFIGS