Convert models from Pytorch to Keras (#99)

* Convert models from Pytorch to Keras

* Convert ResNet 50, 101, 152 models

* Rebase; unify layer names b/t frameworks; produce samples

* Verify accuracy of models using both Keras and Pytorch pipeline

* Add args for arch key and model type; make style

Author: natuan

scripts/keras_classification.py

@@ -304,8 +304,8 @@
 from sparseml.keras.utils import (
     LossesAndMetricsLoggingCallback,
     ModelExporter,
-    keras,
     TensorBoardLogger,
+    keras,
 )
 from sparseml.utils import create_dirs
 

src/sparseml/keras/datasets/classification/__init__.py

@@ -19,4 +19,5 @@
 # flake8: noqa
 
 from .imagefolder import *
+from .imagenet import *
 from .imagenette import *

src/sparseml/keras/datasets/classification/imagefolder.py

@@ -48,7 +48,6 @@
 def imagenet_normalizer(img: tensorflow.Tensor, mode: str):
     """
     Normalize an image using mean and std of the imagenet dataset
-
     :param img: The input image to normalize
     :param mode: either "tf", "caffe", "torch"
     :return: The normalized image
@@ -73,7 +72,7 @@ def imagenet_normalizer(img: tensorflow.Tensor, mode: str):
 
 def default_imagenet_normalizer():
     def normalizer(img: tensorflow.Tensor):
-        # Default to the same preprocessing used by ResNet
+        # Default to the same preprocessing used by Keras Applications ResNet
         return imagenet_normalizer(img, "caffe")
 
     return normalizer
@@ -109,7 +108,7 @@ def __init__(
         self,
         root: str,
         train: bool,
-        image_size: Union[int, Tuple[int, int]] = 224,
+        image_size: Union[None, int, Tuple[int, int]] = 224,
         pre_resize_transforms: Union[SplitsTransforms, None] = SplitsTransforms(
             train=(
                 random_scaling_crop(),
@@ -126,9 +125,14 @@ def __init__(
         if not os.path.exists(self._root):
             raise ValueError("Data set folder {} must exist".format(self._root))
         self._train = train
-        self._image_size = (
-            image_size if isinstance(image_size, tuple) else (image_size, image_size)
-        )
+        if image_size is not None:
+            self._image_size = (
+                image_size
+                if isinstance(image_size, tuple)
+                else (image_size, image_size)
+            )
+        else:
+            self._image_size = None
         self._pre_resize_transforms = pre_resize_transforms
         self._post_resize_transforms = post_resize_transforms
 
@@ -199,7 +203,6 @@ def processor(self, file_path: tensorflow.Tensor, label: tensorflow.Tensor):
         """
         img = tensorflow.io.read_file(file_path)
         img = tensorflow.image.decode_jpeg(img, channels=3)
-
         if self.pre_resize_transforms:
             transforms = (
                 self.pre_resize_transforms.train
@@ -209,7 +212,7 @@ def processor(self, file_path: tensorflow.Tensor, label: tensorflow.Tensor):
             if transforms:
                 for trans in transforms:
                     img = trans(img)
-        if self._image_size:
+        if self._image_size is not None:
             img = tensorflow.image.resize(img, self.image_size)
 
         if self.post_resize_transforms:
@@ -221,7 +224,6 @@ def processor(self, file_path: tensorflow.Tensor, label: tensorflow.Tensor):
             if transforms:
                 for trans in transforms:
                     img = trans(img)
-
         return img, label
 
     def creator(self):

src/sparseml/keras/datasets/classification/imagenet.py

@@ -0,0 +1,130 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Imagenet dataset implementations for the image classification field in computer vision.
+More info for the dataset can be found `here <http://www.image-net.org/>`__.
+"""
+
+import random
+from typing import Tuple, Union
+
+import tensorflow as tf
+
+from sparseml.keras.datasets.classification import (
+    ImageFolderDataset,
+    SplitsTransforms,
+    imagenet_normalizer,
+)
+from sparseml.keras.datasets.helpers import random_scaling_crop
+from sparseml.keras.datasets.registry import DatasetRegistry
+from sparseml.keras.utils import keras
+from sparseml.utils import clean_path
+from sparseml.utils.datasets import (
+    IMAGENET_RGB_MEANS,
+    IMAGENET_RGB_STDS,
+    default_dataset_path,
+)
+
+
+__all__ = ["ImageNetDataset"]
+
+
+def torch_imagenet_normalizer():
+    def normalizer(image: tf.Tensor):
+        return imagenet_normalizer(image, "torch")
+
+    return normalizer
+
+
+def imagenet_pre_resize_processor():
+    def processor(image: tf.Tensor):
+        image_batch = tf.expand_dims(image, axis=0)
+
+        # Resize the image the following way to match torchvision's Resize
+        # transform used by Pytorch code path for Imagenet:
+        #   torchvision.transforms.Resize(256)
+        # which resize the smaller side of images to 256 and the other one based
+        # on the aspect ratio
+        shape = tf.shape(image)
+        h, w = shape[0], shape[1]
+        if h > w:
+            new_h, new_w = tf.cast(256 * h / w, dtype=tf.uint16), tf.constant(
+                256, dtype=tf.uint16
+            )
+        else:
+            new_h, new_w = tf.constant(256, dtype=tf.uint16), tf.cast(
+                256 * w / h, dtype=tf.uint16
+            )
+        resizer = keras.layers.experimental.preprocessing.Resizing(new_h, new_w)
+        image_batch = tf.cast(resizer(image_batch), dtype=tf.uint8)
+
+        # Center crop
+        center_cropper = keras.layers.experimental.preprocessing.CenterCrop(224, 224)
+        image_batch = tf.cast(center_cropper(image_batch), dtype=tf.uint8)
+
+        return image_batch[0, :]
+
+    return processor
+
+
+@DatasetRegistry.register(
+    key=["imagenet"],
+    attributes={
+        "num_classes": 1000,
+        "transform_means": IMAGENET_RGB_MEANS,
+        "transform_stds": IMAGENET_RGB_STDS,
+    },
+)
+class ImageNetDataset(ImageFolderDataset):
+    """
+    Wrapper for the ImageNet dataset to apply standard transforms.
+
+    :param root: The root folder to find the dataset at
+    :param train: True if this is for the training distribution,
+        False for the validation
+    :param rand_trans: True to apply RandomCrop and RandomHorizontalFlip to the data,
+        False otherwise
+    :param image_size: the size of the image to output from the dataset
+    """
+
+    def __init__(
+        self,
+        root: str = default_dataset_path("imagenet"),
+        train: bool = True,
+        rand_trans: bool = False,
+        image_size: Union[None, int, Tuple[int, int]] = 224,
+        pre_resize_transforms=SplitsTransforms(
+            train=(
+                random_scaling_crop(),
+                tf.image.random_flip_left_right,
+            ),
+            val=(imagenet_pre_resize_processor(),),
+        ),
+        post_resize_transforms=SplitsTransforms(
+            train=(torch_imagenet_normalizer(),), val=(torch_imagenet_normalizer(),)
+        ),
+    ):
+        root = clean_path(root)
+        super().__init__(
+            root,
+            train,
+            image_size=image_size,
+            pre_resize_transforms=pre_resize_transforms,
+            post_resize_transforms=post_resize_transforms,
+        )
+
+        if train:
+            # make sure we don't preserve the folder structure class order
+            random.shuffle(self.samples)

src/sparseml/keras/datasets/dataset.py

@@ -46,7 +46,6 @@ def build(
     ) -> tensorflow.data.Dataset:
         """
         Create the dataset in the current graph using tensorflow.data APIs
-
         :param batch_size: the batch size to create the dataset for
         :param repeat_count: the number of times to repeat the dataset,
             if unset or None, will repeat indefinitely

src/sparseml/keras/datasets/helpers.py

@@ -33,7 +33,6 @@ def random_scaling_crop(
     """
     Random crop implementation which also randomly scales the crop taken
     as well as the aspect ratio of the crop.
-
     :param scale_range: the (min, max) of the crop scales to take from the orig image
     :param ratio_range: the (min, max) of the aspect ratios to take from the orig image
     :return: the callable function for random scaling crop op,

src/sparseml/keras/models/__init__.py

@@ -18,5 +18,6 @@
 
 # flake8: noqa
 
+from .classification import *
 from .external import *
 from .registry import *

src/sparseml/keras/models/classification/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .resnet import *