add uniformer

Signed-off-by: gexy5 <xinyu.ge@zilliz.com>

README.md

@@ -1,2 +1,106 @@
-# uniformer
+# Video Classification with Uniformer
 
+*Author: [Xinyu Ge](https://github.com/gexy185)*
+
+<br />
+
+## Description
+
+A video classification operator generates labels (and corresponding scores) and extracts features for the input video.
+It transforms the video into frames and loads pre-trained models by model names.
+This operator has implemented pre-trained models from [Uniformer](https://arxiv.org/abs/2201.09450)
+and maps vectors with labels provided by datasets used for pre-training.
+
+<br />
+
+## Code Example
+
+Use the pretrained Uniformer model to classify and generate a vector for the given video path './archery.mp4' 
+([download](https://dl.fbaipublicfiles.com/pytorchvideo/projects/archery.mp4)).
+
+ *Write the pipeline in simplified style*:
+
+- Predict labels (default):
+```python
+import towhee
+
+(
+    towhee.glob('./archery.mp4') 
+          .video_decode.ffmpeg()
+          .action_classification.uniformer(
+            model_name='omnivore_swinT', topk=5)
+          .show()
+)
+```
+<img src="./result1.png" height="px"/>
+
+*Write a same pipeline with explicit inputs/outputs name specifications*:
+
+```python
+import towhee
+
+(
+    towhee.glob['path']('./archery.mp4')
+          .video_decode.ffmpeg['path', 'frames']()
+          .action_classification.uniformer['frames', ('labels', 'scores', 'features')](
+                model_name='uniformer_k400_s8')
+          .select['path', 'labels', 'scores', 'features']()
+          .show(formatter={'path': 'video_path'})
+)
+```
+
+<img src="./result2.png" height="px"/>
+
+<br />
+
+## Factory Constructor
+
+Create the operator via the following factory method
+
+***video_classification.uniformer(
+model_name='omnivore_swinT', skip_preprocess=False, classmap=None, topk=5)***
+
+**Parameters:**
+
+​	***model_name***: *str*
+
+​	The name of pre-trained uniformer model.
+
+​    Supported model names:
+- uniformer_k400_s8
+- uniformer_k400_s16
+- uniformer_k400_b8
+- uniformer_k400_b16
+
+​	***skip_preprocess***: *bool*
+
+​	Flag to control whether to skip video transforms, defaults to False.
+If set to True, the step to transform videos will be skipped.
+In this case, the user should guarantee that all the input video frames are already reprocessed properly,
+and thus can be fed to model directly.
+
+​	***classmap***: *Dict[str: int]*: 
+
+​	Dictionary that maps class names to one hot vectors.
+If not given, the operator will load the default class map dictionary.
+
+​	***topk***: *int*
+
+​	The topk labels & scores to present in result. The default value is 5.
+
+## Interface
+
+A video classification operator generates a list of class labels
+and a corresponding vector in numpy.ndarray given a video input data.
+
+**Parameters:**
+
+​	***video***: *Union[str, numpy.ndarray]*
+
+​	Input video data using local path in string or video frames in ndarray.
+
+
+**Returns**: *(list, list, torch.Tensor)*
+
+​	A tuple of (labels, scores, features),
+which contains lists of predicted class names and corresponding scores.

__init__.py

@@ -0,0 +1,19 @@
+# Copyright 2022 Zilliz. 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 .uniformer import Uniformer
+
+
+def uniformer(**kwargs):
+    return Uniformer(**kwargs)

uniformer.py

@@ -0,0 +1,131 @@
+import logging
+import os
+import json
+from pathlib import Path
+from typing import List
+
+import torch
+import numpy
+
+from towhee import register
+from towhee.operator.base import NNOperator
+from towhee.types.video_frame import VideoFrame
+from towhee.models.utils.video_transforms import get_configs, transform_video
+from towhee.models.uniformer.uniformer import create_model
+
+log = logging.getLogger()
+
+@register(output_schema=['labels', 'scores', 'features'])
+class Tsm(NNOperator):
+    """
+    Generate a list of class labels given a video input data.
+    Default labels are from [Kinetics400 Dataset](https://deepmind.com/research/open-source/kinetics).
+    Args:
+        model_name (`str`):
+            Supported model names:
+            - uniformer_k400_s8
+            - uniformer_k400_s16
+            - uniformer_k400_b8
+            - uniformer_k400_b16
+        skip_preprocess (`str`):
+            Flag to skip video transforms.
+        predict (`bool`):
+            Flag to control whether predict labels. If False, then return video embedding.
+        classmap (`dict=None`):
+            The dictionary maps classes to integers.
+        topk (`int=5`):
+            The number of classification labels to be returned (ordered by possibility from high to low).
+    """
+    def __init__(self,
+                 model_name: str = 'uniformer_k400_s8',
+                 framework: str = 'pytorch',
+                 skip_preprocess: bool = False,
+                 classmap: dict = None,
+                 topk: int = 5,
+                 ):
+        super().__init__(framework=framework)
+        self.model_name = model_name
+        self.skip_preprocess = skip_preprocess
+        self.topk = topk
+        if 'k400' in model_name:
+            self.dataset_name = 'kinetics_400'
+        if classmap is None:
+            class_file = os.path.join(str(Path(__file__).parent), self.dataset_name+'.json')
+            with open(class_file, "r") as f:
+                kinetics_classes = json.load(f)
+            self.classmap = {}
+            for k, v in kinetics_classes.items():
+                self.classmap[v] = str(k).replace('"', '')
+        else:
+            self.classmap = classmap
+        self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
+        
+        self.input_mean = [0.45, 0.45, 0.45]
+        self.input_std = [0.225, 0.225, 0.225]
+
+        model_path = str(Path(__file__).parent)
+        model_path = {
+            'uniformer_k400_s8': os.path.join(model_path, 'uniformer_small_k400_8x8.pth'),
+            'uniformer_k400_s16': os.path.join(model_path, 'uniformer_small_k400_16x4.pth'),
+            'uniformer_k400_b8': os.path.join(model_path, 'uniformer_base_k400_8x8.pth'),
+            'uniformer_k400_b16': os.path.join(model_path, 'uniformer_base_k400_16x4.pth'),
+        }
+        self.weights_path = model_path[model_name]
+        self.model = create_model(model_name=model_name, pretrained=True, weights_path=self.weights_path, device=self.device)
+        if '8' in model_name:
+            self.transform_cfgs = get_configs(
+                    side_size=256,
+                    crop_size=224,
+                    num_frames=8,
+                    mean=self.input_mean,
+                    std=self.input_std,
+                    )
+        elif '16' in model_name:
+            self.transform_cfgs = get_configs(
+                    side_size=256,
+                    crop_size=224,
+                    num_frames=16,
+                    mean=self.input_mean,
+                    std=self.input_std,
+                    )
+        self.model.eval()
+
+    def __call__(self, video: List[VideoFrame]):
+        """
+        Args:
+            video (`List[VideoFrame]`):
+                Video path in string.
+
+        Returns:
+            (labels, scores)
+                A tuple of lists (labels, scores).
+            OR emb
+                Video embedding.
+        """
+        # Convert list of towhee.types.Image to numpy.ndarray in float32
+        video = numpy.stack([img.astype(numpy.float32)/255. for img in video], axis=0)
+        assert len(video.shape) == 4
+        video = video.transpose(3, 0, 1, 2)  # twhc -> ctwh
+
+        # Transform video data given configs
+        if self.skip_preprocess:
+            self.transform_cfgs.update(num_frames=None)
+
+        data = transform_video(
+                    video=video,
+                    **self.transform_cfgs
+                    )
+        inputs = data.to(self.device)[None, ...]
+
+        feats = self.model.forward_features(inputs)
+        features = feats.to('cpu').squeeze(0).detach().numpy()
+
+        feats = feats.flatten(2).mean(-1)
+        outs = self.model.head(feats)
+        post_act = torch.nn.Softmax(dim=1)
+        preds = post_act(outs)
+        pred_scores, pred_classes = preds.topk(k=self.topk)
+        labels = [self.classmap[int(i)] for i in pred_classes[0]]
+        scores = [round(float(x), 5) for x in pred_scores[0]]
+
+        return labels, scores, features