init

app.py

@@ -0,0 +1,131 @@
+import os
+
+import torch
+import torch.nn as nn
+import numpy as np
+import torch.nn.functional as F
+import torchvision.transforms as T
+from PIL import Image
+from decord import VideoReader
+from decord import cpu
+from uniformerv2 import uniformerv2_b16
+from mitv1_class_index import mitv1_classnames
+from transforms import (
+    GroupNormalize, GroupScale, GroupCenterCrop, 
+    Stack, ToTorchFormatTensor
+)
+
+import gradio as gr
+from huggingface_hub import hf_hub_download
+
+class Uniformerv2(nn.Module):
+    def __init__(self, model):
+        super().__init__()
+        self.backbone = model
+    
+    def forward(self, x):
+        return self.backbone(x)
+
+# Device on which to run the model
+# Set to cuda to load on GPU
+device = "cpu"
+model_path = hf_hub_download(repo_id="Andy1621/uniformerv2", filename="mit_uniformerv2_b16_8x224.pyth")
+# Pick a pretrained model 
+model = Uniformerv2(uniformerv2_b16(pretrained=False, t_size=8, no_lmhra=True, temporal_downsample=False, num_classes=339))
+state_dict = torch.load(model_path, map_location='cpu')
+model.load_state_dict(state_dict)
+
+# Set to eval mode and move to desired device
+model = model.to(device)
+model = model.eval()
+
+# Create an id to label name mapping
+mitv1_id_to_classname = {}
+for k, v in mitv1_classnames.items():
+    mitv1_id_to_classname[k] = v
+
+
+def get_index(num_frames, num_segments=8):
+    seg_size = float(num_frames - 1) / num_segments
+    start = int(seg_size / 2)
+    offsets = np.array([
+        start + int(np.round(seg_size * idx)) for idx in range(num_segments)
+    ])
+    return offsets
+
+
+def load_video(video_path):
+    vr = VideoReader(video_path, ctx=cpu(0))
+    num_frames = len(vr)
+    frame_indices = get_index(num_frames, 8)
+
+    # transform
+    crop_size = 224
+    scale_size = 256
+    input_mean = [0.485, 0.456, 0.406]
+    input_std = [0.229, 0.224, 0.225]
+
+    transform = T.Compose([
+        GroupScale(int(scale_size)),
+        GroupCenterCrop(crop_size),
+        Stack(),
+        ToTorchFormatTensor(),
+        GroupNormalize(input_mean, input_std) 
+    ])
+
+    images_group = list()
+    for frame_index in frame_indices:
+        img = Image.fromarray(vr[frame_index].asnumpy())
+        images_group.append(img)
+    torch_imgs = transform(images_group)
+    return torch_imgs
+    
+
+def inference(video):
+    vid = load_video(video)
+    
+    # The model expects inputs of shape: B x C x H x W
+    TC, H, W = vid.shape
+    inputs = vid.reshape(1, TC//3, 3, H, W).permute(0, 2, 1, 3, 4)
+    
+    prediction = model(inputs)
+    prediction = F.softmax(prediction, dim=1).flatten()
+
+    return {mitv1_id_to_classname[str(i)]: float(prediction[i]) for i in range(400)}
+    
+
+def set_example_video(example: list) -> dict:
+    return gr.Video.update(value=example[0])
+
+
+demo = gr.Blocks()
+with demo:
+    gr.Markdown(
+        """
+        # UniFormerV2-B
+        Gradio demo for <a href='https://github.com/OpenGVLab/UniFormerV2' target='_blank'>UniFormerV2</a>: To use it, simply upload your video, or click one of the examples to load them. Read more at the links below.
+        """
+    )
+
+    with gr.Box():
+        with gr.Row():
+                with gr.Column():
+                    with gr.Row():
+                        input_video = gr.Video(label='Input Video')
+                    with gr.Row():
+                        submit_button = gr.Button('Submit')
+                with gr.Column():
+                    label = gr.Label(num_top_classes=5)
+        with gr.Row():
+            example_videos = gr.Dataset(components=[input_video], samples=[['clapping.mp4'], ['jumping.mp4'], ['swimming.mp4']])
+
+    gr.Markdown(
+        """
+        <p style='text-align: center'><a href='https://arxiv.org/abs/2211.09552' target='_blank'>[Arxiv] UniFormerV2: Spatiotemporal Learning by Arming Image ViTs with Video UniFormer</a> | <a href='https://github.com/OpenGVLab/UniFormerV2' target='_blank'>Github Repo</a></p>
+        """
+    )
+
+    submit_button.click(fn=inference, inputs=input_video, outputs=label)
+    example_videos.click(fn=set_example_video, inputs=example_videos, outputs=example_videos.components)
+
+demo.launch(enable_queue=True)

mitv1_class_index.py

@@ -0,0 +1,341 @@
+mitv1_classnames = {
+    "0": "adult+female+singing",
+    "1": "adult+female+speaking",
+    "2": "adult+male+singing",
+    "3": "adult+male+speaking",
+    "4": "aiming",
+    "5": "applauding",
+    "6": "arresting",
+    "7": "ascending",
+    "8": "asking",
+    "9": "assembling",
+    "10": "attacking",
+    "11": "autographing",
+    "12": "baking",
+    "13": "balancing",
+    "14": "baptizing",
+    "15": "barbecuing",
+    "16": "barking",
+    "17": "bathing",
+    "18": "bending",
+    "19": "bicycling",
+    "20": "biting",
+    "21": "blocking",
+    "22": "blowing",
+    "23": "boarding",
+    "24": "boating",
+    "25": "boiling",
+    "26": "bouncing",
+    "27": "bowing",
+    "28": "bowling",
+    "29": "boxing",
+    "30": "breaking",
+    "31": "brushing",
+    "32": "bubbling",
+    "33": "building",
+    "34": "bulldozing",
+    "35": "burning",
+    "36": "burying",
+    "37": "buttoning",
+    "38": "buying",
+    "39": "calling",
+    "40": "camping",
+    "41": "carrying",
+    "42": "carving",
+    "43": "catching",
+    "44": "celebrating",
+    "45": "chasing",
+    "46": "cheering",
+    "47": "cheerleading",
+    "48": "chewing",
+    "49": "child+singing",
+    "50": "child+speaking",
+    "51": "chopping",
+    "52": "clapping",
+    "53": "clawing",
+    "54": "cleaning",
+    "55": "clearing",
+    "56": "climbing",
+    "57": "clinging",
+    "58": "clipping",
+    "59": "closing",
+    "60": "coaching",
+    "61": "colliding",
+    "62": "combing",
+    "63": "combusting",
+    "64": "competing",
+    "65": "constructing",
+    "66": "cooking",
+    "67": "coughing",
+    "68": "covering",
+    "69": "cracking",
+    "70": "crafting",
+    "71": "cramming",
+    "72": "crashing",
+    "73": "crawling",
+    "74": "crouching",
+    "75": "crushing",
+    "76": "crying",
+    "77": "cuddling",
+    "78": "cutting",
+    "79": "dancing",
+    "80": "descending",
+    "81": "destroying",
+    "82": "digging",
+    "83": "dining",
+    "84": "dipping",
+    "85": "discussing",
+    "86": "diving",
+    "87": "dragging",
+    "88": "draining",
+    "89": "drawing",
+    "90": "drenching",
+    "91": "dressing",
+    "92": "drilling",
+    "93": "drinking",
+    "94": "dripping",
+    "95": "driving",
+    "96": "dropping",
+    "97": "drumming",
+    "98": "drying",
+    "99": "dunking",
+    "100": "dusting",
+    "101": "eating",
+    "102": "emptying",
+    "103": "entering",
+    "104": "erupting",
+    "105": "exercising",
+    "106": "exiting",
+    "107": "extinguishing",
+    "108": "falling",
+    "109": "feeding",
+    "110": "fencing",
+    "111": "fighting",
+    "112": "filling",
+    "113": "filming",
+    "114": "fishing",
+    "115": "flicking",
+    "116": "flipping",
+    "117": "floating",
+    "118": "flooding",
+    "119": "flowing",
+    "120": "flying",
+    "121": "folding",
+    "122": "frowning",
+    "123": "frying",
+    "124": "fueling",
+    "125": "gambling",
+    "126": "gardening",
+    "127": "giggling",
+    "128": "giving",
+    "129": "grilling",
+    "130": "grinning",
+    "131": "gripping",
+    "132": "grooming",
+    "133": "guarding",
+    "134": "hammering",
+    "135": "handcuffing",
+    "136": "handwriting",
+    "137": "hanging",
+    "138": "hiking",
+    "139": "hitchhiking",
+    "140": "hitting",
+    "141": "howling",
+    "142": "hugging",
+    "143": "hunting",
+    "144": "imitating",
+    "145": "inflating",
+    "146": "injecting",
+    "147": "instructing",
+    "148": "interviewing",
+    "149": "jogging",
+    "150": "joining",
+    "151": "juggling",
+    "152": "jumping",
+    "153": "kicking",
+    "154": "kissing",
+    "155": "kneeling",
+    "156": "knitting",
+    "157": "knocking",
+    "158": "landing",
+    "159": "laughing",
+    "160": "launching",
+    "161": "leaking",
+    "162": "leaning",
+    "163": "leaping",
+    "164": "lecturing",
+    "165": "licking",
+    "166": "lifting",
+    "167": "loading",
+    "168": "locking",
+    "169": "manicuring",
+    "170": "marching",
+    "171": "marrying",
+    "172": "massaging",
+    "173": "measuring",
+    "174": "mopping",
+    "175": "mowing",
+    "176": "officiating",
+    "177": "opening",
+    "178": "operating",
+    "179": "overflowing",
+    "180": "packaging",
+    "181": "packing",
+    "182": "painting",
+    "183": "parading",
+    "184": "paying",
+    "185": "pedaling",
+    "186": "peeling",
+    "187": "performing",
+    "188": "photographing",
+    "189": "picking",
+    "190": "piloting",
+    "191": "pitching",
+    "192": "placing",
+    "193": "planting",
+    "194": "playing",
+    "195": "playing+fun",
+    "196": "playing+music",
+    "197": "playing+sports",
+    "198": "playing+videogames",
+    "199": "plugging",
+    "200": "plunging",
+    "201": "pointing",
+    "202": "poking",
+    "203": "pouring",
+    "204": "praying",
+    "205": "preaching",
+    "206": "pressing",
+    "207": "protesting",
+    "208": "pulling",
+    "209": "punching",
+    "210": "punting",
+    "211": "pushing",
+    "212": "putting",
+    "213": "queuing",
+    "214": "racing",
+    "215": "rafting",
+    "216": "raining",
+    "217": "raising",
+    "218": "reaching",
+    "219": "reading",
+    "220": "removing",
+    "221": "repairing",
+    "222": "resting",
+    "223": "riding",
+    "224": "rinsing",
+    "225": "rising",
+    "226": "roaring",
+    "227": "rocking",
+    "228": "rolling",
+    "229": "rowing",
+    "230": "rubbing",
+    "231": "running",
+    "232": "sailing",
+    "233": "saluting",
+    "234": "sanding",
+    "235": "sawing",
+    "236": "scratching",
+    "237": "screwing",
+    "238": "scrubbing",
+    "239": "selling",
+    "240": "serving",
+    "241": "sewing",
+    "242": "shaking",
+    "243": "shaving",
+    "244": "shooting",
+    "245": "shopping",
+    "246": "shouting",
+    "247": "shoveling",
+    "248": "shredding",
+    "249": "shrugging",
+    "250": "signing",
+    "251": "singing",
+    "252": "sitting",
+    "253": "skating",
+    "254": "sketching",
+    "255": "skiing",
+    "256": "skipping",
+    "257": "slapping",
+    "258": "sleeping",
+    "259": "slicing",
+    "260": "sliding",
+    "261": "slipping",
+    "262": "smashing",
+    "263": "smelling",
+    "264": "smiling",
+    "265": "smoking",
+    "266": "snapping",
+    "267": "sneezing",
+    "268": "sniffing",
+    "269": "snowing",
+    "270": "snuggling",
+    "271": "socializing",
+    "272": "sowing",
+    "273": "speaking",
+    "274": "spilling",
+    "275": "spinning",
+    "276": "spitting",
+    "277": "splashing",
+    "278": "spraying",
+    "279": "spreading",
+    "280": "sprinkling",
+    "281": "sprinting",
+    "282": "squatting",
+    "283": "squinting",
+    "284": "stacking",
+    "285": "standing",
+    "286": "starting",
+    "287": "stealing",
+    "288": "steering",
+    "289": "stirring",
+    "290": "stitching",
+    "291": "stomping",
+    "292": "stopping",
+    "293": "storming",
+    "294": "stretching",
+    "295": "stroking",
+    "296": "studying",
+    "297": "submerging",
+    "298": "surfing",
+    "299": "sweeping",
+    "300": "swerving",
+    "301": "swimming",
+    "302": "swinging",
+    "303": "talking",
+    "304": "taping",
+    "305": "tapping",
+    "306": "tattooing",
+    "307": "teaching",
+    "308": "tearing",
+    "309": "telephoning",
+    "310": "throwing",
+    "311": "tickling",
+    "312": "towing",
+    "313": "trimming",
+    "314": "tripping",
+    "315": "tuning",
+    "316": "turning",
+    "317": "twisting",
+    "318": "tying",
+    "319": "typing",
+    "320": "unloading",
+    "321": "unpacking",
+    "322": "vacuuming",
+    "323": "waking",
+    "324": "walking",
+    "325": "washing",
+    "326": "watering",
+    "327": "waving",
+    "328": "waxing",
+    "329": "weeding",
+    "330": "welding",
+    "331": "wetting",
+    "332": "whistling",
+    "333": "winking",
+    "334": "working",
+    "335": "wrapping",
+    "336": "wrestling",
+    "337": "writing",
+    "338": "yawning"
+}

transforms.py

@@ -0,0 +1,443 @@
+import torchvision
+import random
+from PIL import Image, ImageOps
+import numpy as np
+import numbers
+import math
+import torch
+
+
+class GroupRandomCrop(object):
+    def __init__(self, size):
+        if isinstance(size, numbers.Number):
+            self.size = (int(size), int(size))
+        else:
+            self.size = size
+
+    def __call__(self, img_group):
+
+        w, h = img_group[0].size
+        th, tw = self.size
+
+        out_images = list()
+
+        x1 = random.randint(0, w - tw)
+        y1 = random.randint(0, h - th)
+
+        for img in img_group:
+            assert(img.size[0] == w and img.size[1] == h)
+            if w == tw and h == th:
+                out_images.append(img)
+            else:
+                out_images.append(img.crop((x1, y1, x1 + tw, y1 + th)))
+
+        return out_images
+
+
+class MultiGroupRandomCrop(object):
+    def __init__(self, size, groups=1):
+        if isinstance(size, numbers.Number):
+            self.size = (int(size), int(size))
+        else:
+            self.size = size
+        self.groups = groups
+
+    def __call__(self, img_group):
+
+        w, h = img_group[0].size
+        th, tw = self.size
+
+        out_images = list()
+
+        for i in range(self.groups):
+            x1 = random.randint(0, w - tw)
+            y1 = random.randint(0, h - th)
+
+            for img in img_group:
+                assert(img.size[0] == w and img.size[1] == h)
+                if w == tw and h == th:
+                    out_images.append(img)
+                else:
+                    out_images.append(img.crop((x1, y1, x1 + tw, y1 + th)))
+
+        return out_images
+
+
+class GroupCenterCrop(object):
+    def __init__(self, size):
+        self.worker = torchvision.transforms.CenterCrop(size)
+
+    def __call__(self, img_group):
+        return [self.worker(img) for img in img_group]
+
+
+class GroupRandomHorizontalFlip(object):
+    """Randomly horizontally flips the given PIL.Image with a probability of 0.5
+    """
+
+    def __init__(self, is_flow=False):
+        self.is_flow = is_flow
+
+    def __call__(self, img_group, is_flow=False):
+        v = random.random()
+        if v < 0.5:
+            ret = [img.transpose(Image.FLIP_LEFT_RIGHT) for img in img_group]
+            if self.is_flow:
+                for i in range(0, len(ret), 2):
+                    # invert flow pixel values when flipping
+                    ret[i] = ImageOps.invert(ret[i])
+            return ret
+        else:
+            return img_group
+
+
+class GroupNormalize(object):
+    def __init__(self, mean, std):
+        self.mean = mean
+        self.std = std
+
+    def __call__(self, tensor):
+        rep_mean = self.mean * (tensor.size()[0] // len(self.mean))
+        rep_std = self.std * (tensor.size()[0] // len(self.std))
+
+        # TODO: make efficient
+        for t, m, s in zip(tensor, rep_mean, rep_std):
+            t.sub_(m).div_(s)
+
+        return tensor
+
+
+class GroupScale(object):
+    """ Rescales the input PIL.Image to the given 'size'.
+    'size' will be the size of the smaller edge.
+    For example, if height > width, then image will be
+    rescaled to (size * height / width, size)
+    size: size of the smaller edge
+    interpolation: Default: PIL.Image.BILINEAR
+    """
+
+    def __init__(self, size, interpolation=Image.BILINEAR):
+        self.worker = torchvision.transforms.Resize(size, interpolation)
+
+    def __call__(self, img_group):
+        return [self.worker(img) for img in img_group]
+
+
+class GroupOverSample(object):
+    def __init__(self, crop_size, scale_size=None, flip=True):
+        self.crop_size = crop_size if not isinstance(
+            crop_size, int) else (crop_size, crop_size)
+
+        if scale_size is not None:
+            self.scale_worker = GroupScale(scale_size)
+        else:
+            self.scale_worker = None
+        self.flip = flip
+
+    def __call__(self, img_group):
+
+        if self.scale_worker is not None:
+            img_group = self.scale_worker(img_group)
+
+        image_w, image_h = img_group[0].size
+        crop_w, crop_h = self.crop_size
+
+        offsets = GroupMultiScaleCrop.fill_fix_offset(
+            False, image_w, image_h, crop_w, crop_h)
+        oversample_group = list()
+        for o_w, o_h in offsets:
+            normal_group = list()
+            flip_group = list()
+            for i, img in enumerate(img_group):
+                crop = img.crop((o_w, o_h, o_w + crop_w, o_h + crop_h))
+                normal_group.append(crop)
+                flip_crop = crop.copy().transpose(Image.FLIP_LEFT_RIGHT)
+
+                if img.mode == 'L' and i % 2 == 0:
+                    flip_group.append(ImageOps.invert(flip_crop))
+                else:
+                    flip_group.append(flip_crop)
+
+            oversample_group.extend(normal_group)
+            if self.flip:
+                oversample_group.extend(flip_group)
+        return oversample_group
+
+
+class GroupFullResSample(object):
+    def __init__(self, crop_size, scale_size=None, flip=True):
+        self.crop_size = crop_size if not isinstance(
+            crop_size, int) else (crop_size, crop_size)
+
+        if scale_size is not None:
+            self.scale_worker = GroupScale(scale_size)
+        else:
+            self.scale_worker = None
+        self.flip = flip
+
+    def __call__(self, img_group):
+
+        if self.scale_worker is not None:
+            img_group = self.scale_worker(img_group)
+
+        image_w, image_h = img_group[0].size
+        crop_w, crop_h = self.crop_size
+
+        w_step = (image_w - crop_w) // 4
+        h_step = (image_h - crop_h) // 4
+
+        offsets = list()
+        offsets.append((0 * w_step, 2 * h_step))  # left
+        offsets.append((4 * w_step, 2 * h_step))  # right
+        offsets.append((2 * w_step, 2 * h_step))  # center
+
+        oversample_group = list()
+        for o_w, o_h in offsets:
+            normal_group = list()
+            flip_group = list()
+            for i, img in enumerate(img_group):
+                crop = img.crop((o_w, o_h, o_w + crop_w, o_h + crop_h))
+                normal_group.append(crop)
+                if self.flip:
+                    flip_crop = crop.copy().transpose(Image.FLIP_LEFT_RIGHT)
+
+                    if img.mode == 'L' and i % 2 == 0:
+                        flip_group.append(ImageOps.invert(flip_crop))
+                    else:
+                        flip_group.append(flip_crop)
+
+            oversample_group.extend(normal_group)
+            oversample_group.extend(flip_group)
+        return oversample_group
+
+
+class GroupMultiScaleCrop(object):
+
+    def __init__(self, input_size, scales=None, max_distort=1,
+                 fix_crop=True, more_fix_crop=True):
+        self.scales = scales if scales is not None else [1, .875, .75, .66]
+        self.max_distort = max_distort
+        self.fix_crop = fix_crop
+        self.more_fix_crop = more_fix_crop
+        self.input_size = input_size if not isinstance(input_size, int) else [
+            input_size, input_size]
+        self.interpolation = Image.BILINEAR
+
+    def __call__(self, img_group):
+
+        im_size = img_group[0].size
+
+        crop_w, crop_h, offset_w, offset_h = self._sample_crop_size(im_size)
+        crop_img_group = [
+            img.crop(
+                (offset_w,
+                 offset_h,
+                 offset_w +
+                 crop_w,
+                 offset_h +
+                 crop_h)) for img in img_group]
+        ret_img_group = [img.resize((self.input_size[0], self.input_size[1]), self.interpolation)
+                         for img in crop_img_group]
+        return ret_img_group
+
+    def _sample_crop_size(self, im_size):
+        image_w, image_h = im_size[0], im_size[1]
+
+        # find a crop size
+        base_size = min(image_w, image_h)
+        crop_sizes = [int(base_size * x) for x in self.scales]
+        crop_h = [
+            self.input_size[1] if abs(
+                x - self.input_size[1]) < 3 else x for x in crop_sizes]
+        crop_w = [
+            self.input_size[0] if abs(
+                x - self.input_size[0]) < 3 else x for x in crop_sizes]
+
+        pairs = []
+        for i, h in enumerate(crop_h):
+            for j, w in enumerate(crop_w):
+                if abs(i - j) <= self.max_distort:
+                    pairs.append((w, h))
+
+        crop_pair = random.choice(pairs)
+        if not self.fix_crop:
+            w_offset = random.randint(0, image_w - crop_pair[0])
+            h_offset = random.randint(0, image_h - crop_pair[1])
+        else:
+            w_offset, h_offset = self._sample_fix_offset(
+                image_w, image_h, crop_pair[0], crop_pair[1])
+
+        return crop_pair[0], crop_pair[1], w_offset, h_offset
+
+    def _sample_fix_offset(self, image_w, image_h, crop_w, crop_h):
+        offsets = self.fill_fix_offset(
+            self.more_fix_crop, image_w, image_h, crop_w, crop_h)
+        return random.choice(offsets)
+
+    @staticmethod
+    def fill_fix_offset(more_fix_crop, image_w, image_h, crop_w, crop_h):
+        w_step = (image_w - crop_w) // 4
+        h_step = (image_h - crop_h) // 4
+
+        ret = list()
+        ret.append((0, 0))  # upper left
+        ret.append((4 * w_step, 0))  # upper right
+        ret.append((0, 4 * h_step))  # lower left
+        ret.append((4 * w_step, 4 * h_step))  # lower right
+        ret.append((2 * w_step, 2 * h_step))  # center
+
+        if more_fix_crop:
+            ret.append((0, 2 * h_step))  # center left
+            ret.append((4 * w_step, 2 * h_step))  # center right
+            ret.append((2 * w_step, 4 * h_step))  # lower center
+            ret.append((2 * w_step, 0 * h_step))  # upper center
+
+            ret.append((1 * w_step, 1 * h_step))  # upper left quarter
+            ret.append((3 * w_step, 1 * h_step))  # upper right quarter
+            ret.append((1 * w_step, 3 * h_step))  # lower left quarter
+            ret.append((3 * w_step, 3 * h_step))  # lower righ quarter
+
+        return ret
+
+
+class GroupRandomSizedCrop(object):
+    """Random crop the given PIL.Image to a random size of (0.08 to 1.0) of the original size
+    and and a random aspect ratio of 3/4 to 4/3 of the original aspect ratio
+    This is popularly used to train the Inception networks
+    size: size of the smaller edge
+    interpolation: Default: PIL.Image.BILINEAR
+    """
+
+    def __init__(self, size, interpolation=Image.BILINEAR):
+        self.size = size
+        self.interpolation = interpolation
+
+    def __call__(self, img_group):
+        for attempt in range(10):
+            area = img_group[0].size[0] * img_group[0].size[1]
+            target_area = random.uniform(0.08, 1.0) * area
+            aspect_ratio = random.uniform(3. / 4, 4. / 3)
+
+            w = int(round(math.sqrt(target_area * aspect_ratio)))
+            h = int(round(math.sqrt(target_area / aspect_ratio)))
+
+            if random.random() < 0.5:
+                w, h = h, w
+
+            if w <= img_group[0].size[0] and h <= img_group[0].size[1]:
+                x1 = random.randint(0, img_group[0].size[0] - w)
+                y1 = random.randint(0, img_group[0].size[1] - h)
+                found = True
+                break
+        else:
+            found = False
+            x1 = 0
+            y1 = 0
+
+        if found:
+            out_group = list()
+            for img in img_group:
+                img = img.crop((x1, y1, x1 + w, y1 + h))
+                assert(img.size == (w, h))
+                out_group.append(
+                    img.resize(
+                        (self.size, self.size), self.interpolation))
+            return out_group
+        else:
+            # Fallback
+            scale = GroupScale(self.size, interpolation=self.interpolation)
+            crop = GroupRandomCrop(self.size)
+            return crop(scale(img_group))
+
+
+class ConvertDataFormat(object):
+    def __init__(self, model_type):
+        self.model_type = model_type
+
+    def __call__(self, images):
+        if self.model_type == '2D':
+            return images
+        tc, h, w = images.size()
+        t = tc // 3
+        images = images.view(t, 3, h, w)
+        images = images.permute(1, 0, 2, 3)
+        return images
+
+
+class Stack(object):
+
+    def __init__(self, roll=False):
+        self.roll = roll
+
+    def __call__(self, img_group):
+        if img_group[0].mode == 'L':
+            return np.concatenate([np.expand_dims(x, 2)
+                                   for x in img_group], axis=2)
+        elif img_group[0].mode == 'RGB':
+            if self.roll:
+                return np.concatenate([np.array(x)[:, :, ::-1]
+                                       for x in img_group], axis=2)
+            else:
+                #print(np.concatenate(img_group, axis=2).shape)
+                # print(img_group[0].shape)
+                return np.concatenate(img_group, axis=2)
+
+
+class ToTorchFormatTensor(object):
+    """ Converts a PIL.Image (RGB) or numpy.ndarray (H x W x C) in the range [0, 255]
+    to a torch.FloatTensor of shape (C x H x W) in the range [0.0, 1.0] """
+
+    def __init__(self, div=True):
+        self.div = div
+
+    def __call__(self, pic):
+        if isinstance(pic, np.ndarray):
+            # handle numpy array
+            img = torch.from_numpy(pic).permute(2, 0, 1).contiguous()
+        else:
+            # handle PIL Image
+            img = torch.ByteTensor(
+                torch.ByteStorage.from_buffer(
+                    pic.tobytes()))
+            img = img.view(pic.size[1], pic.size[0], len(pic.mode))
+            # put it from HWC to CHW format
+            # yikes, this transpose takes 80% of the loading time/CPU
+            img = img.transpose(0, 1).transpose(0, 2).contiguous()
+        return img.float().div(255) if self.div else img.float()
+
+
+class IdentityTransform(object):
+
+    def __call__(self, data):
+        return data
+
+
+if __name__ == "__main__":
+    trans = torchvision.transforms.Compose([
+        GroupScale(256),
+        GroupRandomCrop(224),
+        Stack(),
+        ToTorchFormatTensor(),
+        GroupNormalize(
+            mean=[.485, .456, .406],
+            std=[.229, .224, .225]
+        )]
+    )
+
+    im = Image.open('../tensorflow-model-zoo.torch/lena_299.png')
+
+    color_group = [im] * 3
+    rst = trans(color_group)
+
+    gray_group = [im.convert('L')] * 9
+    gray_rst = trans(gray_group)
+
+    trans2 = torchvision.transforms.Compose([
+        GroupRandomSizedCrop(256),
+        Stack(),
+        ToTorchFormatTensor(),
+        GroupNormalize(
+            mean=[.485, .456, .406],
+            std=[.229, .224, .225])
+    ])
+    print(trans2(color_group))

uniformerv2.py

@@ -0,0 +1,510 @@
+#!/usr/bin/env python
+import os
+from collections import OrderedDict
+
+from timm.models.layers import DropPath
+import torch
+from torch import nn
+from torch.nn import MultiheadAttention
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+
+MODEL_PATH = './'
+_MODELS = {
+    "ViT-B/16": os.path.join(MODEL_PATH, "vit_b16.pth"),
+    "ViT-L/14": os.path.join(MODEL_PATH, "vit_l14.pth"),
+    "ViT-L/14_336": os.path.join(MODEL_PATH, "vit_l14_336.pth"),
+}
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x):
+        orig_type = x.dtype
+        ret = super().forward(x.type(torch.float32))
+        return ret.type(orig_type)
+
+
+class QuickGELU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class Local_MHRA(nn.Module):
+    def __init__(self, d_model, dw_reduction=1.5, pos_kernel_size=3):
+        super().__init__() 
+
+        padding = pos_kernel_size // 2
+        re_d_model = int(d_model // dw_reduction)
+        self.pos_embed = nn.Sequential(
+            nn.BatchNorm3d(d_model),
+            nn.Conv3d(d_model, re_d_model, kernel_size=1, stride=1, padding=0),
+            nn.Conv3d(re_d_model, re_d_model, kernel_size=(pos_kernel_size, 1, 1), stride=(1, 1, 1), padding=(padding, 0, 0), groups=re_d_model),
+            nn.Conv3d(re_d_model, d_model, kernel_size=1, stride=1, padding=0),
+        )
+
+        # init zero
+        print('Init zero for Conv in pos_emb')
+        nn.init.constant_(self.pos_embed[3].weight, 0)
+        nn.init.constant_(self.pos_embed[3].bias, 0)
+
+    def forward(self, x):
+        return self.pos_embed(x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+            self, d_model, n_head, attn_mask=None, drop_path=0.0, 
+            dw_reduction=1.5, no_lmhra=False, double_lmhra=True
+        ):
+        super().__init__() 
+        
+        self.n_head = n_head
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        print(f'Drop path rate: {drop_path}')
+
+        self.no_lmhra = no_lmhra
+        self.double_lmhra = double_lmhra
+        print(f'No L_MHRA: {no_lmhra}')
+        print(f'Double L_MHRA: {double_lmhra}')
+        if not no_lmhra:
+            self.lmhra1 = Local_MHRA(d_model, dw_reduction=dw_reduction)
+            if double_lmhra:
+                self.lmhra2 = Local_MHRA(d_model, dw_reduction=dw_reduction)
+
+        # spatial
+        self.attn = MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_model * 4)),
+            ("gelu", QuickGELU()),
+            ("c_proj", nn.Linear(d_model * 4, d_model))
+        ]))
+        self.ln_2 = LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+    def attention(self, x):
+        self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
+        return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
+
+    def forward(self, x, T=8, use_checkpoint=False):
+        # x: 1+HW, NT, C
+        if not self.no_lmhra:
+            # Local MHRA
+            tmp_x = x[1:, :, :]
+            L, NT, C = tmp_x.shape
+            N = NT // T
+            H = W = int(L ** 0.5)
+            tmp_x = tmp_x.view(H, W, N, T, C).permute(2, 4, 3, 0, 1).contiguous()
+            tmp_x = tmp_x + self.drop_path(self.lmhra1(tmp_x))
+            tmp_x = tmp_x.view(N, C, T, L).permute(3, 0, 2, 1).contiguous().view(L, NT, C)
+            x = torch.cat([x[:1, :, :], tmp_x], dim=0)
+        # MHSA
+        if use_checkpoint:
+            attn_out = checkpoint.checkpoint(self.attention, self.ln_1(x))
+            x = x + self.drop_path(attn_out)
+        else:
+            x = x + self.drop_path(self.attention(self.ln_1(x)))
+        # Local MHRA
+        if not self.no_lmhra and self.double_lmhra:
+            tmp_x = x[1:, :, :]
+            tmp_x = tmp_x.view(H, W, N, T, C).permute(2, 4, 3, 0, 1).contiguous()
+            tmp_x = tmp_x + self.drop_path(self.lmhra2(tmp_x))
+            tmp_x = tmp_x.view(N, C, T, L).permute(3, 0, 2, 1).contiguous().view(L, NT, C)
+            x = torch.cat([x[:1, :, :], tmp_x], dim=0)
+        # FFN
+        if use_checkpoint:
+            mlp_out = checkpoint.checkpoint(self.mlp, self.ln_2(x))
+            x = x + self.drop_path(mlp_out)
+        else:
+            x = x + self.drop_path(self.mlp(self.ln_2(x)))
+        return x
+
+
+class Extractor(nn.Module):
+    def __init__(
+            self, d_model, n_head, attn_mask=None,
+            mlp_factor=4.0, dropout=0.0, drop_path=0.0,
+        ):
+        super().__init__()
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        print(f'Drop path rate: {drop_path}')
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = nn.LayerNorm(d_model)
+        d_mlp = round(mlp_factor * d_model)
+        self.mlp = nn.Sequential(OrderedDict([
+            ("c_fc", nn.Linear(d_model, d_mlp)),
+            ("gelu", QuickGELU()),
+            ("dropout", nn.Dropout(dropout)),
+            ("c_proj", nn.Linear(d_mlp, d_model))
+        ]))
+        self.ln_2 = nn.LayerNorm(d_model)
+        self.ln_3 = nn.LayerNorm(d_model)
+        self.attn_mask = attn_mask
+
+        # zero init
+        nn.init.xavier_uniform_(self.attn.in_proj_weight)
+        nn.init.constant_(self.attn.out_proj.weight, 0.)
+        nn.init.constant_(self.attn.out_proj.bias, 0.)
+        nn.init.xavier_uniform_(self.mlp[0].weight)
+        nn.init.constant_(self.mlp[-1].weight, 0.)
+        nn.init.constant_(self.mlp[-1].bias, 0.)
+
+    def attention(self, x, y):
+        d_model = self.ln_1.weight.size(0)
+        q = (x @ self.attn.in_proj_weight[:d_model].T) + self.attn.in_proj_bias[:d_model]
+
+        k = (y @ self.attn.in_proj_weight[d_model:-d_model].T) + self.attn.in_proj_bias[d_model:-d_model]
+        v = (y @ self.attn.in_proj_weight[-d_model:].T) + self.attn.in_proj_bias[-d_model:]
+        Tx, Ty, N = q.size(0), k.size(0), q.size(1)
+        q = q.view(Tx, N, self.attn.num_heads, self.attn.head_dim).permute(1, 2, 0, 3)
+        k = k.view(Ty, N, self.attn.num_heads, self.attn.head_dim).permute(1, 2, 0, 3)
+        v = v.view(Ty, N, self.attn.num_heads, self.attn.head_dim).permute(1, 2, 0, 3)
+        aff = (q @ k.transpose(-2, -1) / (self.attn.head_dim ** 0.5))
+
+        aff = aff.softmax(dim=-1)
+        out = aff @ v
+        out = out.permute(2, 0, 1, 3).flatten(2)
+        out = self.attn.out_proj(out)
+        return out
+
+    def forward(self, x, y):
+        x = x + self.drop_path(self.attention(self.ln_1(x), self.ln_3(y)))
+        x = x + self.drop_path(self.mlp(self.ln_2(x)))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+            self, width, layers, heads, attn_mask=None, backbone_drop_path_rate=0., 
+            use_checkpoint=False, checkpoint_num=[0], t_size=8, dw_reduction=2,
+            no_lmhra=False, double_lmhra=True,
+            return_list=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],
+            n_layers=12, n_dim=768, n_head=12, mlp_factor=4.0, drop_path_rate=0.,
+            mlp_dropout=[0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], 
+            cls_dropout=0.5, num_classes=400,
+        ):
+        super().__init__()
+        self.T = t_size
+        self.return_list = return_list
+        # backbone
+        b_dpr = [x.item() for x in torch.linspace(0, backbone_drop_path_rate, layers)]
+        self.resblocks = nn.ModuleList([
+            ResidualAttentionBlock(
+                width, heads, attn_mask, 
+                drop_path=b_dpr[i],
+                dw_reduction=dw_reduction,
+                no_lmhra=no_lmhra,
+                double_lmhra=double_lmhra,
+            ) for i in range(layers)
+        ])
+        # checkpoint
+        self.use_checkpoint = use_checkpoint
+        self.checkpoint_num = checkpoint_num
+        self.n_layers = n_layers
+        print(f'Use checkpoint: {self.use_checkpoint}')
+        print(f'Checkpoint number: {self.checkpoint_num}')
+
+        # global block
+        assert n_layers == len(return_list)
+        if n_layers > 0:
+            self.temporal_cls_token = nn.Parameter(torch.zeros(1, 1, n_dim))
+            self.dpe = nn.ModuleList([
+                nn.Conv3d(n_dim, n_dim, kernel_size=3, stride=1, padding=1, bias=True, groups=n_dim)
+                for i in range(n_layers)
+            ])
+            for m in self.dpe:
+                nn.init.constant_(m.bias, 0.)
+            dpr = [x.item() for x in torch.linspace(0, drop_path_rate, n_layers)]
+            self.dec = nn.ModuleList([
+                Extractor(
+                    n_dim, n_head, mlp_factor=mlp_factor, 
+                    dropout=mlp_dropout[i], drop_path=dpr[i],
+                ) for i in range(n_layers)
+            ])
+            self.balance = nn.Parameter(torch.zeros((n_dim)))
+            self.sigmoid = nn.Sigmoid()
+        # projection
+        self.proj = nn.Sequential(
+            nn.LayerNorm(n_dim),
+            nn.Dropout(cls_dropout),
+            nn.Linear(n_dim, num_classes),
+        )
+
+    def forward(self, x):
+        T_down = self.T
+        L, NT, C = x.shape
+        N = NT // T_down
+        H = W = int((L - 1) ** 0.5)
+
+        if self.n_layers > 0:
+            cls_token = self.temporal_cls_token.repeat(1, N, 1)
+
+        j = -1
+        for i, resblock in enumerate(self.resblocks):
+            if self.use_checkpoint and i < self.checkpoint_num[0]:
+                x = resblock(x, self.T, use_checkpoint=True)
+            else:
+                x = resblock(x, T_down)
+            if i in self.return_list:
+                j += 1
+                tmp_x = x.clone()
+                tmp_x = tmp_x.view(L, N, T_down, C)
+                # dpe
+                _, tmp_feats = tmp_x[:1], tmp_x[1:]
+                tmp_feats = tmp_feats.permute(1, 3, 2, 0).reshape(N, C, T_down, H, W)
+                tmp_feats = self.dpe[j](tmp_feats).view(N, C, T_down, L - 1).permute(3, 0, 2, 1).contiguous()
+                tmp_x[1:] = tmp_x[1:] + tmp_feats
+                # global block
+                tmp_x = tmp_x.permute(2, 0, 1, 3).flatten(0, 1)  # T * L, N, C
+                cls_token = self.dec[j](cls_token, tmp_x)
+
+        if self.n_layers > 0:
+            weight = self.sigmoid(self.balance)
+            residual = x.view(L, N, T_down, C)[0].mean(1)  # L, N, T, C
+            return self.proj((1 - weight) * cls_token[0, :, :] + weight * residual)
+        else:
+            residual = x.view(L, N, T_down, C)[0].mean(1)  # L, N, T, C
+            return self.proj(residual)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(
+        self, 
+        # backbone
+        input_resolution, patch_size, width, layers, heads, output_dim, backbone_drop_path_rate=0.,
+        use_checkpoint=False, checkpoint_num=[0], t_size=8, kernel_size=3, dw_reduction=1.5,
+        temporal_downsample=True,
+        no_lmhra=-False, double_lmhra=True,
+        # global block
+        return_list=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11],
+        n_layers=12, n_dim=768, n_head=12, mlp_factor=4.0, drop_path_rate=0.,
+        mlp_dropout=[0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], 
+        cls_dropout=0.5, num_classes=400,
+    ):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.output_dim = output_dim
+        padding = (kernel_size - 1) // 2
+        if temporal_downsample:
+            self.conv1 = nn.Conv3d(3, width, (kernel_size, patch_size, patch_size), (2, patch_size, patch_size), (padding, 0, 0), bias=False)
+            t_size = t_size // 2
+        else:
+            self.conv1 = nn.Conv3d(3, width, (1, patch_size, patch_size), (1, patch_size, patch_size), (0, 0, 0), bias=False)
+
+        scale = width ** -0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
+        self.ln_pre = LayerNorm(width)
+        
+        self.transformer = Transformer(
+            width, layers, heads, dw_reduction=dw_reduction, 
+            backbone_drop_path_rate=backbone_drop_path_rate, 
+            use_checkpoint=use_checkpoint, checkpoint_num=checkpoint_num, t_size=t_size,
+            no_lmhra=no_lmhra, double_lmhra=double_lmhra,
+            return_list=return_list, n_layers=n_layers, n_dim=n_dim, n_head=n_head, 
+            mlp_factor=mlp_factor, drop_path_rate=drop_path_rate, mlp_dropout=mlp_dropout, 
+            cls_dropout=cls_dropout, num_classes=num_classes,
+        )
+
+    def forward(self, x):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        N, C, T, H, W = x.shape
+        x = x.permute(0, 2, 3, 4, 1).reshape(N * T, H * W, C)
+        
+        x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1)  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        out = self.transformer(x)
+        return out
+
+
+def inflate_weight(weight_2d, time_dim, center=True):
+    print(f'Init center: {center}')
+    if center:
+        weight_3d = torch.zeros(*weight_2d.shape)
+        weight_3d = weight_3d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
+        middle_idx = time_dim // 2
+        weight_3d[:, :, middle_idx, :, :] = weight_2d
+    else:
+        weight_3d = weight_2d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
+        weight_3d = weight_3d / time_dim
+    return weight_3d
+
+
+def load_state_dict(model, state_dict):
+    state_dict_3d = model.state_dict()
+    for k in state_dict.keys():
+        if state_dict[k].shape != state_dict_3d[k].shape:
+            if len(state_dict_3d[k].shape) <= 2:
+                print(f'Ignore: {k}')
+                continue
+            print(f'Inflate: {k}, {state_dict[k].shape} => {state_dict_3d[k].shape}')
+            time_dim = state_dict_3d[k].shape[2]
+            state_dict[k] = inflate_weight(state_dict[k], time_dim)
+    model.load_state_dict(state_dict, strict=False)
+
+
+def uniformerv2_b16(
+    pretrained=True, use_checkpoint=False, checkpoint_num=[0],
+    t_size=16, dw_reduction=1.5, backbone_drop_path_rate=0., 
+    temporal_downsample=True,
+    no_lmhra=False, double_lmhra=True,
+    return_list=[8, 9, 10, 11], 
+    n_layers=4, n_dim=768, n_head=12, mlp_factor=4.0, drop_path_rate=0.,
+    mlp_dropout=[0.5, 0.5, 0.5, 0.5], 
+    cls_dropout=0.5, num_classes=400,
+):
+    model = VisionTransformer(
+        input_resolution=224,
+        patch_size=16,
+        width=768,
+        layers=12,
+        heads=12,
+        output_dim=512,
+        use_checkpoint=use_checkpoint,
+        checkpoint_num=checkpoint_num,
+        t_size=t_size,
+        dw_reduction=dw_reduction, 
+        backbone_drop_path_rate=backbone_drop_path_rate, 
+        temporal_downsample=temporal_downsample,
+        no_lmhra=no_lmhra,
+        double_lmhra=double_lmhra,
+        return_list=return_list, 
+        n_layers=n_layers, 
+        n_dim=n_dim, 
+        n_head=n_head, 
+        mlp_factor=mlp_factor, 
+        drop_path_rate=drop_path_rate, 
+        mlp_dropout=mlp_dropout, 
+        cls_dropout=cls_dropout, 
+        num_classes=num_classes,
+    )
+
+    if pretrained:
+        print('load pretrained weights')
+        state_dict = torch.load(_MODELS["ViT-B/16"], map_location='cpu')
+        load_state_dict(model, state_dict)
+    return model.eval()
+
+
+def uniformerv2_l14(
+    pretrained=True, use_checkpoint=False, checkpoint_num=[0],
+    t_size=16, dw_reduction=1.5, backbone_drop_path_rate=0., 
+    temporal_downsample=True,
+    no_lmhra=False, double_lmhra=True,
+    return_list=[20, 21, 22, 23],
+    n_layers=4, n_dim=1024, n_head=16, mlp_factor=4.0, drop_path_rate=0.,
+    mlp_dropout=[0.5, 0.5, 0.5, 0.5], 
+    cls_dropout=0.5, num_classes=400,
+):
+    model = VisionTransformer(
+        input_resolution=224,
+        patch_size=14,
+        width=1024,
+        layers=24,
+        heads=16,
+        output_dim=768,
+        use_checkpoint=use_checkpoint,
+        checkpoint_num=checkpoint_num,
+        t_size=t_size,
+        dw_reduction=dw_reduction, 
+        backbone_drop_path_rate=backbone_drop_path_rate, 
+        temporal_downsample=temporal_downsample,
+        no_lmhra=no_lmhra,
+        double_lmhra=double_lmhra,
+        return_list=return_list, 
+        n_layers=n_layers, 
+        n_dim=n_dim, 
+        n_head=n_head, 
+        mlp_factor=mlp_factor, 
+        drop_path_rate=drop_path_rate, 
+        mlp_dropout=mlp_dropout, 
+        cls_dropout=cls_dropout, 
+        num_classes=num_classes,
+    )
+
+    if pretrained:
+        print('load pretrained weights')
+        state_dict = torch.load(_MODELS["ViT-L/14"], map_location='cpu')
+        load_state_dict(model, state_dict)
+    return model.eval()
+
+
+def uniformerv2_l14_336(
+    pretrained=True, use_checkpoint=False, checkpoint_num=[0],
+    t_size=16, dw_reduction=1.5, backbone_drop_path_rate=0., 
+    no_temporal_downsample=True,
+    no_lmhra=False, double_lmhra=True,
+    return_list=[20, 21, 22, 23],
+    n_layers=4, n_dim=1024, n_head=16, mlp_factor=4.0, drop_path_rate=0.,
+    mlp_dropout=[0.5, 0.5, 0.5, 0.5], 
+    cls_dropout=0.5, num_classes=400,
+):
+    model = VisionTransformer(
+        input_resolution=336,
+        patch_size=14,
+        width=1024,
+        layers=24,
+        heads=16,
+        output_dim=768,
+        use_checkpoint=use_checkpoint,
+        checkpoint_num=checkpoint_num,
+        t_size=t_size,
+        dw_reduction=dw_reduction, 
+        backbone_drop_path_rate=backbone_drop_path_rate, 
+        no_temporal_downsample=no_temporal_downsample,
+        no_lmhra=no_lmhra,
+        double_lmhra=double_lmhra,
+        return_list=return_list, 
+        n_layers=n_layers, 
+        n_dim=n_dim, 
+        n_head=n_head, 
+        mlp_factor=mlp_factor, 
+        drop_path_rate=drop_path_rate, 
+        mlp_dropout=mlp_dropout, 
+        cls_dropout=cls_dropout, 
+        num_classes=num_classes,
+    )
+
+    if pretrained:
+        print('load pretrained weights')
+        state_dict = torch.load(_MODELS["ViT-L/14_336"], map_location='cpu')
+        load_state_dict(model, state_dict)
+    return model.eval()
+
+
+if __name__ == '__main__':
+    import time
+    from fvcore.nn import FlopCountAnalysis
+    from fvcore.nn import flop_count_table
+    import numpy as np
+
+    seed = 4217
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)
+    num_frames = 16
+
+    model = uniformerv2_l14(
+        pretrained=False, 
+        t_size=num_frames, backbone_drop_path_rate=0., drop_path_rate=0.,
+        dw_reduction=1.5,
+        no_lmhra=False,
+        temporal_downsample=True,
+        return_list=[8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23],
+        mlp_dropout=[0.5]*16,
+        n_layers=16
+    )
+    print(model)
+
+    flops = FlopCountAnalysis(model, torch.rand(1, 3, num_frames, 224, 224))
+    s = time.time()
+    print(flop_count_table(flops, max_depth=1))
+    print(time.time()-s)