init video

app.py

@@ -0,0 +1,125 @@
+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 uniformer_light_video import uniformer_xxs_video
+from kinetics_class_index import kinetics_classnames
+from transforms import (
+    GroupNormalize, GroupScale, GroupCenterCrop, 
+    Stack, ToTorchFormatTensor
+)
+
+import gradio as gr
+from huggingface_hub import hf_hub_download
+
+
+# Device on which to run the model
+# Set to cuda to load on GPU
+device = "cpu"
+model_path = hf_hub_download(repo_id="Andy1621/uniformer_light", filename="uniformer_xxs16_160_k400.pth")
+# Pick a pretrained model 
+model = uniformer_xxs_video()
+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
+kinetics_id_to_classname = {}
+for k, v in kinetics_classnames.items():
+    kinetics_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 = 160
+    scale_size = 160
+    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)
+    
+    with torch.no_grad():
+        prediction = model(inputs)
+        prediction = F.softmax(prediction, dim=1).flatten()
+
+    return {kinetics_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(
+        """
+        # UniFormer Light
+        Gradio demo for <a href='https://github.com/Sense-X/UniFormer' target='_blank'>UniFormer</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=[['./videos/hitting_baseball.mp4'], ['./videos/hoverboarding.mp4'], ['./videos/yoga.mp4']])
+
+    gr.Markdown(
+        """
+        <p style='text-align: center'><a href='https://arxiv.org/abs/2201.09450' target='_blank'>[TPAMI] UniFormer: Unifying Convolution and Self-attention for Visual Recognition</a> | <a href='https://github.com/Sense-X/UniFormer' 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)

kinetics_class_index.py

@@ -0,0 +1,402 @@
+kinetics_classnames = {
+    "0": "riding a bike", 
+    "1": "marching", 
+    "2": "dodgeball", 
+    "3": "playing cymbals", 
+    "4": "checking tires", 
+    "5": "roller skating", 
+    "6": "tasting beer", 
+    "7": "clapping", 
+    "8": "drawing", 
+    "9": "juggling fire", 
+    "10": "bobsledding", 
+    "11": "petting animal (not cat)", 
+    "12": "spray painting", 
+    "13": "training dog", 
+    "14": "eating watermelon", 
+    "15": "building cabinet", 
+    "16": "applauding", 
+    "17": "playing harp", 
+    "18": "balloon blowing", 
+    "19": "sled dog racing", 
+    "20": "wrestling", 
+    "21": "pole vault", 
+    "22": "hurling (sport)", 
+    "23": "riding scooter", 
+    "24": "shearing sheep", 
+    "25": "sweeping floor", 
+    "26": "eating carrots", 
+    "27": "skateboarding", 
+    "28": "dunking basketball", 
+    "29": "disc golfing", 
+    "30": "eating spaghetti", 
+    "31": "playing flute", 
+    "32": "riding mechanical bull", 
+    "33": "making sushi", 
+    "34": "trapezing", 
+    "35": "picking fruit", 
+    "36": "stretching leg", 
+    "37": "playing ukulele", 
+    "38": "tying tie", 
+    "39": "skydiving", 
+    "40": "playing cello", 
+    "41": "jumping into pool", 
+    "42": "shooting goal (soccer)", 
+    "43": "trimming trees", 
+    "44": "bookbinding", 
+    "45": "ski jumping", 
+    "46": "walking the dog", 
+    "47": "riding unicycle", 
+    "48": "shaving head", 
+    "49": "hopscotch", 
+    "50": "playing piano", 
+    "51": "parasailing", 
+    "52": "bartending", 
+    "53": "kicking field goal", 
+    "54": "finger snapping", 
+    "55": "dining", 
+    "56": "yawning", 
+    "57": "peeling potatoes", 
+    "58": "canoeing or kayaking", 
+    "59": "front raises", 
+    "60": "laughing", 
+    "61": "dancing macarena", 
+    "62": "digging", 
+    "63": "reading newspaper", 
+    "64": "hitting baseball", 
+    "65": "clay pottery making", 
+    "66": "exercising with an exercise ball", 
+    "67": "playing saxophone", 
+    "68": "shooting basketball", 
+    "69": "washing hair", 
+    "70": "lunge", 
+    "71": "brushing hair", 
+    "72": "curling hair", 
+    "73": "kitesurfing", 
+    "74": "tapping guitar", 
+    "75": "bending back", 
+    "76": "skipping rope", 
+    "77": "situp", 
+    "78": "folding paper", 
+    "79": "cracking neck", 
+    "80": "assembling computer", 
+    "81": "cleaning gutters", 
+    "82": "blowing out candles", 
+    "83": "shaking hands", 
+    "84": "dancing gangnam style", 
+    "85": "windsurfing", 
+    "86": "tap dancing", 
+    "87": "skiing (not slalom or crosscountry)", 
+    "88": "bandaging", 
+    "89": "push up", 
+    "90": "doing nails", 
+    "91": "punching person (boxing)", 
+    "92": "bouncing on trampoline", 
+    "93": "scrambling eggs", 
+    "94": "singing", 
+    "95": "cleaning floor", 
+    "96": "krumping", 
+    "97": "drumming fingers", 
+    "98": "snowmobiling", 
+    "99": "gymnastics tumbling", 
+    "100": "headbanging", 
+    "101": "catching or throwing frisbee", 
+    "102": "riding elephant", 
+    "103": "bee keeping", 
+    "104": "feeding birds", 
+    "105": "snatch weight lifting", 
+    "106": "mowing lawn", 
+    "107": "fixing hair", 
+    "108": "playing trumpet", 
+    "109": "flying kite", 
+    "110": "crossing river", 
+    "111": "swinging legs", 
+    "112": "sanding floor", 
+    "113": "belly dancing", 
+    "114": "sneezing", 
+    "115": "clean and jerk", 
+    "116": "side kick", 
+    "117": "filling eyebrows", 
+    "118": "shuffling cards", 
+    "119": "recording music", 
+    "120": "cartwheeling", 
+    "121": "feeding fish", 
+    "122": "folding clothes", 
+    "123": "water skiing", 
+    "124": "tobogganing", 
+    "125": "blowing leaves", 
+    "126": "smoking", 
+    "127": "unboxing", 
+    "128": "tai chi", 
+    "129": "waxing legs", 
+    "130": "riding camel", 
+    "131": "slapping", 
+    "132": "tossing salad", 
+    "133": "capoeira", 
+    "134": "playing cards", 
+    "135": "playing organ", 
+    "136": "playing violin", 
+    "137": "playing drums", 
+    "138": "tapping pen", 
+    "139": "vault", 
+    "140": "shoveling snow", 
+    "141": "playing tennis", 
+    "142": "getting a tattoo", 
+    "143": "making a sandwich", 
+    "144": "making tea", 
+    "145": "grinding meat", 
+    "146": "squat", 
+    "147": "eating doughnuts", 
+    "148": "ice fishing", 
+    "149": "snowkiting", 
+    "150": "kicking soccer ball", 
+    "151": "playing controller", 
+    "152": "giving or receiving award", 
+    "153": "welding", 
+    "154": "throwing discus", 
+    "155": "throwing axe", 
+    "156": "ripping paper", 
+    "157": "swimming butterfly stroke", 
+    "158": "air drumming", 
+    "159": "blowing nose", 
+    "160": "hockey stop", 
+    "161": "taking a shower", 
+    "162": "bench pressing", 
+    "163": "planting trees", 
+    "164": "pumping fist", 
+    "165": "climbing tree", 
+    "166": "tickling", 
+    "167": "high kick", 
+    "168": "waiting in line", 
+    "169": "slacklining", 
+    "170": "tango dancing", 
+    "171": "hurdling", 
+    "172": "carrying baby", 
+    "173": "celebrating", 
+    "174": "sharpening knives", 
+    "175": "passing American football (in game)", 
+    "176": "headbutting", 
+    "177": "playing recorder", 
+    "178": "brush painting", 
+    "179": "garbage collecting", 
+    "180": "robot dancing", 
+    "181": "shredding paper", 
+    "182": "pumping gas", 
+    "183": "rock climbing", 
+    "184": "hula hooping", 
+    "185": "braiding hair", 
+    "186": "opening present", 
+    "187": "texting", 
+    "188": "decorating the christmas tree", 
+    "189": "answering questions", 
+    "190": "playing keyboard", 
+    "191": "writing", 
+    "192": "bungee jumping", 
+    "193": "sniffing", 
+    "194": "eating burger", 
+    "195": "playing accordion", 
+    "196": "making pizza", 
+    "197": "playing volleyball", 
+    "198": "tasting food", 
+    "199": "pushing cart", 
+    "200": "spinning poi", 
+    "201": "cleaning windows", 
+    "202": "arm wrestling", 
+    "203": "changing oil", 
+    "204": "swimming breast stroke", 
+    "205": "tossing coin", 
+    "206": "deadlifting", 
+    "207": "hoverboarding", 
+    "208": "cutting watermelon", 
+    "209": "cheerleading", 
+    "210": "snorkeling", 
+    "211": "washing hands", 
+    "212": "eating cake", 
+    "213": "pull ups", 
+    "214": "surfing water", 
+    "215": "eating hotdog", 
+    "216": "holding snake", 
+    "217": "playing harmonica", 
+    "218": "ironing", 
+    "219": "cutting nails", 
+    "220": "golf chipping", 
+    "221": "shot put", 
+    "222": "hugging", 
+    "223": "playing clarinet", 
+    "224": "faceplanting", 
+    "225": "trimming or shaving beard", 
+    "226": "drinking shots", 
+    "227": "riding mountain bike", 
+    "228": "tying bow tie", 
+    "229": "swinging on something", 
+    "230": "skiing crosscountry", 
+    "231": "unloading truck", 
+    "232": "cleaning pool", 
+    "233": "jogging", 
+    "234": "ice climbing", 
+    "235": "mopping floor", 
+    "236": "making bed", 
+    "237": "diving cliff", 
+    "238": "washing dishes", 
+    "239": "grooming dog", 
+    "240": "weaving basket", 
+    "241": "frying vegetables", 
+    "242": "stomping grapes", 
+    "243": "moving furniture", 
+    "244": "cooking sausages", 
+    "245": "doing laundry", 
+    "246": "dying hair", 
+    "247": "knitting", 
+    "248": "reading book", 
+    "249": "baby waking up", 
+    "250": "punching bag", 
+    "251": "surfing crowd", 
+    "252": "cooking chicken", 
+    "253": "pushing car", 
+    "254": "springboard diving", 
+    "255": "swing dancing", 
+    "256": "massaging legs", 
+    "257": "beatboxing", 
+    "258": "breading or breadcrumbing", 
+    "259": "somersaulting", 
+    "260": "brushing teeth", 
+    "261": "stretching arm", 
+    "262": "juggling balls", 
+    "263": "massaging person's head", 
+    "264": "eating ice cream", 
+    "265": "extinguishing fire", 
+    "266": "hammer throw", 
+    "267": "whistling", 
+    "268": "crawling baby", 
+    "269": "using remote controller (not gaming)", 
+    "270": "playing cricket", 
+    "271": "opening bottle", 
+    "272": "playing xylophone", 
+    "273": "motorcycling", 
+    "274": "driving car", 
+    "275": "exercising arm", 
+    "276": "passing American football (not in game)", 
+    "277": "playing kickball", 
+    "278": "sticking tongue out", 
+    "279": "flipping pancake", 
+    "280": "catching fish", 
+    "281": "eating chips", 
+    "282": "shaking head", 
+    "283": "sword fighting", 
+    "284": "playing poker", 
+    "285": "cooking on campfire", 
+    "286": "doing aerobics", 
+    "287": "paragliding", 
+    "288": "using segway", 
+    "289": "folding napkins", 
+    "290": "playing bagpipes", 
+    "291": "gargling", 
+    "292": "skiing slalom", 
+    "293": "strumming guitar", 
+    "294": "javelin throw", 
+    "295": "waxing back", 
+    "296": "riding or walking with horse", 
+    "297": "plastering", 
+    "298": "long jump", 
+    "299": "parkour", 
+    "300": "wrapping present", 
+    "301": "egg hunting", 
+    "302": "archery", 
+    "303": "cleaning toilet", 
+    "304": "swimming backstroke", 
+    "305": "snowboarding", 
+    "306": "catching or throwing baseball", 
+    "307": "massaging back", 
+    "308": "blowing glass", 
+    "309": "playing guitar", 
+    "310": "playing chess", 
+    "311": "golf driving", 
+    "312": "presenting weather forecast", 
+    "313": "rock scissors paper", 
+    "314": "high jump", 
+    "315": "baking cookies", 
+    "316": "using computer", 
+    "317": "washing feet", 
+    "318": "arranging flowers", 
+    "319": "playing bass guitar", 
+    "320": "spraying", 
+    "321": "cutting pineapple", 
+    "322": "waxing chest", 
+    "323": "auctioning", 
+    "324": "jetskiing", 
+    "325": "drinking", 
+    "326": "busking", 
+    "327": "playing monopoly", 
+    "328": "salsa dancing", 
+    "329": "waxing eyebrows", 
+    "330": "watering plants", 
+    "331": "zumba", 
+    "332": "chopping wood", 
+    "333": "pushing wheelchair", 
+    "334": "carving pumpkin", 
+    "335": "building shed", 
+    "336": "making jewelry", 
+    "337": "catching or throwing softball", 
+    "338": "bending metal", 
+    "339": "ice skating", 
+    "340": "dancing charleston", 
+    "341": "abseiling", 
+    "342": "climbing a rope", 
+    "343": "crying", 
+    "344": "cleaning shoes", 
+    "345": "dancing ballet", 
+    "346": "driving tractor", 
+    "347": "triple jump", 
+    "348": "throwing ball", 
+    "349": "getting a haircut", 
+    "350": "running on treadmill", 
+    "351": "climbing ladder", 
+    "352": "blasting sand", 
+    "353": "playing trombone", 
+    "354": "drop kicking", 
+    "355": "country line dancing", 
+    "356": "changing wheel", 
+    "357": "feeding goats", 
+    "358": "tying knot (not on a tie)", 
+    "359": "setting table", 
+    "360": "shaving legs", 
+    "361": "kissing", 
+    "362": "riding mule", 
+    "363": "counting money", 
+    "364": "laying bricks", 
+    "365": "barbequing", 
+    "366": "news anchoring", 
+    "367": "smoking hookah", 
+    "368": "cooking egg", 
+    "369": "peeling apples", 
+    "370": "yoga", 
+    "371": "sharpening pencil", 
+    "372": "dribbling basketball", 
+    "373": "petting cat", 
+    "374": "playing ice hockey", 
+    "375": "milking cow", 
+    "376": "shining shoes", 
+    "377": "juggling soccer ball", 
+    "378": "scuba diving", 
+    "379": "playing squash or racquetball", 
+    "380": "drinking beer", 
+    "381": "sign language interpreting", 
+    "382": "playing basketball", 
+    "383": "breakdancing", 
+    "384": "testifying", 
+    "385": "making snowman", 
+    "386": "golf putting", 
+    "387": "playing didgeridoo", 
+    "388": "biking through snow", 
+    "389": "sailing", 
+    "390": "jumpstyle dancing", 
+    "391": "water sliding", 
+    "392": "grooming horse", 
+    "393": "massaging feet", 
+    "394": "playing paintball", 
+    "395": "making a cake", 
+    "396": "bowling", 
+    "397": "contact juggling", 
+    "398": "applying cream", 
+    "399": "playing badminton"
+}

requirements.txt

@@ -0,0 +1,6 @@
+torch==1.9.1
+torchvision 
+einops 
+timm
+Pillow
+decord

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))

uniformer_light_video.py

@@ -0,0 +1,595 @@
+# All rights reserved.
+from math import ceil, sqrt
+from collections import OrderedDict
+import torch
+import torch.nn as nn
+from functools import partial
+from timm.models.vision_transformer import _cfg
+from timm.models.layers import trunc_normal_, DropPath, to_2tuple
+import os
+
+
+global_attn = None
+token_indices = None
+
+model_path = 'path_to_models'
+model_path = {
+    'uniformer_xxs_128_in1k': os.path.join(model_path, 'uniformer_xxs_128_in1k.pth'),
+    'uniformer_xxs_160_in1k': os.path.join(model_path, 'uniformer_xxs_160_in1k.pth'),
+    'uniformer_xxs_192_in1k': os.path.join(model_path, 'uniformer_xxs_192_in1k.pth'),
+    'uniformer_xxs_224_in1k': os.path.join(model_path, 'uniformer_xxs_224_in1k.pth'),
+    'uniformer_xs_192_in1k': os.path.join(model_path, 'uniformer_xs_192_in1k.pth'),
+    'uniformer_xs_224_in1k': os.path.join(model_path, 'uniformer_xs_224_in1k.pth'),
+}
+
+
+def conv_3xnxn(inp, oup, kernel_size=3, stride=3, groups=1):
+    return nn.Conv3d(inp, oup, (3, kernel_size, kernel_size), (2, stride, stride), (1, 0, 0), groups=groups)
+    
+def conv_1xnxn(inp, oup, kernel_size=3, stride=3, groups=1):
+    return nn.Conv3d(inp, oup, (1, kernel_size, kernel_size), (1, stride, stride), (0, 0, 0), groups=groups)
+
+def conv_3xnxn_std(inp, oup, kernel_size=3, stride=3, groups=1):
+    return nn.Conv3d(inp, oup, (3, kernel_size, kernel_size), (1, stride, stride), (1, 0, 0), groups=groups)
+
+def conv_1x1x1(inp, oup, groups=1):
+    return nn.Conv3d(inp, oup, (1, 1, 1), (1, 1, 1), (0, 0, 0), groups=groups)
+
+def conv_3x3x3(inp, oup, groups=1):
+    return nn.Conv3d(inp, oup, (3, 3, 3), (1, 1, 1), (1, 1, 1), groups=groups)
+
+def conv_5x5x5(inp, oup, groups=1):
+    return nn.Conv3d(inp, oup, (5, 5, 5), (1, 1, 1), (2, 2, 2), groups=groups)
+
+def bn_3d(dim):
+    return nn.BatchNorm3d(dim)
+
+
+# code is from https://github.com/YifanXu74/Evo-ViT
+def easy_gather(x, indices):
+    # x => B x N x C
+    # indices => B x N
+    B, N, C = x.shape
+    N_new = indices.shape[1]
+    offset = torch.arange(B, dtype=torch.long, device=x.device).view(B, 1) * N
+    indices = indices + offset
+    # only select the informative tokens
+    out = x.reshape(B * N, C)[indices.view(-1)].reshape(B, N_new, C)
+    return out
+
+
+# code is from https://github.com/YifanXu74/Evo-ViT
+def merge_tokens(x_drop, score):
+    # x_drop => B x N_drop
+    # score => B x N_drop
+    weight = score / torch.sum(score, dim=1, keepdim=True)
+    x_drop = weight.unsqueeze(-1) * x_drop
+    return torch.sum(x_drop, dim=1, keepdim=True)
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., trade_off=1):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        # updating weight for global score
+        self.trade_off = trade_off
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]   # make torchscript happy (cannot use tensor as tuple)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+
+        # update global score
+        global global_attn
+        tradeoff = self.trade_off
+        if isinstance(global_attn, int):
+            global_attn = torch.mean(attn[:, :, 0, 1:], dim=1)
+        elif global_attn.shape[1] == N - 1:
+            # no additional token and no pruning, update all global scores
+            cls_attn = torch.mean(attn[:, :, 0, 1:], dim=1)
+            global_attn = (1 - tradeoff) * global_attn + tradeoff * cls_attn
+        else:
+            # only update the informative tokens
+            # the first one is class token
+            # the last one is rrepresentative token
+            cls_attn = torch.mean(attn[:, :, 0, 1:-1], dim=1)
+            if self.training:
+                temp_attn = (1 - tradeoff) * global_attn[:, :(N - 2)] + tradeoff * cls_attn
+                global_attn = torch.cat((temp_attn, global_attn[:, (N - 2):]), dim=1)
+            else:
+                # no use torch.cat() for fast inference
+                global_attn[:, :(N - 2)] = (1 - tradeoff) * global_attn[:, :(N - 2)] + tradeoff * cls_attn
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class CMlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = conv_1x1x1(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = conv_1x1x1(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+    
+    
+class CBlock(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
+        self.norm1 = bn_3d(dim)
+        self.conv1 = conv_1x1x1(dim, dim, 1)
+        self.conv2 = conv_1x1x1(dim, dim, 1)
+        self.attn = conv_5x5x5(dim, dim, groups=dim)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = bn_3d(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = CMlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.pos_embed(x)
+        x = x + self.drop_path(self.conv2(self.attn(self.conv1(self.norm1(x)))))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x   
+
+
+class EvoSABlock(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, prune_ratio=1,
+                 trade_off=0, downsample=False):
+        super().__init__()
+        self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop, trade_off=trade_off)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+        self.prune_ratio = prune_ratio
+        self.downsample = downsample
+        if downsample:
+            self.avgpool = nn.AvgPool3d(kernel_size=(1, 2, 2), stride=(1, 2, 2))
+
+    def forward(self, cls_token, x):
+        x = x + self.pos_embed(x)
+        B, C, T, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)  
+
+        if self.prune_ratio == 1:
+            x = torch.cat([cls_token, x], dim=1)
+            x = x + self.drop_path(self.attn(self.norm1(x)))
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+            cls_token, x = x[:, :1], x[:, 1:]
+            x = x.transpose(1, 2).reshape(B, C, T, H, W)
+            return cls_token, x  
+        else:
+            global global_attn, token_indices
+            # calculate the number of informative tokens
+            N = x.shape[1]
+            N_ = int(N * self.prune_ratio)
+            # sort global attention
+            indices = torch.argsort(global_attn, dim=1, descending=True)
+
+            # concatenate x, global attention and token indices => x_ga_ti
+            # rearrange the tensor according to new indices 
+            x_ga_ti = torch.cat((x, global_attn.unsqueeze(-1), token_indices.unsqueeze(-1)), dim=-1)
+            x_ga_ti = easy_gather(x_ga_ti, indices)
+            x_sorted, global_attn, token_indices = x_ga_ti[:, :, :-2], x_ga_ti[:, :, -2], x_ga_ti[:, :, -1]
+
+            # informative tokens
+            x_info = x_sorted[:, :N_]
+            # merge dropped tokens
+            x_drop = x_sorted[:, N_:]
+            score = global_attn[:, N_:]
+            #  B x N_drop x C => B x 1 x C
+            rep_token = merge_tokens(x_drop, score)
+            # concatenate new tokens
+            x = torch.cat((cls_token, x_info, rep_token), dim=1)
+
+            # slow update
+            fast_update = 0
+            tmp_x = self.attn(self.norm1(x))
+            fast_update = fast_update + tmp_x[:, -1:]
+            x = x + self.drop_path(tmp_x)
+            tmp_x = self.mlp(self.norm2(x))
+            fast_update = fast_update + tmp_x[:, -1:]
+            x = x + self.drop_path(tmp_x)
+            # fast update
+            x_drop = x_drop + fast_update.expand(-1, N - N_, -1)
+
+            cls_token, x = x[:, :1, :], x[:, 1:-1, :]
+            if self.training:
+                x_sorted = torch.cat((x, x_drop), dim=1)
+            else:
+                x_sorted[:, N_:] = x_drop
+                x_sorted[:, :N_] = x
+
+            # recover token
+            # scale for normalization
+            old_global_scale = torch.sum(global_attn, dim=1, keepdim=True)
+            # recover order
+            indices = torch.argsort(token_indices, dim=1)
+            x_ga_ti = torch.cat((x_sorted, global_attn.unsqueeze(-1), token_indices.unsqueeze(-1)), dim=-1)
+            x_ga_ti = easy_gather(x_ga_ti, indices)
+            x_patch, global_attn, token_indices = x_ga_ti[:, :, :-2], x_ga_ti[:, :, -2], x_ga_ti[:, :, -1]
+            x_patch = x_patch.transpose(1, 2).reshape(B, C, T, H, W)
+
+            if self.downsample:
+                # downsample global attention
+                global_attn = global_attn.reshape(B, 1, T, H, W)
+                global_attn = self.avgpool(global_attn).view(B, -1)
+                # normalize global attention
+                new_global_scale = torch.sum(global_attn, dim=1, keepdim=True)
+                scale = old_global_scale / new_global_scale
+                global_attn = global_attn * scale
+            
+            return cls_token, x_patch
+
+
+class SABlock(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.pos_embed(x)
+        B, C, T, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = x + self.drop_path(self.attn(self.norm1(x)))
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        x = x.transpose(1, 2).reshape(B, C, T, H, W)
+        return x    
+
+
+class SplitSABlock(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.pos_embed = conv_3x3x3(dim, dim, groups=dim)
+        self.t_norm = norm_layer(dim)
+        self.t_attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop)
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
+            attn_drop=attn_drop, proj_drop=drop)
+        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+    def forward(self, x):
+        x = x + self.pos_embed(x)
+        B, C, T, H, W = x.shape
+        attn = x.view(B, C, T, H * W).permute(0, 3, 2, 1).contiguous()
+        attn = attn.view(B * H * W, T, C)
+        attn = attn + self.drop_path(self.t_attn(self.t_norm(attn)))
+        attn = attn.view(B, H * W, T, C).permute(0, 2, 1, 3).contiguous()
+        attn = attn.view(B * T, H * W, C)
+        residual = x.view(B, C, T, H * W).permute(0, 2, 3, 1).contiguous()
+        residual = residual.view(B * T, H * W, C)
+        attn = residual + self.drop_path(self.attn(self.norm1(attn)))
+        attn = attn.view(B, T * H * W, C)
+        out = attn + self.drop_path(self.mlp(self.norm2(attn)))
+        out = out.transpose(1, 2).reshape(B, C, T, H, W)
+        return out
+
+
+class SpeicalPatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+        
+        self.proj = nn.Sequential(
+            nn.Conv3d(in_chans, embed_dim // 2, kernel_size=(3, 3, 3), stride=(1, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(embed_dim // 2),
+            nn.GELU(),
+            nn.Conv3d(embed_dim // 2, embed_dim, kernel_size=(3, 3, 3), stride=(2, 2, 2), padding=(1, 1, 1)),
+            nn.BatchNorm3d(embed_dim),
+        )
+
+    def forward(self, x):
+        B, C, T, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        # assert H == self.img_size[0] and W == self.img_size[1], \
+        #     f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x)
+        B, C, T, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = x.reshape(B, T, H, W, -1).permute(0, 4, 1, 2, 3).contiguous()
+        return x
+    
+
+class PatchEmbed(nn.Module):
+    """ Image to Patch Embedding
+    """
+    def __init__(self, patch_size=16, in_chans=3, embed_dim=768):
+        super().__init__()
+        patch_size = to_2tuple(patch_size)
+        self.patch_size = patch_size
+        self.norm = nn.LayerNorm(embed_dim)
+        self.proj = conv_1xnxn(in_chans, embed_dim, kernel_size=patch_size[0], stride=patch_size[0])
+
+    def forward(self, x):
+        B, C, T, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        # assert H == self.img_size[0] and W == self.img_size[1], \
+        #     f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x)
+        B, C, T, H, W = x.shape
+        x = x.flatten(2).transpose(1, 2)
+        x = self.norm(x)
+        x = x.reshape(B, T, H, W, -1).permute(0, 4, 1, 2, 3).contiguous()
+        return x
+
+
+class Uniformer_light(nn.Module):
+    """ Vision Transformer
+    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`  -
+        https://arxiv.org/abs/2010.11929
+    """
+    def __init__(self, depth=[3, 4, 8, 3], in_chans=3, num_classes=400, embed_dim=[64, 128, 320, 512],
+                 head_dim=64, mlp_ratio=[4., 4., 4., 4.], qkv_bias=True, qk_scale=None, representation_size=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=None, 
+                 prune_ratio=[[], [], [1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], [0.5, 0.5, 0.5]],
+                 trade_off=[[], [], [1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], [0.5, 0.5, 0.5]]
+                 ):
+        super().__init__()
+
+        self.num_classes = num_classes
+        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models
+        norm_layer = partial(nn.LayerNorm, eps=1e-6) 
+        
+        self.patch_embed1 = SpeicalPatchEmbed(
+            patch_size=4, in_chans=in_chans, embed_dim=embed_dim[0])
+        self.patch_embed2 = PatchEmbed(
+            patch_size=2, in_chans=embed_dim[0], embed_dim=embed_dim[1])
+        self.patch_embed3 = PatchEmbed(
+            patch_size=2, in_chans=embed_dim[1], embed_dim=embed_dim[2])
+        self.patch_embed4 = PatchEmbed(
+            patch_size=2, in_chans=embed_dim[2], embed_dim=embed_dim[3])
+
+        # class token
+        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim[2]))
+        self.cls_upsample = nn.Linear(embed_dim[2], embed_dim[3])
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depth))]  # stochastic depth decay rule
+        num_heads = [dim // head_dim for dim in embed_dim]
+        self.blocks1 = nn.ModuleList([
+            CBlock(
+                dim=embed_dim[0], num_heads=num_heads[0], mlp_ratio=mlp_ratio[0], qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer)
+            for i in range(depth[0])])
+        self.blocks2 = nn.ModuleList([
+            CBlock(
+                dim=embed_dim[1], num_heads=num_heads[1], mlp_ratio=mlp_ratio[1], qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]], norm_layer=norm_layer)
+            for i in range(depth[1])])
+        self.blocks3 = nn.ModuleList([
+            EvoSABlock(
+                dim=embed_dim[2], num_heads=num_heads[2], mlp_ratio=mlp_ratio[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]+depth[1]], norm_layer=norm_layer,
+                prune_ratio=prune_ratio[2][i], trade_off=trade_off[2][i],
+                downsample=True if i == depth[2] - 1 else False)
+            for i in range(depth[2])])
+        self.blocks4 = nn.ModuleList([
+            EvoSABlock(
+                dim=embed_dim[3], num_heads=num_heads[3], mlp_ratio=mlp_ratio[3], qkv_bias=qkv_bias, qk_scale=qk_scale,
+                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i+depth[0]+depth[1]+depth[2]], norm_layer=norm_layer,
+                prune_ratio=prune_ratio[3][i], trade_off=trade_off[3][i])
+        for i in range(depth[3])])
+        self.norm = bn_3d(embed_dim[-1])
+        self.norm_cls = nn.LayerNorm(embed_dim[-1])
+        
+        # Representation layer
+        if representation_size:
+            self.num_features = representation_size
+            self.pre_logits = nn.Sequential(OrderedDict([
+                ('fc', nn.Linear(embed_dim, representation_size)),
+                ('act', nn.Tanh())
+            ]))
+        else:
+            self.pre_logits = nn.Identity()
+        
+        # Classifier head
+        self.head = nn.Linear(embed_dim[-1], num_classes) if num_classes > 0 else nn.Identity()
+        self.head_cls = nn.Linear(embed_dim[-1], num_classes) if num_classes > 0 else nn.Identity()
+        
+        self.apply(self._init_weights)
+
+        for name, p in self.named_parameters():
+            # fill proj weight with 1 here to improve training dynamics. Otherwise temporal attention inputs
+            # are multiplied by 0*0, which is hard for the model to move out of.
+            if 't_attn.qkv.weight' in name:
+                nn.init.constant_(p, 0)
+            if 't_attn.qkv.bias' in name:
+                nn.init.constant_(p, 0)
+            if 't_attn.proj.weight' in name:
+                nn.init.constant_(p, 1)
+            if 't_attn.proj.bias' in name:
+                nn.init.constant_(p, 0)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'pos_embed', 'cls_token'}
+
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=''):
+        self.num_classes = num_classes
+        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+
+    def inflate_weight(self, weight_2d, time_dim, center=False):
+        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 forward_features(self, x):
+        x = self.patch_embed1(x)
+        x = self.pos_drop(x)
+        for blk in self.blocks1:
+            x = blk(x)
+        x = self.patch_embed2(x)
+        for blk in self.blocks2:
+            x = blk(x)
+        x = self.patch_embed3(x)
+        # add cls_token in stage3
+        cls_token = self.cls_token.expand(x.shape[0], -1, -1)
+        global global_attn, token_indices
+        global_attn = 0
+        token_indices = torch.arange(x.shape[2] * x.shape[3] * x.shape[4], dtype=torch.long, device=x.device).unsqueeze(0)
+        token_indices = token_indices.expand(x.shape[0], -1)
+        for blk in self.blocks3:
+            cls_token, x = blk(cls_token, x)
+        # upsample cls_token before stage4
+        cls_token = self.cls_upsample(cls_token)
+        x = self.patch_embed4(x)
+        # whether reset global attention? Now simple avgpool
+        token_indices = torch.arange(x.shape[2] * x.shape[3] * x.shape[4], dtype=torch.long, device=x.device).unsqueeze(0)
+        token_indices = token_indices.expand(x.shape[0], -1)
+        for blk in self.blocks4:
+            cls_token, x = blk(cls_token, x)
+        if self.training:
+            # layer normalization for cls_token
+            cls_token = self.norm_cls(cls_token)
+        x = self.norm(x)
+        x = self.pre_logits(x)
+        return cls_token, x
+
+    def forward(self, x):
+        cls_token, x = self.forward_features(x)
+        x = x.flatten(2).mean(-1)
+        if self.training:
+            x = self.head(x), self.head_cls(cls_token.squeeze(1))
+        else:
+            x = self.head(x)
+        return x
+
+
+def uniformer_xxs_video(**kwargs):
+    model = Uniformer_light(
+        depth=[2, 5, 8, 2],
+        prune_ratio=[[], [], [1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], [0.5, 0.5]],
+        trade_off=[[], [], [1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], [0.5, 0.5]],
+        embed_dim=[56, 112, 224, 448], head_dim=28, mlp_ratio=[3, 3, 3, 3], qkv_bias=True,
+        **kwargs)
+    model.default_cfg = _cfg()
+    return model
+
+
+def uniformer_xs_video(**kwargs):
+    model = Uniformer_light(
+        depth=[3, 5, 9, 3],
+        prune_ratio=[[], [], [1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], [0.5, 0.5, 0.5]],
+        trade_off=[[], [], [1, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5], [0.5, 0.5, 0.5]],
+        embed_dim=[64, 128, 256, 512], head_dim=32, mlp_ratio=[3, 3, 3, 3], qkv_bias=True,
+        **kwargs)
+    model.default_cfg = _cfg()
+    return model
+
+
+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 = uniformer_xxs_video()
+    # print(model)
+
+    flops = FlopCountAnalysis(model, torch.rand(1, 3, num_frames, 160, 160))
+    s = time.time()
+    print(flop_count_table(flops, max_depth=1))
+    print(time.time()-s)