eval_knn.py

"""
KNN monitor for test tracking the current feature quality.

Copied from
https://colab.research.google.com/github/facebookresearch/moco/blob/colab-notebook/colab/moco_cifar10_demo.ipynb
Note:
    This moco demo add F.normalize instead of directly using the features as the below two code bases.

The code is further from https://github.com/leftthomas/SimCLR and https://github.com/zhirongw/lemniscate.pytorch.
"""

import pickle
import torch
import torch.nn.functional as F
import torch.distributed as dist
import tqdm

import argparse
import datetime
import numpy as np
import time
import torch
import torch.backends.cudnn as cudnn
import json
import os
from functools import partial
from pathlib import Path
from collections import OrderedDict

from mixup import Mixup
from timm.models import create_model
from timm.loss import LabelSmoothingCrossEntropy, SoftTargetCrossEntropy
from timm.utils import ModelEma
from optim_factory import create_optimizer, get_parameter_groups, LayerDecayValueAssigner

from datasets import build_dataset
from engine_for_finetuning import train_one_epoch, validation_one_epoch, final_test, merge, merge_mean_per_class
from utils import NativeScalerWithGradNormCount as NativeScaler
from utils import  multiple_samples_collate
import utils
import modeling_finetune




def get_world_size():
    if not dist.is_available():
        return 1
    if not dist.is_initialized():
        return 1
    return dist.get_world_size()


def all_gather(data):
    """
    Run all_gather on arbitrary picklable data (not necessarily tensors)
    Args:
        data: any picklable object
    Returns:
        list[data]: list of data gathered from each rank
    """
    world_size = get_world_size()
    if world_size == 1:
        return [data]

    if type(data) is torch.Tensor:
        data = data.cpu()
    # serialized to a Tensor
    buffer = pickle.dumps(data)
    storage = torch.ByteStorage.from_buffer(buffer)
    tensor = torch.ByteTensor(storage).to("cuda")

    # obtain Tensor size of each rank
    local_size = torch.LongTensor([tensor.numel()]).to("cuda")
    size_list = [torch.LongTensor([0]).to("cuda") for _ in range(world_size)]
    dist.all_gather(size_list, local_size)
    size_list = [int(size.item()) for size in size_list]
    max_size = max(size_list)

    # receiving Tensor from all ranks
    # we pad the tensor because torch all_gather does not support
    # gathering tensors of different shapes
    tensor_list = []
    for _ in size_list:
        tensor_list.append(torch.ByteTensor(size=(max_size,)).to("cuda"))
    if local_size != max_size:
        padding = torch.ByteTensor(size=(max_size - local_size,)).to("cuda")
        tensor = torch.cat((tensor, padding), dim=0)
    dist.all_gather(tensor_list, tensor)

    data_list = []
    for size, tensor in zip(size_list, tensor_list):
        buffer = tensor.cpu().numpy().tobytes()[:size]
        data = pickle.loads(buffer)
        if type(data) is torch.Tensor:
            data = data.to("cuda")
        data_list.append(data)

    return data_list


class KNNMonitor:
    def __init__(self, device: torch.device, visible: bool = False, tqdm: bool = False, dist=False, num_classes=400, train_loader=None, test_loader=None):
        """

        :param dataset: the dataset name of this monitor.
        """
        self.memory_data_loader = train_loader
        self.test_data_loader = test_loader

        self.num_classes = num_classes
        # This temperature param was tuned for Imagnet, from https://github.com/zhirongw/lemniscate.pytorch
        self.knn_t = 0.07

        self.visible = visible
        self.tqdm = tqdm
        self.dist = dist
        self.device = device
        self.knn_k = 200

    def test(self, net, return_top5=False, **model_args):
        classes = self.num_classes
        total_top1, total_top5, total_num, feature_bank, target_bank = 0.0, 0.0, 0, [], []
        with torch.no_grad():
            # generate feature bank
            for data_iter_step, (data, target, _, _, _) in enumerate(self.test_data_loader):
                # print(data.shape)
                target = torch.Tensor([int(i) for i in target])
                # print(target)
                data = data.to(self.device, non_blocking=True)
                feature = net(data)
                feature = F.normalize(feature, dim=1)
                feature_bank.append(feature)
                target_bank.append(target)
            # [D, N]
            feature_bank = torch.cat(feature_bank, dim=0).t().contiguous()
            target_bank = torch.cat(target_bank, dim=0).contiguous()

            if self.dist:
                feature_banks = all_gather(feature_bank)
                target_banks = all_gather(target_bank)
                first_device = feature_banks[0].device
                for bank in feature_banks:
                    if first_device != bank.device:
                        raise Warning(f"Device are not equal {first_device} and {bank.device}.")
                # [[feature_dim, batch_size]] --> [feature_dim, sum(batch_size)]
                feature_bank = torch.cat(feature_banks, -1)
                target_bank = torch.cat(target_banks, 0)
                if self.visible:
                    print(f"Gather feature bank for device {self.device}: {feature_bank.shape}")

            # [N]
            # feature_labels = torch.tensor(self.memory_data_loader.dataset.targets, device=feature_bank.device)
            feature_labels = target_bank

            # if self.dist:
            #     feature_labels = feature_labels.reshape(-1, dist.get_world_size()).t().flatten().contiguous()

            # loop test data to predict the label by weighted knn search
            test_bar = tqdm.tqdm(self.test_data_loader, disable=not self.visible or not self.tqdm)
            for data_iter_step, (data, target, _, _, _) in enumerate(self.test_data_loader):
                # print(f"Data iter step = {data_iter_step}")
                # print(data.shape)
                # print(target)
                data, target = data.to(self.device, non_blocking=True), target.to(self.device, non_blocking=True)
                feature = net(data)
                feature = F.normalize(feature, dim=1)

                pred_labels = self.knn_predict(feature, feature_bank, feature_labels, classes)

                total_num += data.size(0)
                total_top1 += (pred_labels[:, 0] == target).float().sum().item()
                total_top5 += torch.sum((pred_labels[:, :5] == target.unsqueeze(dim=-1)).any(dim=-1).float()).item()

        if self.dist:
            total_top1 = sum(all_gather(total_top1))
            total_top5 = sum(all_gather(total_top5))
            total_num = sum(all_gather(total_num))

        if return_top5:
            return total_top1 / total_num * 100, total_top5 / total_num * 100
        else:
            return total_top1 / total_num * 100

    # knn monitor as in InstDisc https://arxiv.org/abs/1805.01978
    # implementation follows http://github.com/zhirongw/lemniscate.pytorch and https://github.com/leftthomas/SimCLR
    def knn_predict(self, feature, feature_bank, feature_labels, classes):
        # compute cos similarity between each feature vector and feature bank ---> [B, N]
        #   Note: [batch_size, feature_dim] x [feature_dim, memory_bank_size] --> [batch_size, memory_bank_size]
        sim_matrix = torch.mm(feature, feature_bank)
        # [B, K]
        sim_weight, sim_indices = sim_matrix.topk(k=self.knn_k, dim=-1)
        # [B, K]
        sim_labels = torch.gather(feature_labels.expand(feature.size(0), -1), dim=-1, index=sim_indices)
        sim_weight = (sim_weight / self.knn_t).exp()

        # counts for each class
        one_hot_label = torch.zeros(feature.size(0) * self.knn_k, classes, device=sim_labels.device)
        # [B*K, C]
        # print(sim_labels)
        sim_labels = sim_labels.long()
        one_hot_label = one_hot_label.scatter(dim=-1, index=sim_labels.view(-1, 1), value=1.0)
        # weighted score ---> [B, C]
        pred_scores = torch.sum(one_hot_label.view(feature.size(0), -1, classes) * sim_weight.unsqueeze(dim=-1), dim=1)

        pred_labels = pred_scores.argsort(dim=-1, descending=True)
        return pred_labels


def get_args():
    parser = argparse.ArgumentParser('VideoMAE fine-tuning and evaluation script for video classification', add_help=False)
    parser.add_argument('--batch_size', default=64, type=int)
    parser.add_argument('--epochs', default=30, type=int)
    parser.add_argument('--update_freq', default=1, type=int)
    parser.add_argument('--save_ckpt_freq', default=100, type=int)

    # Model parameters
    parser.add_argument('--model', default='vit_base_patch16_224', type=str, metavar='MODEL',
                        help='Name of model to train')
    parser.add_argument('--tubelet_size', type=int, default= 2)
    parser.add_argument('--input_size', default=224, type=int,
                        help='videos input size')

    parser.add_argument('--fc_drop_rate', type=float, default=0.0, metavar='PCT',
                        help='Dropout rate (default: 0.)')
    parser.add_argument('--drop', type=float, default=0.0, metavar='PCT',
                        help='Dropout rate (default: 0.)')
    parser.add_argument('--attn_drop_rate', type=float, default=0.0, metavar='PCT',
                        help='Attention dropout rate (default: 0.)')
    parser.add_argument('--drop_path', type=float, default=0.1, metavar='PCT',
                        help='Drop path rate (default: 0.1)')

    parser.add_argument('--disable_eval_during_finetuning', action='store_true', default=False)
    parser.add_argument('--model_ema', action='store_true', default=False)
    parser.add_argument('--model_ema_decay', type=float, default=0.9999, help='')
    parser.add_argument('--model_ema_force_cpu', action='store_true', default=False, help='')

    # Optimizer parameters
    parser.add_argument('--opt', default='adamw', type=str, metavar='OPTIMIZER',
                        help='Optimizer (default: "adamw"')
    parser.add_argument('--opt_eps', default=1e-8, type=float, metavar='EPSILON',
                        help='Optimizer Epsilon (default: 1e-8)')
    parser.add_argument('--opt_betas', default=None, type=float, nargs='+', metavar='BETA',
                        help='Optimizer Betas (default: None, use opt default)')
    parser.add_argument('--clip_grad', type=float, default=None, metavar='NORM',
                        help='Clip gradient norm (default: None, no clipping)')
    parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
                        help='SGD momentum (default: 0.9)')
    parser.add_argument('--weight_decay', type=float, default=0.05,
                        help='weight decay (default: 0.05)')
    parser.add_argument('--weight_decay_end', type=float, default=None, help="""Final value of the
        weight decay. We use a cosine schedule for WD and using a larger decay by
        the end of training improves performance for ViTs.""")

    parser.add_argument('--lr', type=float, default=1e-3, metavar='LR',
                        help='learning rate (default: 1e-3)')
    parser.add_argument('--layer_decay', type=float, default=0.75)

    parser.add_argument('--warmup_lr', type=float, default=1e-6, metavar='LR',
                        help='warmup learning rate (default: 1e-6)')
    parser.add_argument('--min_lr', type=float, default=1e-6, metavar='LR',
                        help='lower lr bound for cyclic schedulers that hit 0 (1e-5)')

    parser.add_argument('--warmup_epochs', type=int, default=5, metavar='N',
                        help='epochs to warmup LR, if scheduler supports')
    parser.add_argument('--warmup_steps', type=int, default=-1, metavar='N',
                        help='num of steps to warmup LR, will overload warmup_epochs if set > 0')

    # Augmentation parameters
    parser.add_argument('--color_jitter', type=float, default=0.4, metavar='PCT',
                        help='Color jitter factor (default: 0.4)')
    parser.add_argument('--num_sample', type=int, default=2,
                        help='Repeated_aug (default: 2)')
    parser.add_argument('--aa', type=str, default='rand-m7-n4-mstd0.5-inc1', metavar='NAME',
                        help='Use AutoAugment policy. "v0" or "original". " + "(default: rand-m7-n4-mstd0.5-inc1)'),
    parser.add_argument('--smoothing', type=float, default=0.1,
                        help='Label smoothing (default: 0.1)')
    parser.add_argument('--train_interpolation', type=str, default='bicubic',
                        help='Training interpolation (random, bilinear, bicubic default: "bicubic")')

    # Evaluation parameters
    parser.add_argument('--crop_pct', type=float, default=None)
    parser.add_argument('--short_side_size', type=int, default=224)
    parser.add_argument('--test_num_segment', type=int, default=5)
    parser.add_argument('--test_num_crop', type=int, default=3)
    
    # Random Erase params
    parser.add_argument('--reprob', type=float, default=0.25, metavar='PCT',
                        help='Random erase prob (default: 0.25)')
    parser.add_argument('--remode', type=str, default='pixel',
                        help='Random erase mode (default: "pixel")')
    parser.add_argument('--recount', type=int, default=1,
                        help='Random erase count (default: 1)')
    parser.add_argument('--resplit', action='store_true', default=False,
                        help='Do not random erase first (clean) augmentation split')

    # Mixup params
    parser.add_argument('--mixup', type=float, default=0.8,
                        help='mixup alpha, mixup enabled if > 0.')
    parser.add_argument('--cutmix', type=float, default=1.0,
                        help='cutmix alpha, cutmix enabled if > 0.')
    parser.add_argument('--cutmix_minmax', type=float, nargs='+', default=None,
                        help='cutmix min/max ratio, overrides alpha and enables cutmix if set (default: None)')
    parser.add_argument('--mixup_prob', type=float, default=1.0,
                        help='Probability of performing mixup or cutmix when either/both is enabled')
    parser.add_argument('--mixup_switch_prob', type=float, default=0.5,
                        help='Probability of switching to cutmix when both mixup and cutmix enabled')
    parser.add_argument('--mixup_mode', type=str, default='batch',
                        help='How to apply mixup/cutmix params. Per "batch", "pair", or "elem"')

    # Finetuning params
    parser.add_argument('--finetune', default='', help='finetune from checkpoint')
    parser.add_argument('--model_key', default='model|module', type=str)
    parser.add_argument('--model_prefix', default='', type=str)
    parser.add_argument('--init_scale', default=0.001, type=float)
    parser.add_argument('--use_checkpoint', action='store_true')
    parser.set_defaults(use_checkpoint=False)
    parser.add_argument('--use_mean_pooling', action='store_true')
    parser.set_defaults(use_mean_pooling=True)
    parser.add_argument('--use_cls', action='store_false', dest='use_mean_pooling')

    # Dataset parameters
    parser.add_argument('--data_path', default='/path/to/list_kinetics-400', type=str,
                        help='dataset path')
    parser.add_argument('--eval_data_path', default=None, type=str,
                        help='dataset path for evaluation')
    parser.add_argument('--nb_classes', default=400, type=int,
                        help='number of the classification types')
    parser.add_argument('--imagenet_default_mean_and_std', default=True, action='store_true')
    parser.add_argument('--num_segments', type=int, default= 1)
    parser.add_argument('--num_frames', type=int, default= 16)
    parser.add_argument('--sampling_rate', type=int, default= 2)
    parser.add_argument('--data_set', default='SSV2', choices=['Kinetics-400', 'SSV2', 'UCF101', 'HMDB51','image_folder','DIV48','GYM99','FXS1','UBS1'],
                        type=str, help='dataset')
    parser.add_argument('--output_dir', default='',
                        help='path where to save, empty for no saving')
    parser.add_argument('--log_dir', default=None,
                        help='path where to tensorboard log')
    parser.add_argument('--device', default='cuda',
                        help='device to use for training / testing')
    parser.add_argument('--seed', default=0, type=int)
    parser.add_argument('--resume', default='',
                        help='resume from checkpoint')
    parser.add_argument('--auto_resume', action='store_true')
    parser.add_argument('--no_auto_resume', action='store_false', dest='auto_resume')
    parser.set_defaults(auto_resume=True)

    parser.add_argument('--save_ckpt', action='store_true')
    parser.add_argument('--no_save_ckpt', action='store_false', dest='save_ckpt')
    parser.set_defaults(save_ckpt=True)

    parser.add_argument('--start_epoch', default=0, type=int, metavar='N',
                        help='start epoch')
    parser.add_argument('--eval', action='store_true',
                        help='Perform evaluation only')
    parser.add_argument('--dist_eval', action='store_true', default=True,
                        help='Enabling distributed evaluation')
    parser.add_argument('--num_workers', default=10, type=int)
    parser.add_argument('--pin_mem', action='store_true',
                        help='Pin CPU memory in DataLoader for more efficient (sometimes) transfer to GPU.')
    parser.add_argument('--no_pin_mem', action='store_false', dest='pin_mem')
    parser.set_defaults(pin_mem=True)

    # distributed training parameters
    parser.add_argument('--world_size', default=1, type=int,
                        help='number of distributed processes')
    parser.add_argument('--local-rank', default=-1, type=int)
    parser.add_argument('--dist_on_itp', action='store_true')
    parser.add_argument('--dist_url', default='env://',
                        help='url used to set up distributed training')

    parser.add_argument('--enable_deepspeed', action='store_true', default=False)
    parser.add_argument('--examples', default=0, type=int)

    known_args, _ = parser.parse_known_args()

    if known_args.enable_deepspeed:
        try:
            import deepspeed
            from deepspeed import DeepSpeedConfig
            parser = deepspeed.add_config_arguments(parser)
            ds_init = deepspeed.initialize
        except:
            print("Please 'pip install deepspeed'")
            exit(0)
    else:
        ds_init = None

    return parser.parse_args(), ds_init


def main(args, ds_init):
    utils.init_distributed_mode(args)

    if ds_init is not None:
        utils.create_ds_config(args)

    print(args)

    device = torch.device(args.device)

    # fix the seed for reproducibility
    seed = args.seed + utils.get_rank()
    torch.manual_seed(seed)
    np.random.seed(seed)
    # random.seed(seed)

    cudnn.benchmark = True

    dataset_train, args.nb_classes = build_dataset(is_train=True, test_mode=False, args=args)
    if args.disable_eval_during_finetuning:
        dataset_val = None
    else:
        dataset_val, _ = build_dataset(is_train=False, test_mode=False, args=args)
    dataset_test, _ = build_dataset(is_train=False, test_mode=True, args=args)
    # print(dataset_test)
    

    num_tasks = utils.get_world_size()
    global_rank = utils.get_rank()
    sampler_train = torch.utils.data.DistributedSampler(
        dataset_train, num_replicas=num_tasks, rank=global_rank, shuffle=True
    )
    print("Sampler_train = %s" % str(sampler_train))
    if args.dist_eval:
        if len(dataset_val) % num_tasks != 0:
            print('Warning: Enabling distributed evaluation with an eval dataset not divisible by process number. '
                    'This will slightly alter validation results as extra duplicate entries are added to achieve '
                    'equal num of samples per-process.')
        sampler_val = torch.utils.data.DistributedSampler(
            dataset_val, num_replicas=num_tasks, rank=global_rank, shuffle=False)
        sampler_test = torch.utils.data.DistributedSampler(
            dataset_test, num_replicas=num_tasks, rank=global_rank, shuffle=False)
    else:
        sampler_val = torch.utils.data.SequentialSampler(dataset_val)

    if global_rank == 0 and args.log_dir is not None:
        os.makedirs(args.log_dir, exist_ok=True)
        log_writer = utils.TensorboardLogger(log_dir=args.log_dir)
    else:
        log_writer = None

    if args.num_sample > 1:
        collate_func = partial(multiple_samples_collate, fold=False)
    else:
        collate_func = None

    data_loader_train = torch.utils.data.DataLoader(
        dataset_train, sampler=sampler_train,
        batch_size=args.batch_size,
        num_workers=args.num_workers,
        pin_memory=args.pin_mem,
        drop_last=True,
        collate_fn=collate_func,
    )

    if dataset_val is not None:
        data_loader_val = torch.utils.data.DataLoader(
            dataset_val, sampler=sampler_val,
            batch_size=int(1.5 * args.batch_size),
            num_workers=args.num_workers,
            pin_memory=args.pin_mem,
            drop_last=False
        )
    else:
        data_loader_val = None

    if dataset_test is not None:
        data_loader_test = torch.utils.data.DataLoader(
            dataset_test, sampler=sampler_test,
            batch_size=args.batch_size,
            num_workers=args.num_workers,
            pin_memory=args.pin_mem,
            drop_last=False
        )
    else:
        data_loader_test = None

    mixup_fn = None
    mixup_active = args.mixup > 0 or args.cutmix > 0. or args.cutmix_minmax is not None
    if mixup_active:
        print("Mixup is activated!")
        mixup_fn = Mixup(
            mixup_alpha=args.mixup, cutmix_alpha=args.cutmix, cutmix_minmax=args.cutmix_minmax,
            prob=args.mixup_prob, switch_prob=args.mixup_switch_prob, mode=args.mixup_mode,
            label_smoothing=args.smoothing, num_classes=args.nb_classes)

    model = create_model(
        args.model,
        pretrained=False,
        num_classes=args.nb_classes,
        all_frames=args.num_frames * args.num_segments,
        tubelet_size=args.tubelet_size,
        fc_drop_rate=args.fc_drop_rate,
        drop_rate=args.drop,
        drop_path_rate=args.drop_path,
        attn_drop_rate=args.attn_drop_rate,
        drop_block_rate=None,
        use_checkpoint=args.use_checkpoint,
        use_mean_pooling=args.use_mean_pooling,
        init_scale=args.init_scale,
    )

    patch_size = model.patch_embed.patch_size
    print("Patch size = %s" % str(patch_size))
    args.window_size = (args.num_frames // 2, args.input_size // patch_size[0], args.input_size // patch_size[1])
    args.patch_size = patch_size

    if args.finetune:
        if args.finetune.startswith('https'):
            checkpoint = torch.hub.load_state_dict_from_url(
                args.finetune, map_location='cpu', check_hash=True)
        else:
            checkpoint = torch.load(args.finetune, map_location='cpu')

        print("Load ckpt from %s" % args.finetune)
        checkpoint_model = None
        for model_key in args.model_key.split('|'):
            if model_key in checkpoint:
                checkpoint_model = checkpoint[model_key]
                print("Load state_dict by model_key = %s" % model_key)
                break
        if checkpoint_model is None:
            checkpoint_model = checkpoint
        state_dict = model.state_dict()
        for k in ['head.weight', 'head.bias']:
            if k in checkpoint_model and checkpoint_model[k].shape != state_dict[k].shape:
                print(f"Removing key {k} from pretrained checkpoint")
                del checkpoint_model[k]

        all_keys = list(checkpoint_model.keys())
        new_dict = OrderedDict()
        for key in all_keys:
            if key.startswith('backbone.'):
                new_dict[key[9:]] = checkpoint_model[key]
            elif key.startswith('encoder.'):
                new_dict[key[8:]] = checkpoint_model[key]
            else:
                new_dict[key] = checkpoint_model[key]
        checkpoint_model = new_dict

        # interpolate position embedding
        if 'pos_embed' in checkpoint_model:
            pos_embed_checkpoint = checkpoint_model['pos_embed']
            embedding_size = pos_embed_checkpoint.shape[-1] # channel dim
            num_patches = model.patch_embed.num_patches # 
            num_extra_tokens = model.pos_embed.shape[-2] - num_patches # 0/1

            # height (== width) for the checkpoint position embedding 
            orig_size = int(((pos_embed_checkpoint.shape[-2] - num_extra_tokens)//(args.num_frames // model.patch_embed.tubelet_size)) ** 0.5)
            # height (== width) for the new position embedding
            new_size = int((num_patches // (args.num_frames // model.patch_embed.tubelet_size) )** 0.5)
            # class_token and dist_token are kept unchanged
            if orig_size != new_size:
                print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
                extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
                # only the position tokens are interpolated
                pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
                # B, L, C -> BT, H, W, C -> BT, C, H, W
                pos_tokens = pos_tokens.reshape(-1, args.num_frames // model.patch_embed.tubelet_size, orig_size, orig_size, embedding_size)
                pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
                pos_tokens = torch.nn.functional.interpolate(
                    pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
                # BT, C, H, W -> BT, H, W, C ->  B, T, H, W, C
                pos_tokens = pos_tokens.permute(0, 2, 3, 1).reshape(-1, args.num_frames // model.patch_embed.tubelet_size, new_size, new_size, embedding_size) 
                pos_tokens = pos_tokens.flatten(1, 3) # B, L, C
                new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
                checkpoint_model['pos_embed'] = new_pos_embed

        utils.load_state_dict(model, checkpoint_model, prefix=args.model_prefix)

    model.to(device)
    dist = True if args.world_size > 1 else False
    knn_eval = KNNMonitor(device = device, visible=True, tqdm=True, dist=dist, num_classes=args.nb_classes, train_loader=data_loader_train, test_loader=data_loader_test)
    top1, top5 = knn_eval.test(model, return_top5=True, tubelet_size=args.tubelet_size, num_frames=args.num_frames, num_segments=args.num_segments)
    print(f"Top-1 accuracy: {top1:.2f}")
    print(f"Top-5 accuracy: {top5:.2f}")


if __name__ == '__main__':
    args, ds_init = get_args()
    main(args, ds_init)