kaggle视频行为分析1st and Future - Player Contact Detection

这次比赛的目标是检测美式橄榄球NFL比赛中球员经历的外部接触。您将使用视频和球员追踪数据来识别发生接触的时刻，以帮助提高球员的安全。两种接触，一种是人与人的，另一种是人与地面，不包括脚底和地面的，跟我之前做的这个是同一个主办方举行的

kaggle视频追踪NFL Health & Safety - Helmet Assignment-CSDN博客

之前做的是视频追踪，用的deepsort，这一场比赛用的2.5DCNN。

EDA部分

eda可以参考这一个notebook，用的fasteda，挺方便的

NFL Player Contact Detection EDA 🏈 | Kaggle

视频数据在test和train文件夹里面，还提供了这一个train_baseline_helmets.csv， 是由上一次比赛的冠军方案产生的，是我之前做的视频追踪，train_player_tracking.csv 的频率是10HZ，视频是59.94HZ，之后要进行转换，snap 事件也就是比赛开始发生在视频的第五秒

train_labels.csv

• step: A number representing each each timestep for each play, starting at 0 at the moment of the play starting, and incrementing by 1 every 0.1 seconds.
• 之前说的比赛第5秒开始，一个step是0.1秒

接触发生以10HZ记录

[train/test]_player_tracking.csv

• datetime: timestamp at 10 Hz.

[train/test]_video_metadata.csv

be used to sync with player tracking data.和视频是同步的

训练部分

我自己租卡跑，20多个小时，10个epoch，我上传到kaggle，链接如下

track_weight | Kaggle

额外要用的一个数据集如下，我用的的4090显卡20核跑的，你要自己训练的话要自己修改一下

timm-0.6.9 | Kaggle

导入包

import os
import sys
import glob
import numpy as np
import pandas as pd
import random
import math
import gc
import cv2
from tqdm import tqdm
import time
from functools import lru_cache
import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import Dataset, DataLoader
from torch.cuda.amp import autocast, GradScaler
import timm
import albumentations as A
from albumentations.pytorch import ToTensorV2
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from timm.scheduler import CosineLRScheduler
sys.path.append('../input/timm-0-6-9/pytorch-image-models-master')

配置

CFG = {
    'seed': 42,
    'model': 'convnext_small.fb_in1k',
    'img_size': 256,
    'epochs': 10,
    'train_bs': 48, 
    'valid_bs': 32,
    'lr': 1e-3, 
    'weight_decay': 1e-6,
    'num_workers': 20,
    'max_grad_norm' : 1000,
    'epochs_warmup' : 3.0
}

我用的convnext，这个网络是原本的cnn根据vit模型去反复修改的，有兴趣自己去找论文看，但论文也就是在那反复调

设置种子和device

def seed_everything(seed):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

seed_everything(CFG['seed'])
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

添加一些额外的列和读取数据

def expand_contact_id(df):
    """
    Splits out contact_id into seperate columns.
    """
    df["game_play"] = df["contact_id"].str[:12]
    df["step"] = df["contact_id"].str.split("_").str[-3].astype("int")
    df["nfl_player_id_1"] = df["contact_id"].str.split("_").str[-2]
    df["nfl_player_id_2"] = df["contact_id"].str.split("_").str[-1]
    return df
labels = expand_contact_id(pd.read_csv("../input/nfl-player-contact-detection/train_labels.csv"))
train_tracking = pd.read_csv("../input/nfl-player-contact-detection/train_player_tracking.csv")
train_helmets = pd.read_csv("../input/nfl-player-contact-detection/train_baseline_helmets.csv")
train_video_metadata = pd.read_csv("../input/nfl-player-contact-detection/train_video_metadata.csv")

将视频数据转化为图像数据

import subprocess
from tqdm import tqdm
 
# 假设 train_helmets 是一个包含视频文件名的 DataFrame
for video in tqdm(train_helmets.video.unique()):
    if 'Endzone2' not in video:
        # 输入视频路径
        input_path = f'/openbayes/home/train/{video}'
        # 输出帧路径
        output_path = f'/openbayes/train/frames/{video}_%04d.jpg'
        
        # 构建 ffmpeg 命令
        command = [
            'ffmpeg',
            '-i', input_path,  # 输入视频文件
            '-q:v', '5',       # 设置输出图像质量
            '-f', 'image2',    # 输出为图像序列
            output_path,       # 输出图像路径
            '-hide_banner',    # 隐藏 ffmpeg 的 banner 信息
            '-loglevel', 'error'  # 只显示错误日志
        ]
        
        # 执行命令
        subprocess.run(command, check=True)

可以自己修改那里的质量，在kaggle上不能训练，要你自己租卡才跑的动

创建一些特征

def create_features(df, tr_tracking, merge_col="step", use_cols=["x_position", "y_position"]):
    output_cols = []
    df_combo = (
        df.astype({"nfl_player_id_1": "str"})
        .merge(
            tr_tracking.astype({"nfl_player_id": "str"})[
                ["game_play", merge_col, "nfl_player_id",] + use_cols
            ],
            left_on=["game_play", merge_col, "nfl_player_id_1"],
            right_on=["game_play", merge_col, "nfl_player_id"],
            how="left",
        )
        .rename(columns={c: c+"_1" for c in use_cols})
        .drop("nfl_player_id", axis=1)
        .merge(
            tr_tracking.astype({"nfl_player_id": "str"})[
                ["game_play", merge_col, "nfl_player_id"] + use_cols
            ],
            left_on=["game_play", merge_col, "nfl_player_id_2"],
            right_on=["game_play", merge_col, "nfl_player_id"],
            how="left",
        )
        .drop("nfl_player_id", axis=1)
        .rename(columns={c: c+"_2" for c in use_cols})
        .sort_values(["game_play", merge_col, "nfl_player_id_1", "nfl_player_id_2"])
        .reset_index(drop=True)
    )
    output_cols += [c+"_1" for c in use_cols]
    output_cols += [c+"_2" for c in use_cols]
    
    if ("x_position" in use_cols) & ("y_position" in use_cols):
        index = df_combo['x_position_2'].notnull()
        
        distance_arr = np.full(len(index), np.nan)
        tmp_distance_arr = np.sqrt(
            np.square(df_combo.loc[index, "x_position_1"] - df_combo.loc[index, "x_position_2"])
            + np.square(df_combo.loc[index, "y_position_1"]- df_combo.loc[index, "y_position_2"])
        )
        
        distance_arr[index] = tmp_distance_arr
        df_combo['distance'] = distance_arr
        output_cols += ["distance"]
        
    df_combo['G_flug'] = (df_combo['nfl_player_id_2']=="G")
    output_cols += ["G_flug"]
    return df_combo, output_cols


use_cols = [
    'x_position', 'y_position', 'speed', 'distance',
    'direction', 'orientation', 'acceleration', 'sa'
]

train, feature_cols = create_features(labels, train_tracking, use_cols=use_cols)

label和train_tracking进行合并，这里的feature_cols后面训练要用到

和视频的频率进行同步，过滤一部分数据

train_filtered = train.query('not distance>2').reset_index(drop=True)
train_filtered['frame'] = (train_filtered['step']/10*59.94+5*59.94).astype('int')+1
train_filtered.head()

视频频率是59.94,而数据集是10,这里将距离过大的pair去除

数据增强

train_aug = A.Compose([
    A.HorizontalFlip(p=0.5),
    A.ShiftScaleRotate(p=0.5),
    A.RandomBrightnessContrast(brightness_limit=(-0.1, 0.1), contrast_limit=(-0.1, 0.1), p=0.5),
    A.Normalize(mean=[0.], std=[1.]),
    ToTensorV2()
])

valid_aug = A.Compose([
    A.Normalize(mean=[0.], std=[1.]),
    ToTensorV2()
])

创建字典

video2helmets = {}
train_helmets_new = train_helmets.set_index('video')
for video in tqdm(train_helmets.video.unique()):
    video2helmets[video] = train_helmets_new.loc[video].reset_index(drop=True)
video2frames = {}

for game_play in tqdm(train_video_metadata.game_play.unique()):
    for view in ['Endzone', 'Sideline']:
        video = game_play + f'_{view}.mp4'
        video2frames[video] = max(list(map(lambda x:int(x.split('_')[-1].split('.')[0]), \
                                           glob.glob(f'../train/frames/{video}*'))))

取出视频对应的检测数据和每个视频的最大帧数，检测数据后面用来截取图像用的，最大帧数确保抽取的帧不超过这个范围

数据集

class MyDataset(Dataset):
    def __init__(self, df, aug=train_aug, mode='train'):
        self.df = df
        self.frame = df.frame.values
        self.feature = df[feature_cols].fillna(-1).values
        self.players = df[['nfl_player_id_1','nfl_player_id_2']].values
        self.game_play = df.game_play.values
        self.aug = aug
        self.mode = mode
        
    def __len__(self):
        return len(self.df)
    
    # @lru_cache(1024)
    # def read_img(self, path):
    #     return cv2.imread(path, 0)
   
    def __getitem__(self, idx):   
        window = 24
        frame = self.frame[idx]
        
        if self.mode == 'train':
            frame = frame + random.randint(-6, 6)

        players = []
        for p in self.players[idx]:
            if p == 'G':
                players.append(p)
            else:
                players.append(int(p))
        
        imgs = []
        for view in ['Endzone', 'Sideline']:
            video = self.game_play[idx] + f'_{view}.mp4'

            tmp = video2helmets[video]
#             tmp = tmp.query('@frame-@window<=frame<=@frame+@window')
            tmp[tmp['frame'].between(frame-window, frame+window)]
            tmp = tmp[tmp.nfl_player_id.isin(players)]#.sort_values(['nfl_player_id', 'frame'])
            tmp_frames = tmp.frame.values
            tmp = tmp.groupby('frame')[['left','width','top','height']].mean()
#0.002s

            bboxes = []
            for f in range(frame-window, frame+window+1, 1):
                if f in tmp_frames:
                    x, w, y, h = tmp.loc[f][['left','width','top','height']]
                    bboxes.append([x, w, y, h])
                else:
                    bboxes.append([np.nan, np.nan, np.nan, np.nan])
            bboxes = pd.DataFrame(bboxes).interpolate(limit_direction='both').values
            bboxes = bboxes[::4]

            if bboxes.sum() > 0:
                flag = 1
            else:
                flag = 0
#0.03s
                    
            for i, f in enumerate(range(frame-window, frame+window+1, 4)):
                img_new = np.zeros((256, 256), dtype=np.float32)

                if flag == 1 and f <= video2frames[video]:
                    img = cv2.imread(f'../train/frames/{video}_{f:04d}.jpg', 0)

                    x, w, y, h = bboxes[i]

                    img = img[int(y+h/2)-128:int(y+h/2)+128,int(x+w/2)-128:int(x+w/2)+128].copy()
                    img_new[:img.shape[0], :img.shape[1]] = img
                    
                imgs.append(img_new)
#0.06s
                
        feature = np.float32(self.feature[idx])

        img = np.array(imgs).transpose(1, 2, 0)    
        img = self.aug(image=img)["image"]
        label = np.float32(self.df.contact.values[idx])

        return img, feature, label

模型

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.backbone = timm.create_model(CFG['model'], pretrained=True, num_classes=500, in_chans=13)
        self.mlp = nn.Sequential(
            nn.Linear(18, 64),
            nn.LayerNorm(64),
            nn.ReLU(),
            nn.Dropout(0.2),
        )
        self.fc = nn.Linear(64+500*2, 1)

    def forward(self, img, feature):
        b, c, h, w = img.shape
        img = img.reshape(b*2, c//2, h, w)
        img = self.backbone(img).reshape(b, -1)
        feature = self.mlp(feature)
        y = self.fc(torch.cat([img, feature], dim=1))
        return y

这里len(feature_cols)是18,所以mlp输入是18,在上面

            for i, f in enumerate(range(frame-window, frame+window+1, 4)):
                img_new = np.zeros((256, 256), dtype=np.float32)
 
                if flag == 1 and f <= video2frames[video]:
                    img = cv2.imread(f'/openbayes/train/frames/{video}_{f:04d}.jpg', 0)
 
                    x, w, y, h = bboxes[i]
 
                    img = img[int(y+h/2)-128:int(y+h/2)+128,int(x+w/2)-128:int(x+w/2)+128].copy()
                    img_new[:img.shape[0], :img.shape[1]] = img
                    
                imgs.append(img_new)

进行了抽帧，每个视角抽了13帧，两个视角，总计26帧，所以输入通道26，跟之前的比赛一样，也是提供两个视角

for view in ['Endzone', 'Sideline']:

损失函数

model = Model()
model.to(device)
model.train()
import torch.nn as nn
criterion = nn.BCEWithLogitsLoss()

这里用的交叉熵

评估指标

def evaluate(model, loader_val, *, compute_score=True, pbar=None):
    """
    Predict and compute loss and score
    """
    tb = time.time()
    in_training = model.training
    model.eval()

    loss_sum = 0.0
    n_sum = 0
    y_all = []
    y_pred_all = []

    if pbar is not None:
        pbar = tqdm(desc='Predict', nrows=78, total=pbar)
        
    total= len(loader_val)

    for ibatch,(img, feature, label) in tqdm(enumerate(loader_val),total = total):
        # img, feature, label = [x.to(device) for x in batch]
        img = img.to(device)
        feature = feature.to(device)
        n = label.size(0)
        label = label.to(device)

        with torch.no_grad():
            y_pred = model(img, feature)
        loss = criterion(y_pred.view(-1), label)

        n_sum += n
        loss_sum += n * loss.item()
        
        if pbar is not None:
            pbar.update(len(img))
        
        del loss, img, label
        gc.collect()

    loss_val = loss_sum / n_sum


    ret = {'loss': loss_val,
           'time': time.time() - tb}
    
    model.train(in_training) 
    gc.collect()
    return ret

载入数据，设置学习率计划和优化器

train_set,valid_set = train_test_split(train_filtered,test_size=0.05, random_state=42,stratify = train_filtered['contact'])
train_set = MyDataset(train_set, train_aug, 'train')
train_loader = DataLoader(train_set, batch_size=CFG['train_bs'], shuffle=True, num_workers=12, pin_memory=True,drop_last=True)
valid_set = MyDataset(valid_set, valid_aug, 'test')
valid_loader = DataLoader(valid_set, batch_size=CFG['valid_bs'], shuffle=False, num_workers=12, pin_memory=True)
optimizer = torch.optim.AdamW(model.parameters(), lr=CFG['lr'], weight_decay=CFG['weight_decay'])
nbatch = len(train_loader)
warmup = CFG['epochs_warmup'] * nbatch
nsteps = CFG['epochs'] * nbatch 
scheduler = CosineLRScheduler(optimizer,warmup_t=warmup, warmup_lr_init=0.0, warmup_prefix=True,t_initial=(nsteps - warmup), lr_min=1e-6)    

开始训练，这里保存整个模型

for iepoch in range(CFG['epochs']):
    print('Epoch:', iepoch+1)
    loss_sum = 0.0
    n_sum = 0
    total = len(train_loader)
 
    # Train
    for ibatch,(img, feature, label) in tqdm(enumerate(train_loader),total = total):
        img = img.to(device)
        feature = feature.to(device)
        n = label.size(0)
        label = label.to(device)
        
 
        optimizer.zero_grad()
        y_pred = model(img, feature).squeeze(-1)
        loss = criterion(y_pred, label)
        loss_train = loss.item()
        loss_sum += n * loss_train
        n_sum += n
 
        loss.backward()
        grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(),CFG['max_grad_norm'])
 
        optimizer.step()
        scheduler.step(iepoch * nbatch + ibatch + 1)
        
    val = evaluate(model, valid_loader)
    time_val += val['time']
    loss_train = loss_sum / n_sum
    dt = (time.time() - tb) / 60
    print('Epoch: %d Train Loss: %.4f Test Loss: %.4f Time: %.2f min' %
          (iepoch + 1, loss_train, val['loss'],dt))
    if val['loss'] < best_loss:
        best_loss = val['loss']
        # Save model
        ofilename = '/openbayes/home/best_model.pt'
        torch.save(model, ofilename)
        print(ofilename, 'written')
    del val
    gc.collect()
 
dt = time.time() - tb
print(' %.2f min total, %.2f min val' % (dt / 60, time_val / 60))
gc.collect()

只保留权重可能会出现一些bug，保留整个模型比较稳妥

推理部分

这里我用TTA的版本

导入包

import os
import sys
sys.path.append('/kaggle/input/timm-0-6-9/pytorch-image-models-master')
import glob
import numpy as np
import pandas as pd
import random
import math
import gc
import cv2
from tqdm import tqdm
import time
from functools import lru_cache
import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.data import Dataset, DataLoader
from torch.cuda.amp import autocast, GradScaler
import timm
import albumentations as A
from albumentations.pytorch import ToTensorV2
import matplotlib.pyplot as plt
from sklearn.metrics import matthews_corrcoef

数据处理

这里基本和前面一样，我全部放一起了

CFG = {
    'seed': 42,
    'model': 'convnext_small.fb_in1k',
    'img_size': 256,
    'epochs': 10,
    'train_bs': 100, 
    'valid_bs': 64,
    'lr': 1e-3, 
    'weight_decay': 1e-6,
    'num_workers': 4
}
def seed_everything(seed):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

seed_everything(CFG['seed'])
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
def expand_contact_id(df):
    """
    Splits out contact_id into seperate columns.
    """
    df["game_play"] = df["contact_id"].str[:12]
    df["step"] = df["contact_id"].str.split("_").str[-3].astype("int")
    df["nfl_player_id_1"] = df["contact_id"].str.split("_").str[-2]
    df["nfl_player_id_2"] = df["contact_id"].str.split("_").str[-1]
    return df

labels = expand_contact_id(pd.read_csv("/kaggle/input/nfl-player-contact-detection/sample_submission.csv"))

test_tracking = pd.read_csv("/kaggle/input/nfl-player-contact-detection/test_player_tracking.csv")

test_helmets = pd.read_csv("/kaggle/input/nfl-player-contact-detection/test_baseline_helmets.csv")

test_video_metadata = pd.read_csv("/kaggle/input/nfl-player-contact-detection/test_video_metadata.csv")
!mkdir -p ../work/frames

for video in tqdm(test_helmets.video.unique()):
    if 'Endzone2' not in video:
        !ffmpeg -i /kaggle/input/nfl-player-contact-detection/test/{video} -q:v 2 -f image2 /kaggle/work/frames/{video}_%04d.jpg -hide_banner -loglevel error
def create_features(df, tr_tracking, merge_col="step", use_cols=["x_position", "y_position"]):
    output_cols = []
    df_combo = (
        df.astype({"nfl_player_id_1": "str"})
        .merge(
            tr_tracking.astype({"nfl_player_id": "str"})[
                ["game_play", merge_col, "nfl_player_id",] + use_cols
            ],
            left_on=["game_play", merge_col, "nfl_player_id_1"],
            right_on=["game_play", merge_col, "nfl_player_id"],
            how="left",
        )
        .rename(columns={c: c+"_1" for c in use_cols})
        .drop("nfl_player_id", axis=1)
        .merge(
            tr_tracking.astype({"nfl_player_id": "str"})[
                ["game_play", merge_col, "nfl_player_id"] + use_cols
            ],
            left_on=["game_play", merge_col, "nfl_player_id_2"],
            right_on=["game_play", merge_col, "nfl_player_id"],
            how="left",
        )
        .drop("nfl_player_id", axis=1)
        .rename(columns={c: c+"_2" for c in use_cols})
        .sort_values(["game_play", merge_col, "nfl_player_id_1", "nfl_player_id_2"])
        .reset_index(drop=True)
    )
    output_cols += [c+"_1" for c in use_cols]
    output_cols += [c+"_2" for c in use_cols]
    
    if ("x_position" in use_cols) & ("y_position" in use_cols):
        index = df_combo['x_position_2'].notnull()
        
        distance_arr = np.full(len(index), np.nan)
        tmp_distance_arr = np.sqrt(
            np.square(df_combo.loc[index, "x_position_1"] - df_combo.loc[index, "x_position_2"])
            + np.square(df_combo.loc[index, "y_position_1"]- df_combo.loc[index, "y_position_2"])
        )
        
        distance_arr[index] = tmp_distance_arr
        df_combo['distance'] = distance_arr
        output_cols += ["distance"]
        
    df_combo['G_flug'] = (df_combo['nfl_player_id_2']=="G")
    output_cols += ["G_flug"]
    return df_combo, output_cols


use_cols = [
    'x_position', 'y_position', 'speed', 'distance',
    'direction', 'orientation', 'acceleration', 'sa'
]

test, feature_cols = create_features(labels, test_tracking, use_cols=use_cols)
test
test_filtered = test.query('not distance>2').reset_index(drop=True)
test_filtered['frame'] = (test_filtered['step']/10*59.94+5*59.94).astype('int')+1
test_filtered
del test, labels, test_tracking
gc.collect()
train_aug = A.Compose([
    A.HorizontalFlip(p=0.5),
    A.ShiftScaleRotate(p=0.5),
    A.RandomBrightnessContrast(brightness_limit=(-0.1, 0.1), contrast_limit=(-0.1, 0.1), p=0.5),
    A.Normalize(mean=[0.], std=[1.]),
    ToTensorV2()
])

valid_aug = A.Compose([
    A.Normalize(mean=[0.], std=[1.]),
    ToTensorV2()
])
video2helmets = {}
test_helmets_new = test_helmets.set_index('video')
for video in tqdm(test_helmets.video.unique()):
    video2helmets[video] = test_helmets_new.loc[video].reset_index(drop=True)
    
del test_helmets, test_helmets_new
gc.collect()
video2frames = {}

for game_play in tqdm(test_video_metadata.game_play.unique()):
    for view in ['Endzone', 'Sideline']:
        video = game_play + f'_{view}.mp4'
        video2frames[video] = max(list(map(lambda x:int(x.split('_')[-1].split('.')[0]), \
                                           glob.glob(f'/kaggle/work/frames/{video}*'))))
class MyDataset(Dataset):
    def __init__(self, df, aug=valid_aug, mode='train'):
        self.df = df
        self.frame = df.frame.values
        self.feature = df[feature_cols].fillna(-1).values
        self.players = df[['nfl_player_id_1','nfl_player_id_2']].values
        self.game_play = df.game_play.values
        self.aug = aug
        self.mode = mode
        
    def __len__(self):
        return len(self.df)
    
    # @lru_cache(1024)
    # def read_img(self, path):
    #     return cv2.imread(path, 0)
   
    def __getitem__(self, idx):   
        window = 24
        frame = self.frame[idx]
        
        if self.mode == 'train':
            frame = frame + random.randint(-6, 6)

        players = []
        for p in self.players[idx]:
            if p == 'G':
                players.append(p)
            else:
                players.append(int(p))
        
        imgs = []
        for view in ['Endzone', 'Sideline']:
            video = self.game_play[idx] + f'_{view}.mp4'

            tmp = video2helmets[video]
#             tmp = tmp.query('@frame-@window<=frame<=@frame+@window')
            tmp[tmp['frame'].between(frame-window, frame+window)]
            tmp = tmp[tmp.nfl_player_id.isin(players)]#.sort_values(['nfl_player_id', 'frame'])
            tmp_frames = tmp.frame.values
            tmp = tmp.groupby('frame')[['left','width','top','height']].mean()
#0.002s

            bboxes = []
            for f in range(frame-window, frame+window+1, 1):
                if f in tmp_frames:
                    x, w, y, h = tmp.loc[f][['left','width','top','height']]
                    bboxes.append([x, w, y, h])
                else:
                    bboxes.append([np.nan, np.nan, np.nan, np.nan])
            bboxes = pd.DataFrame(bboxes).interpolate(limit_direction='both').values
            bboxes = bboxes[::4]

            if bboxes.sum() > 0:
                flag = 1
            else:
                flag = 0
#0.03s
                    
            for i, f in enumerate(range(frame-window, frame+window+1, 4)):
                img_new = np.zeros((256, 256), dtype=np.float32)

                if flag == 1 and f <= video2frames[video]:
                    img = cv2.imread(f'/kaggle/work/frames/{video}_{f:04d}.jpg', 0)

                    x, w, y, h = bboxes[i]

                    img = img[int(y+h/2)-128:int(y+h/2)+128,int(x+w/2)-128:int(x+w/2)+128].copy()
                    img_new[:img.shape[0], :img.shape[1]] = img
                    
                imgs.append(img_new)
#0.06s
                
        feature = np.float32(self.feature[idx])

        img = np.array(imgs).transpose(1, 2, 0)    
        img = self.aug(image=img)["image"]
        label = np.float32(self.df.contact.values[idx])

        return img, feature, label

查看截取出来的图片

img, feature, label = MyDataset(test_filtered, valid_aug, 'test')[0]
plt.imshow(img.permute(1,2,0)[:,:,7])
plt.show()
img.shape, feature, label

进行推理

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.backbone = timm.create_model(CFG['model'], pretrained=False, num_classes=500, in_chans=13)
        self.mlp = nn.Sequential(
            nn.Linear(18, 64),
            nn.LayerNorm(64),
            nn.ReLU(),
            nn.Dropout(0.2),
            # nn.Linear(64, 64),
            # nn.LayerNorm(64),
            # nn.ReLU(),
            # nn.Dropout(0.2)
        )
        self.fc = nn.Linear(64+500*2, 1)

    def forward(self, img, feature):
        b, c, h, w = img.shape
        img = img.reshape(b*2, c//2, h, w)
        img = self.backbone(img).reshape(b, -1)
        feature = self.mlp(feature)
        y = self.fc(torch.cat([img, feature], dim=1))
        return y
test_set = MyDataset(test_filtered, valid_aug, 'test')
test_loader = DataLoader(test_set, batch_size=CFG['valid_bs'], shuffle=False, num_workers=CFG['num_workers'], pin_memory=True)

model = Model().to(device)
model = torch.load('/kaggle/input/track-weight/best_model.pt')

model.eval()
    
y_pred = []
with torch.no_grad():
    tk = tqdm(test_loader, total=len(test_loader))
    for step, batch in enumerate(tk):
        if(step % 4 != 3):
            img, feature, label = [x.to(device) for x in batch]
            output1 = model(img, feature).squeeze(-1)
            output2 = model(img.flip(-1), feature).squeeze(-1)
            
            y_pred.extend(0.2*(output1.sigmoid().cpu().numpy()) + 0.8*(output2.sigmoid().cpu().numpy()))
        else:
            img, feature, label = [x.to(device) for x in batch]
            output = model(img.flip(-1), feature).squeeze(-1)
            y_pred.extend(output.sigmoid().cpu().numpy())    

y_pred = np.array(y_pred)

这里用了翻转，tta算是一种隐式模型集成

提交

th = 0.29

test_filtered['contact'] = (y_pred >= th).astype('int')

sub = pd.read_csv('/kaggle/input/nfl-player-contact-detection/sample_submission.csv')

sub = sub.drop("contact", axis=1).merge(test_filtered[['contact_id', 'contact']], how='left', on='contact_id')
sub['contact'] = sub['contact'].fillna(0).astype('int')

sub[["contact_id", "contact"]].to_csv("submission.csv", index=False)

sub.head()

推理代码链接和成绩

infer_code | Kaggle

修改版本

之前的，效果不是很好，我还是换成resnet50进行训练，结果如下，链接和权重如下

infer_code | Kaggle

best_weight | Kaggle