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
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
导入包
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
for view in ['Endzone', 'Sideline']:
损失函数
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
python
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
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
