1.赛题解析
赛题地址:https://tianchi.aliyun.com/competition/entrance/532155
任务:
- 输入:元宇宙仿真平台生成的前视摄像头虚拟视频数据(8-10秒左右);
- 输出:对视频中的信息进行综合理解,以指定的json文件格式,按照数据说明中的关键词(key)填充描述型的文本信息(value,中文/英文均可以)
评分标准:
系统会针对参赛者提交的json文件,通过描述型的文本信息与真值进行对比,综合得出分数;其中,“距离最近的交通参与者的行为”的题目为2分,其它题目为1分;每个视频的满分为10分。每一个视频结果中的key值,需要参考数据说明的json格式示例,请勿进行修改。
2.Baseline详解
深度学习框架搭建:
可参考
Paddle版:
首先导入库
import paddle
from PIL import Image
from clip import tokenize, load_model
import glob, json, os
import cv2
from PIL import Image
from tqdm import tqdm_notebook
import numpy as np
from sklearn.preprocessing import normalize
import matplotlib.pyplot as plt
paddle:PaddlePaddle深度学习框架。
PIL:Python Imaging Library,用于图像处理。
clip:包含CLIP模型的库。
glob:用于获取文件路径。
json:用于处理JSON数据。
os:用于操作文件和目录。
cv2:OpenCV库,用于读取视频帧。
tqdm_notebook:用于显示进度条。
numpy:用于数值计算。
sklearn.preprocessing.normalize:用于归一化数据。
matplotlib.pyplot:用于绘图。
加载CLIP模型和转换器:
model, transforms = load_model('ViT_B_32', pretrained=True)
定义匹配词典:
en_match_words = {"scerario" : ["suburbs","city street","expressway","tunnel","parking-lot","gas or charging stations","unknown"],"weather" : ["clear","cloudy","raining","foggy","snowy","unknown"],"period" : ["daytime","dawn or dusk","night","unknown"],"road_structure" : ["normal","crossroads","T-junction","ramp","lane merging","parking lot entrance","round about","unknown"],"general_obstacle" : ["nothing","speed bumper","traffic cone","water horse","stone","manhole cover","nothing","unknown"],"abnormal_condition" : ["uneven","oil or water stain","standing water","cracked","nothing","unknown"],"ego_car_behavior" : ["slow down","go straight","turn right","turn left","stop","U-turn","speed up","lane change","others"],"closest_participants_type" : ["passenger car","bus","truck","pedestrain","policeman","nothing","others","unknown"],"closest_participants_behavior" : ["slow down","go straight","turn right","turn left","stop","U-turn","speed up","lane change","others"],
}
定义结果JSON对象
submit_json = {"author" : "abc" ,"time" : "231011","model" : "model_name","test_results" : []
}
获取视频文件路径并排序:
paths = glob.glob('./初赛测试视频/*')
paths.sort()
遍历每个视频文件:
for video_path in paths:print(video_path)
读取视频帧并进行预处理
cap = cv2.VideoCapture(video_path)
img = cap.read()[1]
image = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
image = Image.fromarray(image)
image = transforms(image).unsqueeze(0)
定义单个视频结果的字典,并设置默认值:
single_video_result = {"clip_id": clip_id,"scerario" : "cityroad","weather":"unknown","period":"night","road_structure":"ramp","general_obstacle":"nothing","abnormal_condition":"nothing","ego_car_behavior":"turning right","closest_participants_type":"passenger car","closest_participants_behavior":"braking"
}
对于每个关键词,使用CLIP模型进行分类:
for keyword in en_match_words.keys():if keyword not in ["weather", "road_structure"]:#只有当关键词为"weather"或"road_structure"时才会执行后续的操作,其他关键词则会跳过continuetexts = np.array(en_match_words[keyword])#先将关键词对应的文本转换为一个NumPy数组
#先使用了一个名为tokenize的函数,它将关键词转换为模型能够理解的标记序列,然后调用CLIP模型的model方法,传入图像和标记化的文本,获取图像和文本的logits(预测得分)。其中,logits_per_image表示图像的logits,logits_per_text表示文本的logitswith paddle.no_grad():logits_per_image, logits_per_text = model(image, tokenize(en_match_words[keyword]))probs = paddle.nn.functional.softmax(logits_per_image, axis=-1)#使用softmax函数对图像的logits进行归一化处理,得到每个类别的概率probs = probs.numpy() single_video_result[keyword] = texts[probs[0].argsort()[::-1][0]]#将概率值转换为NumPy数组,并根据概率值从高到低进行排序。然后将对应的文本赋值给single_video_result字典中的相应关键词
将单个视频结果添加到结果JSON对象中:
submit_json["test_results"].append(single_video_result)
将结果JSON对象保存为文件:
with open('clip_result.json', 'w', encoding='utf-8') as up:json.dump(submit_json, up, ensure_ascii=False)
对一系列视频进行分类,并将结果保存在一个JSON文件中。其中使用了PaddlePaddle的深度学习框架和OpenAI的CLIP模型来进行图像和文本的匹配和分类。
CLIP模型原理
可参考另一位助教写的博客点击直达
参考链接1
CLIP的训练数据是文本-图像对:一张图像和它对应的文本描述,这里希望通过对比学习,模型能够学习到文本-图像对的匹配关系。如下图所示,CLIP包括两个模型:Text Encoder和Image Encoder,其中Text Encoder用来提取文本的特征,可以采用NLP中常用的text transformer模型;而Image Encoder用来提取图像的特征,可以采用常用CNN模型或者vision transformer。
CLIP的思想非常简单,只需要看懂这幅图就可以了,左边是训练的原理,CLIP一共有两个模态,一个是文本模态,一个是视觉模态,分别对应了Text Encoder和Image Encoder。
- Text Encoder用于对文本进行编码,获得其Embedding;
- Image Encoder用于对图片编码,获得其Embedding。
- 两个Embedding均为一定长度的单一向量。
pytorch版:
全部代码及注释
#导入库
import glob, json, os
import cv2
from PIL import Image
from tqdm import tqdm_notebook
import numpy as np
from sklearn.preprocessing import normalize
import matplotlib.pyplot as pltfrom PIL import Image
import requests
from transformers import CLIPProcessor, CLIPModel
#加载预训练的CLIP模型:
#通过使用Hugging Face的transformers库,该代码从预训练模型 "openai/clip-vit-large-patch14-336" 中加载了一个CLIP模型,并创建了一个处理器processor
model = CLIPModel.from_pretrained("openai/clip-vit-large-patch14-336")
processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14-336")
#通过URL下载一张图像,并使用PIL库打开该图像,处理图像数据
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
#准备模型输入并获取输出,使用CLIP处理器processor对图像和文本进行处理,准备成模型需要的张量格式,然后将输入传递给CLIP模型,获取模型的输出
inputs = processor(text=["a photo of a cat", "a photo of a dog"], images=image, return_tensors="pt", padding=True)outputs = model(**inputs)logits_per_image = outputs.logits_per_image # this is the image-text similarity score# 得到图像-文本相似度得分
logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities# 对结果进行softmax处理,得到标签概率
#我们得到了图像和文本之间的相似度得分logits_per_image,并对其进行了softmax处理,得到了标签的概率分布#分别定义了中文和英文对应的关键词匹配字典。
cn_match_words = {"工况描述": ["高速/城市快速路", "城区", "郊区", "隧道", "停车场", "加油站/充电站", "未知"],"天气": ["晴天", "雨天", "多云", "雾天", "下雪", "未知"],"时间": ["白天", "夜晚", "拂晓/日暮", "未知"],"道路结构": ["十字路口", "丁字路口", "上下匝道", "车道汇入", "进出停车场", "环岛", "正常车道", "未知"],"一般障碍物": ["雉桶", "水马", "碎石/石块", "井盖", "减速带", "没有"],"道路异常情况": ["油污/水渍", "积水", "龟裂", "起伏不平", "没有", "未知"],"自车行为": ["直行", "左转", "右转", "停止", "掉头", "加速", "减速", "变道", "其它"],"最近的交通参与者": ["行人", "小型汽车", "卡车", "交警", "没有", "未知", "其它"],"最近的交通参与者行为": ["直行", "左转", "右转", "停止", "掉头", "加速", "减速", "变道", "其它"],
}en_match_words = {
"scerario" : ["suburbs","city street","expressway","tunnel","parking-lot","gas or charging stations","unknown"],
"weather" : ["clear","cloudy","raining","foggy","snowy","unknown"],
"period" : ["daytime","dawn or dusk","night","unknown"],
"road_structure" : ["normal","crossroads","T-junction","ramp","lane merging","parking lot entrance","round about","unknown"],
"general_obstacle" : ["nothing","speed bumper","traffic cone","water horse","stone","manhole cover","nothing","unknown"],
"abnormal_condition" : ["uneven","oil or water stain","standing water","cracked","nothing","unknown"],
"ego_car_behavior" : ["slow down","go straight","turn right","turn left","stop","U-turn","speed up","lane change","others"],
"closest_participants_type" : ["passenger car","bus","truck","pedestrain","policeman","nothing","others","unknown"],
"closest_participants_behavior" : ["slow down","go straight","turn right","turn left","stop","U-turn","speed up","lane change","others"],
}
#读取视频文件
cap = cv2.VideoCapture(r'D:/D/Download/360安全浏览器下载/初赛测试视频/初赛测试视频/41.avi')
#读取视频帧图像读取第一帧图片可以考虑一下怎么改进
img = cap.read()[1]
image = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)#转换图像颜色格式,将图像从BGR格式转换为RGB格式
#创建PIL图像对象
image = Image.fromarray(image)#使用PIL库的fromarray函数将NumPy数组(即经过颜色格式转换后的图像数据)转换为PIL的Image对象image.resize((600, 300))
#两个Python字典的定义和赋值操作
submit_json = {"作者" : "阿水" ,"时间" : "231011","模型名字" : "model_name","测试结果" : []
}submit_json = {"author" : "abc" ,"time" : "231011","model" : "model_name","test_results" : []
}
#使用了Python中的glob模块来匹配文件路径,获取指定目录下所有文件的路径并将结果进行排序
paths = glob.glob(r'.\chusai\*')
paths.sort()
#对paths列表中的每个视频路径进行处理,并生成相应的结果
for video_path in paths:print(video_path)# clip_id = video_path.split('/')[-1]clip_id = os.path.split(video_path)[-1]print(clip_id)# clip_id = video_path.split('/')[-1][:-4]cap = cv2.VideoCapture(video_path)#打开视频文件img = cap.read()[1]#读取视频帧,将得到的图像赋值给变量imgimage = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)# image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)image = Image.fromarray(image) #将BGR格式的图像转换为RGB格式,并将结果赋值给变量image #定义了一个字典single_video_result,包含了一些视频相关的属性single_video_result = {"clip_id": clip_id,"scerario" : "cityroad","weather":"clear","period":"daytime","road_structure":"namal","general_obstacle":"nothing","abnormal_condition":"nothing","ego_car_behavior":"go straight","closest_participants_type":"passenger car","closest_participants_behavior":"braking"}#历en_match_words字典中的关键字,对除了weather、road_structure、scerario和period以外的关键字进行处理for keyword in en_match_words.keys():if keyword not in ["weather", "road_structure", 'scerario', 'road_structure', 'period']:continue texts = np.array(en_match_words[keyword])#转化为Numpy数组inputs = processor(text=list(texts), images=image, return_tensors="pt", padding=True)#使用模型的处理器(processor)对文本和图像进行处理,生成模型输入所需的格式。这里的处理器负责将文本和图像转换为模型能够接受的输入格式print(inputs)outputs = model(**inputs)#调用模型,将处理后的输入传入模型中,得到模型的输出。模型的输出包括了图像和文本之间的相似度得分logits_per_image = outputs.logits_per_image # this is the image-text similarity scoreprobs = logits_per_image.softmax(dim=1) # probs: [[1.2686e-03, 5.4499e-02, 6.7968e-04, 9.4355e-01]]将相似度得分进行softmax处理,得到每个类别的概率分布single_video_result[keyword] = texts[probs[0].argsort().numpy()[::-1][0]]#根据概率最大的类别,从文本数组中取出相应的文本,并将其作为该关键字对应的值,赋给single_video_result字典submit_json["test_results"].append(single_video_result)#将处理得到的single_video_result字典添加到submit_json["test_results"]列表中,最终得到一个JSON对象,其中包含了对每个视频的预测结果len(paths)# 遍历每个数据条目,对clip_id进行修改#首先,这个循环遍历了名为submit_json的JSON对象中的'test_results'字段对应的列表。对于列表中的每个条目(在代码中称为entry),它检查是否存在名为'clip_id'的字段。#如果存在'clip_id'字段,代码会执行entry['clip_id'].split("\\")[-1]操作。这行代码的作用是将'clip_id'字段的值按照反斜杠\分割,并取分割后的结果的最后一个部分。这样的操作通常用于获取文件路径中的文件名部分。
for entry in submit_json['test_results']:if 'clip_id' in entry:entry['clip_id'] = entry['clip_id'].split("\\")[-1]
#代码打开了一个名为coggle_result5.json的文件(路径为D:/D/Download/360安全浏览器下载/)以供写入,并使用json.dump将经过处理的submit_json对象写入到这个文件中。参数ensure_ascii=False表示在生成的JSON文件中允许非ASCII字符的存在,通常用于处理非英文文本。
with open(r'D:/D/Download/360安全浏览器下载/coggle_result5.json', 'w', encoding='utf-8') as up:json.dump(submit_json, up, ensure_ascii=False)
# "作者" : "abc" ,
# "时间" : "YYMMDD",
# "模型名字" : "model_name",
# "测试结果" :[
# {
# "视频ID" : "xxxx_1",
# "工况描述" : "城市道路",
# "天气":"未知",
# "时间":"夜晚",
# "道路结构":"匝道",
# "一般障碍物":"无",
# "道路异常情况":"无",
# "自车行为":"右转",
# "最近的交通参与者":"小轿车",
# "最近的交通参与者行为":"制动"
# },submit_json
目前分数132
在pytorch版上做的改进
1.换模型
clip-vit-large-patch14 是两年前的模型
clip-vit-large-patch14-336是一年前的模型
秉着新的肯定会比老的效果好的想法分数从121->126不错!
2.然后还尝试改了一下single_video_result
然后就132了然后就没有了
第一次直播:
第二次直播: