{ "id": 578354, "title": "9th Solution: ResNet18 + GRU | 0.824", "url": "/competitions/nexar-collision-prediction/discussion/578354", "author_name": "Ángel Jacinto SÑnchez Ruiz", "post_date": "2025-05-10T08:56:05.608000", "votes": 7, "comment_count": 0, "views": 0, "content": "

πŸš— Collision Prediction with CNN-GRU | 5-Fold CV + TTA

\n

Public Notebook Link
\nScore: 0.824 | Training Time: 2h 8m 31s Β· GPU T4 x2

\n
\n

Thanks to @SLi for Script and dataset of resized videos

\n
\n

πŸ” Approach Overview

\n

We predict vehicle collisions using spatiotemporal video analysis with a hybrid CNN-GRU architecture.

\n\n
\n

🧠 Model Architecture

\n
 (nn.Module):\n     ():\n        ().__init__()\n        \n        .cnn = resnet18(weights=ResNet18_Weights.IMAGENET1K_V1)\n        .cnn.fc = nn.Identity()  \n        \n        .gru = nn.GRU(input_size=, hidden_size=, bidirectional=)\n        .classifier = nn.Linear(, )  \n\n     ():\n        \n        batch_size, T = x.shape[:]\n        x = x.view(-, *x.shape[:])  \n        x = .cnn(x)  \n        x = x.view(batch_size, T, -)  \n        x, _ = .gru(x)  \n        x = x[:, -]  \n        x = .classifier(x)\n         x\n
\n
\n

Training

\n

During training we cut random clips of T=64 frames each S=3 of them ending .5-1.5 seconds before time_of_event for positives. Leading to about 6.4 seconds long at about 10 fps clips. For negatives we just random cut at any possible time. Afterwards, a random 30 degrees rotation is applied (to avoid too unrealistic images). Next, we cut 224x224 random crops from initial 256x256 resized clips. Finally we apply Gaussian Noise, Brightness and Contrast augmentations. We experimented with Mixup too but without clear improvements.

\n

Plane Cross Entropy Loss have been used.

", "messages": [ { "id": 3198952, "postDate": "2025-05-10T08:56:05.610Z", "content": "

πŸš— Collision Prediction with CNN-GRU | 5-Fold CV + TTA

\n

Public Notebook Link
\nScore: 0.824 | Training Time: 2h 8m 31s Β· GPU T4 x2

\n
\n

Thanks to @SLi for Script and dataset of resized videos

\n
\n

πŸ” Approach Overview

\n

We predict vehicle collisions using spatiotemporal video analysis with a hybrid CNN-GRU architecture.

\n\n
\n

🧠 Model Architecture

\n
 (nn.Module):\n     ():\n        ().__init__()\n        \n        .cnn = resnet18(weights=ResNet18_Weights.IMAGENET1K_V1)\n        .cnn.fc = nn.Identity()  \n        \n        .gru = nn.GRU(input_size=, hidden_size=, bidirectional=)\n        .classifier = nn.Linear(, )  \n\n     ():\n        \n        batch_size, T = x.shape[:]\n        x = x.view(-, *x.shape[:])  \n        x = .cnn(x)  \n        x = x.view(batch_size, T, -)  \n        x, _ = .gru(x)  \n        x = x[:, -]  \n        x = .classifier(x)\n         x\n
\n
\n

Training

\n

During training we cut random clips of T=64 frames each S=3 of them ending .5-1.5 seconds before time_of_event for positives. Leading to about 6.4 seconds long at about 10 fps clips. For negatives we just random cut at any possible time. Afterwards, a random 30 degrees rotation is applied (to avoid too unrealistic images). Next, we cut 224x224 random crops from initial 256x256 resized clips. Finally we apply Gaussian Noise, Brightness and Contrast augmentations. We experimented with Mixup too but without clear improvements.

\n

Plane Cross Entropy Loss have been used.

", "rawMarkdown": "# πŸš— Collision Prediction with CNN-GRU | 5-Fold CV + TTA\n\n**[Public Notebook Link](https://www.kaggle.com/code/sacuscreed/fork-of-nexar-cnngru?scriptVersionId=225218134)** \n*Score: 0.824 | Training Time: 2h 8m 31s Β· GPU T4 x2*\n\n---\n\nThanks to @SLi for [**Script and dataset of resized videos**](https://www.kaggle.com/competitions/nexar-collision-prediction/discussion/564606)\n\n---\n\n## πŸ” Approach Overview \nWe predict vehicle collisions using **spatiotemporal video analysis** with a hybrid CNN-GRU architecture. \n- 🎯 **Stratified 5-fold CV** Although best score was from fold 2, full CV achieved **0.819**\n- πŸŒ€ **Test-Time Augmentation** Horizontal Flips\n- πŸ–ΌοΈ **ResNet18** for spatial features + **BiGRU** for temporal modeling \n\n---\n\n## 🧠 Model Architecture \n\n \n class myCNNGRU(nn.Module):\n def __init__(self):\n super().__init__()\n # 2D Backbone\n self.cnn = resnet18(weights=ResNet18_Weights.IMAGENET1K_V1)\n self.cnn.fc = nn.Identity() # Remove classification head\n # Temporal Model\n self.gru = nn.GRU(input_size=512, hidden_size=128, bidirectional=True)\n self.classifier = nn.Linear(256, 2) # Bidirectional β†’ 256\n\n def forward(self, x):\n # x: [batch, frames, C, H, W]\n batch_size, T = x.shape[:2]\n x = x.view(-1, *x.shape[2:]) # Merge batch + time\n x = self.cnn(x) # [batch*T, 512]\n x = x.view(batch_size, T, -1) # Unmerge\n x, _ = self.gru(x) # [batch, T, 256]\n x = x[:, -1] # Last timestep\n x = self.classifier(x)\n return x\n\n---\n\n## Training\n\nDuring training we cut random clips of **T=64** frames each **S=3** of them ending **.5-1.5 seconds** before **time_of_event** for positives. Leading to about **6.4 seconds** long at about **10 fps** clips. For negatives we just random cut at any possible time. Afterwards, a random **30 degrees** rotation is applied (to avoid too unrealistic images). Next, we cut 224x224 random crops from initial 256x256 resized clips. Finally we apply **Gaussian Noise**, **Brightness** and **Contrast** augmentations. We experimented with **Mixup** too but without clear improvements.\n\nPlane **Cross Entropy Loss** have been used.", "votes": 7 } ], "comments": [], "raw_markdown_by_id": { "3198952": "# πŸš— Collision Prediction with CNN-GRU | 5-Fold CV + TTA\n\n**[Public Notebook Link](https://www.kaggle.com/code/sacuscreed/fork-of-nexar-cnngru?scriptVersionId=225218134)** \n*Score: 0.824 | Training Time: 2h 8m 31s Β· GPU T4 x2*\n\n---\n\nThanks to @SLi for [**Script and dataset of resized videos**](https://www.kaggle.com/competitions/nexar-collision-prediction/discussion/564606)\n\n---\n\n## πŸ” Approach Overview \nWe predict vehicle collisions using **spatiotemporal video analysis** with a hybrid CNN-GRU architecture. \n- 🎯 **Stratified 5-fold CV** Although best score was from fold 2, full CV achieved **0.819**\n- πŸŒ€ **Test-Time Augmentation** Horizontal Flips\n- πŸ–ΌοΈ **ResNet18** for spatial features + **BiGRU** for temporal modeling \n\n---\n\n## 🧠 Model Architecture \n\n \n class myCNNGRU(nn.Module):\n def __init__(self):\n super().__init__()\n # 2D Backbone\n self.cnn = resnet18(weights=ResNet18_Weights.IMAGENET1K_V1)\n self.cnn.fc = nn.Identity() # Remove classification head\n # Temporal Model\n self.gru = nn.GRU(input_size=512, hidden_size=128, bidirectional=True)\n self.classifier = nn.Linear(256, 2) # Bidirectional β†’ 256\n\n def forward(self, x):\n # x: [batch, frames, C, H, W]\n batch_size, T = x.shape[:2]\n x = x.view(-1, *x.shape[2:]) # Merge batch + time\n x = self.cnn(x) # [batch*T, 512]\n x = x.view(batch_size, T, -1) # Unmerge\n x, _ = self.gru(x) # [batch, T, 256]\n x = x[:, -1] # Last timestep\n x = self.classifier(x)\n return x\n\n---\n\n## Training\n\nDuring training we cut random clips of **T=64** frames each **S=3** of them ending **.5-1.5 seconds** before **time_of_event** for positives. Leading to about **6.4 seconds** long at about **10 fps** clips. For negatives we just random cut at any possible time. Afterwards, a random **30 degrees** rotation is applied (to avoid too unrealistic images). Next, we cut 224x224 random crops from initial 256x256 resized clips. Finally we apply **Gaussian Noise**, **Brightness** and **Contrast** augmentations. We experimented with **Mixup** too but without clear improvements.\n\nPlane **Cross Entropy Loss** have been used." } }