Upload models

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+i3d_r50/i3d_training_history.png filter=lfs diff=lfs merge=lfs -text
+r2plus1d_18/r2plus1d_training_history.png filter=lfs diff=lfs merge=lfs -text
+slowfast_r50/slowfast_training_history.png filter=lfs diff=lfs merge=lfs -text
+x3d_m/x3d_training_history.png filter=lfs diff=lfs merge=lfs -text

i3d_r50/MODEL_CARD.md

@@ -0,0 +1,58 @@
+# I3D-R50 for Violence Detection
+
+Fine-tuned **I3D (Inflated 3D ConvNet)** với backbone **ResNet-50** cho bài toán phát hiện bạo lực.
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| Validation Accuracy | ~90.75% |
+| Parameters | 27.2M |
+| Checkpoint size | ~312 MB |
+| GPU Memory (inference) | ~0.6 GB |
+| Latency (CPU, 32 frames) | ~1.5-2.5 s |
+
+## Architecture
+
+- Backbone: **i3d_r50** (PyTorchVideo), pretrained Kinetics-400
+- "Inflated" 2D ResNet kernels thành 3D
+- Input: `(B, 3, 32, 224, 224)`, mean=0.45, std=0.225
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `i3d_r50_best.pt` | Best checkpoint |
+| `i3d_summary.json` | Training/eval metrics |
+| `i3d_confusion_matrix.png` | Confusion matrix |
+| `i3d_training_history.png` | Training curves |
+
+## Usage
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+from pytorchvideo.models.hub import i3d_r50
+
+ckpt_path = hf_hub_download(
+    repo_id="nauthui7/school-violence-detection-models",
+    filename="i3d_r50/i3d_r50_best.pt",
+)
+
+model = i3d_r50(pretrained=False)
+# Replace head — chi tiết xem model/i3d_model.py trong source code
+...
+```
+
+## Use Case
+
+I3D là model "kinh điển" cho video action recognition. Phù hợp:
+
+- Baseline so sánh với các kiến trúc khác
+- Khi cần inference trên GPU
+- Nghiên cứu / phân tích feature transfer learning
+
+## Limitations
+
+- Accuracy thấp nhất trong 4 models (~90.75%).
+- Checkpoint lớn (312 MB), không phù hợp deploy edge.

i3d_r50/i3d_r50_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:367e6e69ca934402711b1b66e14684f7ab2ef6faa501714f5d79b7525be2dd07
+size 327213702

i3d_r50/i3d_summary.json

@@ -0,0 +1,28 @@
+{
+  "model": "i3d_r50",
+  "best_accuracy": 0.9075000214576722,
+  "best_epoch": 10,
+  "total_epochs": 16,
+  "evaluation": {
+    "accuracy": 0.9075,
+    "precision": 0.9222797927461139,
+    "recall": 0.89,
+    "f1_score": 0.905852417302799
+  },
+  "config": {
+    "epochs": 35,
+    "batch_size": 24,
+    "learning_rate": 0.0002,
+    "weight_decay": 0.01,
+    "dropout": 0.4,
+    "mixup_alpha": 0.2,
+    "warmup_epochs": 3,
+    "num_frames": 32,
+    "frame_size": 224
+  },
+  "parameters": {
+    "total": 27227970,
+    "trainable": 27227970,
+    "frozen": 0
+  }
+}

r2plus1d_18/MODEL_CARD.md

@@ -0,0 +1,65 @@
+# R(2+1)D-18 for Violence Detection
+
+Fine-tuned **R(2+1)D-18** - factorized 3D convolution (decompose 3D conv thành 2D spatial + 1D temporal).
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| Validation Accuracy | **~93.0%** (best in this model zoo) |
+| Parameters | 31.3M |
+| Checkpoint size | ~358 MB |
+| GPU Memory (inference) | ~0.5 GB |
+| Latency (CPU, 32 frames) | ~2-3 s |
+
+## Architecture
+
+- Backbone: **r2plus1d_18** (torchvision), pretrained Kinetics-400
+- Factorized: 3D conv (t×h×w) → 2D conv (1×h×w) + 1D conv (t×1×1)
+- Tăng độ phi tuyến giữa spatial/temporal modeling
+- Input: `(B, 3, 32, 224, 224)`, mean=0.45, std=0.225
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `r2plus1d_best.pt` | Best checkpoint |
+| `r2plus1d_confusion_matrix.png` | Confusion matrix |
+| `r2plus1d_training_history.png` | Training curves |
+
+## Usage
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+from torchvision.models.video import r2plus1d_18
+
+ckpt_path = hf_hub_download(
+    repo_id="nauthui7/school-violence-detection-models",
+    filename="r2plus1d_18/r2plus1d_best.pt",
+)
+
+model = r2plus1d_18(weights=None)
+model.fc = torch.nn.Linear(model.fc.in_features, 2)
+
+ckpt = torch.load(ckpt_path, map_location="cpu")
+state_dict = ckpt.get("model", ckpt.get("model_state_dict", ckpt))
+model.load_state_dict(state_dict)
+model.eval()
+```
+
+> Khuyến nghị: dùng `ViolenceR2Plus1D` wrapper trong `model/r2plus1d_model.py`.
+
+## Use Case
+
+R(2+1)D-18 đạt accuracy **cao nhất** trong bộ 4 model. Phù hợp:
+
+- Khi accuracy là ưu tiên hàng đầu
+- Có GPU để inference (CPU vẫn chạy được nhưng chậm)
+- Production-like deployment với SLA chấp nhận ~2-3s/clip
+
+## Limitations
+
+- Checkpoint lớn (358 MB).
+- Vẫn cần input clip 32 frames (không streaming).
+- Tốc độ CPU không nhanh bằng X3D-M.

r2plus1d_18/r2plus1d_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa5aa0d7d4427ea7b81d410f5b5186ed64285ff8bebdd12462d976a36e86cdc0
+size 375895599

slowfast_r50/MODEL_CARD.md

@@ -0,0 +1,61 @@
+# SlowFast-R50 for Violence Detection
+
+Fine-tuned **SlowFast** với ResNet-50 backbone - kiến trúc dual-pathway cho video action recognition.
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| Validation Accuracy | ~92.5% |
+| Parameters | 33.6M |
+| Checkpoint size | ~386 MB |
+| GPU Memory (inference) | ~1.8 GB |
+| Latency (CPU, 32 frames) | ~3-5 s |
+
+## Architecture
+
+- Backbone: **slowfast_r50** (PyTorchVideo), pretrained Kinetics-400
+- **Slow pathway**: low frame rate, high channel - capture spatial semantic
+- **Fast pathway**: high frame rate, low channel - capture motion
+- Input: list `[slow_clip, fast_clip]`, slow=8 frames, fast=32 frames
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `slowfast_best.pt` | Best checkpoint |
+| `slowfast_summary.json` | Training/eval metrics |
+| `slowfast_confusion_matrix.png` | Confusion matrix |
+| `slowfast_training_history.png` | Training curves |
+
+## Usage
+
+> ⚠️ SlowFast cần input đặc biệt (pack pathway). Dùng `ViolenceSlowFast` wrapper trong `model/slowfast_model.py` để tự handle pathway splitting.
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+
+ckpt_path = hf_hub_download(
+    repo_id="nauthui7/school-violence-detection-models",
+    filename="slowfast_r50/slowfast_best.pt",
+)
+
+# Khuyến nghị: dùng wrapper từ source code
+# from model.slowfast_model import create_slowfast_model
+# model = create_slowfast_model(num_classes=2, pretrained=False)
+```
+
+## Use Case
+
+SlowFast là model có accuracy cao thứ 2, đặc biệt mạnh khi:
+
+- Video có cả chi tiết tĩnh và chuyển động nhanh (xô đẩy, đánh nhau)
+- Cần phân tích kỹ temporal dynamics
+- Có GPU đủ mạnh (≥6GB VRAM)
+
+## Limitations
+
+- Inference chậm trên CPU do hai pathway.
+- Memory footprint lớn nhất (~1.8GB GPU).
+- Cần preprocessing đặc biệt (pack pathways).

slowfast_r50/slowfast_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c33809b5d72e015c391bc9bb4afb3fa2f7319bc60e8e848984ab33215470dc3
+size 404601565

slowfast_r50/slowfast_summary.json

@@ -0,0 +1,29 @@
+{
+  "model": "slowfast_r50",
+  "best_accuracy": 0.9250000238418579,
+  "best_epoch": 3,
+  "total_epochs": 10,
+  "evaluation": {
+    "accuracy": 0.925,
+    "precision": 0.9047619047619048,
+    "recall": 0.95,
+    "f1_score": 0.926829268292683
+  },
+  "config": {
+    "epochs": 40,
+    "batch_size": 48,
+    "learning_rate": 0.0003,
+    "weight_decay": 0.02,
+    "dropout": 0.4,
+    "gradient_clip": 1.0,
+    "mixup_alpha": 0.2,
+    "warmup_epochs": 5,
+    "num_frames": 32,
+    "frame_size": 224
+  },
+  "parameters": {
+    "total": 33649098,
+    "trainable": 33649098,
+    "frozen": 0
+  }
+}

x3d_m/MODEL_CARD.md

@@ -0,0 +1,64 @@
+# X3D-M for Violence Detection
+
+Fine-tuned **X3D-M** (Expand-3D Medium) cho bài toán nhận diện hành vi bạo lực 2 lớp trên video.
+
+## Performance
+
+| Metric | Value |
+|--------|-------|
+| Validation Accuracy | ~92.0% |
+| Parameters | 3.0M |
+| Checkpoint size | ~35 MB |
+| GPU Memory (inference) | ~0.2 GB |
+| Latency (CPU, 32 frames) | ~300-500 ms |
+
+## Architecture
+
+- Backbone: **x3d_m** (PyTorchVideo), pretrained Kinetics-400
+- Head: Linear(2048 → 2) thay cho head 400-class gốc
+- Input: `(B, 3, 32, 224, 224)`, normalized với mean=0.45, std=0.225
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `x3d_m_best.pt` | Best checkpoint (state_dict + metadata) |
+| `x3d_summary.json` | Training/eval metrics summary |
+| `x3d_confusion_matrix.png` | Confusion matrix trên test set |
+| `x3d_training_history.png` | Loss/accuracy theo epoch |
+
+## Usage
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+from pytorchvideo.models.hub import x3d_m
+
+ckpt_path = hf_hub_download(
+    repo_id="nauthui7/school-violence-detection-models",
+    filename="x3d_m/x3d_m_best.pt",
+)
+
+model = x3d_m(pretrained=False)
+model.blocks[5].proj = torch.nn.Linear(2048, 2)
+
+ckpt = torch.load(ckpt_path, map_location="cpu")
+state_dict = ckpt.get("model", ckpt.get("model_state_dict", ckpt))
+model.load_state_dict(state_dict)
+model.eval()
+```
+
+> Khuyến nghị: dùng class wrapper `ViolenceX3D` trong `model/x3d_model.py` (repo source) thay vì tự build head, để đảm bảo state_dict key khớp 100%.
+
+## Use Case
+
+X3D-M là model **nhỏ nhất và nhanh nhất** trong bộ 4 model. Phù hợp:
+
+- Real-time inference trên CPU (laptop, edge devices)
+- Demo trên Hugging Face Spaces free tier
+- Khi cần throughput cao
+
+## Limitations
+
+- Accuracy thấp hơn ~1% so với R(2+1)D-18 và SlowFast-R50.
+- Vẫn dùng input clip 32 frames - không phải streaming.

x3d_m/x3d_m_best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b8e4f95d3b7aa0d9df941f3fffbf07b8b35dac82ded4e994c60b91daa138a50
+size 36353579

x3d_m/x3d_summary.json

@@ -0,0 +1,28 @@
+{
+  "model": "x3d_m",
+  "best_accuracy": 0.92,
+  "best_epoch": 14,
+  "total_epochs": 18,
+  "evaluation": {
+    "accuracy": 0.92,
+    "precision": 0.9421052631578948,
+    "recall": 0.895,
+    "f1_score": 0.9179487179487179
+  },
+  "config": {
+    "epochs": 30,
+    "batch_size": 32,
+    "learning_rate": 0.0001,
+    "weight_decay": 0.01,
+    "dropout": 0.3,
+    "mixup_alpha": 0.2,
+    "warmup_epochs": 3,
+    "num_frames": 32,
+    "frame_size": 224
+  },
+  "parameters": {
+    "total": 2978772,
+    "trainable": 2978772,
+    "frozen": 0
+  }
+}