Upload 4 files

README.md

@@ -0,0 +1,276 @@
+# TwinCar: Deep Learning for Automotive Classification
+
+TwinCar is a deep learning project focused on vehicle classification and automotive attribute prediction. It explores and compares multiple state-of-the-art convolutional neural networks (CNNs) and Vision Transformer architectures using transfer learning techniques on public automotive datasets.
+
+The project evaluates model performance across different datasets and training strategies, including fine-tuning and feature extraction, while providing a reproducible workflow for data preparation, training, evaluation, and inference.
+
+## Features
+
+* Comparison of multiple deep learning architectures:
+
+  * EfficientNet-B0
+  * ConvNeXt-Tiny
+  * Vision Transformer (ViT-B/16)
+  * Swin Transformer (Swin-T)
+  * DeiT
+* Support for multiple automotive datasets:
+
+  * Stanford Cars
+  * CompCars
+* Transfer learning experiments with both fine-tuning and frozen-backbone approaches
+* Comprehensive evaluation metrics and visualizations
+* Batch inference demonstrations
+* Structured and reproducible notebook-based workflow
+
+---
+
+## Project Structure
+
+```text
+TwinCar/
+├── notebooks/
+│   ├── 01_data_exploration.ipynb
+│   ├── 02_data_preparation.ipynb
+│   │
+│   ├── EfficientNet Experiments
+│   ├── 03a_efficientnet_b0_v1.ipynb
+│   ├── 03a_efficientnet_b0_v1_evaluation.ipynb
+│   ├── 03b_efficientnet_b0_v2.ipynb
+│   ├── 03b_efficientnet_b0_v2_evaluation.ipynb
+│   │
+│   ├── ConvNeXt Experiments
+│   ├── 04_convnext.ipynb
+│   ├── 04_convnext_tiny_evaluation.ipynb
+│   │
+│   ├── Stanford Cars Models
+│   ├── ConvNeXt_Tiny - Stanford Cars.ipynb
+│   ├── ConvNeXt_Tiny with freezing - Stanford Cars.ipynb
+│   ├── ConvNeXt_Tiny with freezing v2 - Stanford Cars.ipynb
+│   ├── EfficientNet_B0 - Stanford Cars.ipynb
+│   ├── Swin_T - Stanford Cars.ipynb
+│   ├── ViT_B_16_StanfordCars_model.ipynb
+│   ├── deit_tiny_patch16_224 - Stanford Cars.ipynb
+│   │
+│   ├── CompCars Models
+│   ├── 06_compcars_efficientnet_b0_make_model_year.ipynb
+│   ├── EfficientNet_B0 - Comp Cars.ipynb
+│   ├── ConvNeXt_Tiny - Comp Cars.ipynb
+│   ├── CompCars_ViT_model.ipynb
+│   │
+│   ├── 05_batch_prediction_demo.ipynb
+│   └── ml-final-project.ipynb
+│
+├── models/
+├── reports/
+├── scripts/
+├── requirements.txt
+└── README.md
+```
+
+---
+
+## Datasets
+
+### Stanford Cars
+
+A fine-grained vehicle classification dataset containing approximately 16,000 images across 196 vehicle categories. It is widely used for benchmarking car recognition models.
+
+### CompCars
+
+A large-scale automotive dataset containing over 30,000 images with detailed annotations, including vehicle make, model, and year. It is suitable for both classification and attribute prediction tasks.
+
+---
+
+## Model Architectures
+
+### CNN-Based Models
+
+#### EfficientNet-B0
+
+EfficientNet uses compound scaling to balance network depth, width, and resolution. Multiple versions are included to evaluate the impact of training and optimization strategies.
+
+#### ConvNeXt-Tiny
+
+A modern CNN architecture inspired by Vision Transformers while retaining the efficiency and simplicity of convolutional networks. Experiments include both fully trainable and partially frozen variants.
+
+### Transformer-Based Models
+
+#### Vision Transformer (ViT-B/16)
+
+A pure transformer architecture that processes images as sequences of patches for image classification.
+
+#### Swin Transformer (Swin-T)
+
+A hierarchical transformer architecture that uses shifted-window attention for efficient feature extraction.
+
+#### DeiT
+
+A data-efficient transformer model designed to achieve strong performance with reduced training requirements.
+
+---
+
+## Installation
+
+### Prerequisites
+
+* Python 3.8+
+* CUDA-capable GPU (recommended)
+* Jupyter Notebook
+
+### Clone the Repository
+
+```bash
+git clone https://github.com/dragicakostoska/TwinCar.git
+cd TwinCar
+```
+
+### Install Dependencies
+
+```bash
+pip install -r requirements.txt
+```
+
+---
+
+## Dependencies
+
+Core libraries used throughout the project include:
+
+* PyTorch and TorchVision
+* NumPy and Pandas
+* Pillow
+* Hugging Face Datasets
+* Scikit-learn
+* Matplotlib
+* tqdm
+* Jupyter Notebook and IPython Kernel
+
+---
+
+## Workflow
+
+### 1. Data Exploration
+
+`01_data_exploration.ipynb`
+
+* Explore dataset characteristics
+* Visualize class distributions
+* Inspect image samples and dataset statistics
+
+### 2. Data Preparation
+
+`02_data_preparation.ipynb`
+
+* Apply preprocessing and augmentation techniques
+* Create training, validation, and test splits
+* Build dataset loaders and transformations
+
+### 3. Model Training
+
+* Select the desired architecture notebook
+* Configure hyperparameters
+* Train using transfer learning or fine-tuning
+* Monitor performance throughout training
+
+### 4. Evaluation
+
+* Analyze classification metrics
+* Generate confusion matrices
+* Visualize training and validation curves
+* Compare model performance across architectures
+
+### 5. Inference
+
+`05_batch_prediction_demo.ipynb`
+
+* Load trained models
+* Run predictions on image batches
+* Visualize outputs and confidence scores
+
+---
+
+## Training Strategies
+
+The project investigates two common transfer learning approaches:
+
+### Fine-Tuning
+
+All network layers are trained starting from pretrained weights, allowing the model to adapt fully to the target dataset.
+
+### Feature Extraction
+
+Earlier layers are frozen while only the classification head is trained. This reduces training time and helps preserve pretrained feature representations.
+
+---
+
+## Evaluation Metrics
+
+Each evaluation notebook provides:
+
+* Top-1 and Top-5 Accuracy
+* Precision, Recall, and F1 Score
+* Per-class performance analysis
+* Confusion matrices
+* Training and validation loss curves
+* Inference speed comparisons
+* Prediction visualizations
+
+---
+
+## Key Observations
+
+* EfficientNet-B0 v2 improves upon the baseline v1 configuration.
+* ConvNeXt-Tiny achieves strong performance while maintaining computational efficiency.
+* Transformer-based architectures provide competitive results and different representational advantages compared to CNNs.
+* Transfer learning significantly reduces training requirements while maintaining strong classification accuracy.
+
+---
+
+## Customization
+
+The project can be extended in several ways:
+
+* Integrate additional automotive datasets
+* Add new model architectures
+* Experiment with alternative hyperparameters
+* Explore multi-task learning objectives
+* Implement custom data augmentation pipelines
+
+---
+
+## Future Work
+
+* Convert notebook workflows into modular Python packages
+* Implement model ensembling techniques
+* Add advanced augmentation methods such as MixUp and RandAugment
+* Explore knowledge distillation strategies
+* Optimize deployment using ONNX or TensorRT
+* Develop an inference API
+* Create a unified benchmark report across all experiments
+
+---
+
+## References
+
+* EfficientNet — *Scaling Convolutional Neural Networks Efficiently*
+* ConvNeXt — *A ConvNet for the 2020s*
+* Vision Transformer — *An Image is Worth 16×16 Words*
+* Swin Transformer — *Hierarchical Vision Transformer Using Shifted Windows*
+* DeiT — *Data-efficient Image Transformers*
+
+---
+
+## Contributing
+
+Contributions, suggestions, and bug reports are welcome. Feel free to open an issue or submit a pull request.
+
+---
+
+## License
+
+This project is provided for educational and research purposes.
+
+
+---
+
+**Last Updated:** June 2026

best_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:634362d34fb6c9dcd1484278157f395482f4ba0b732bf114e2cff1a46ec43aa3
+size 111947591

idx_to_class.json

@@ -0,0 +1,197 @@
+{
+  "0": "AM_General_Hummer_SUV_2000",
+  "1": "Acura_Integra_Type_R_2001",
+  "2": "Acura_RL_Sedan_2012",
+  "3": "Acura_TL_Sedan_2012",
+  "4": "Acura_TL_Type-S_2008",
+  "5": "Acura_TSX_Sedan_2012",
+  "6": "Acura_ZDX_Hatchback_2012",
+  "7": "Aston_Martin_V8_Vantage_Convertible_2012",
+  "8": "Aston_Martin_V8_Vantage_Coupe_2012",
+  "9": "Aston_Martin_Virage_Convertible_2012",
+  "10": "Aston_Martin_Virage_Coupe_2012",
+  "11": "Audi_100_Sedan_1994",
+  "12": "Audi_100_Wagon_1994",
+  "13": "Audi_A5_Coupe_2012",
+  "14": "Audi_R8_Coupe_2012",
+  "15": "Audi_RS_4_Convertible_2008",
+  "16": "Audi_S4_Sedan_2007",
+  "17": "Audi_S4_Sedan_2012",
+  "18": "Audi_S5_Convertible_2012",
+  "19": "Audi_S5_Coupe_2012",
+  "20": "Audi_S6_Sedan_2011",
+  "21": "Audi_TTS_Coupe_2012",
+  "22": "Audi_TT_Hatchback_2011",
+  "23": "Audi_TT_RS_Coupe_2012",
+  "24": "Audi_V8_Sedan_1994",
+  "25": "BMW_1_Series_Convertible_2012",
+  "26": "BMW_1_Series_Coupe_2012",
+  "27": "BMW_3_Series_Sedan_2012",
+  "28": "BMW_3_Series_Wagon_2012",
+  "29": "BMW_6_Series_Convertible_2007",
+  "30": "BMW_ActiveHybrid_5_Sedan_2012",
+  "31": "BMW_M3_Coupe_2012",
+  "32": "BMW_M5_Sedan_2010",
+  "33": "BMW_M6_Convertible_2010",
+  "34": "BMW_X3_SUV_2012",
+  "35": "BMW_X5_SUV_2007",
+  "36": "BMW_X6_SUV_2012",
+  "37": "BMW_Z4_Convertible_2012",
+  "38": "Bentley_Arnage_Sedan_2009",
+  "39": "Bentley_Continental_Flying_Spur_Sedan_2007",
+  "40": "Bentley_Continental_GT_Coupe_2007",
+  "41": "Bentley_Continental_GT_Coupe_2012",
+  "42": "Bentley_Continental_Supersports_Conv._Convertible_2012",
+  "43": "Bentley_Mulsanne_Sedan_2011",
+  "44": "Bugatti_Veyron_16.4_Convertible_2009",
+  "45": "Bugatti_Veyron_16.4_Coupe_2009",
+  "46": "Buick_Enclave_SUV_2012",
+  "47": "Buick_Rainier_SUV_2007",
+  "48": "Buick_Regal_GS_2012",
+  "49": "Buick_Verano_Sedan_2012",
+  "50": "Cadillac_CTS-V_Sedan_2012",
+  "51": "Cadillac_Escalade_EXT_Crew_Cab_2007",
+  "52": "Cadillac_SRX_SUV_2012",
+  "53": "Chevrolet_Avalanche_Crew_Cab_2012",
+  "54": "Chevrolet_Camaro_Convertible_2012",
+  "55": "Chevrolet_Cobalt_SS_2010",
+  "56": "Chevrolet_Corvette_Convertible_2012",
+  "57": "Chevrolet_Corvette_Ron_Fellows_Edition_Z06_2007",
+  "58": "Chevrolet_Corvette_ZR1_2012",
+  "59": "Chevrolet_Express_Cargo_Van_2007",
+  "60": "Chevrolet_Express_Van_2007",
+  "61": "Chevrolet_HHR_SS_2010",
+  "62": "Chevrolet_Impala_Sedan_2007",
+  "63": "Chevrolet_Malibu_Hybrid_Sedan_2010",
+  "64": "Chevrolet_Malibu_Sedan_2007",
+  "65": "Chevrolet_Monte_Carlo_Coupe_2007",
+  "66": "Chevrolet_Silverado_1500_Classic_Extended_Cab_2007",
+  "67": "Chevrolet_Silverado_1500_Extended_Cab_2012",
+  "68": "Chevrolet_Silverado_1500_Hybrid_Crew_Cab_2012",
+  "69": "Chevrolet_Silverado_1500_Regular_Cab_2012",
+  "70": "Chevrolet_Silverado_2500HD_Regular_Cab_2012",
+  "71": "Chevrolet_Sonic_Sedan_2012",
+  "72": "Chevrolet_Tahoe_Hybrid_SUV_2012",
+  "73": "Chevrolet_TrailBlazer_SS_2009",
+  "74": "Chevrolet_Traverse_SUV_2012",
+  "75": "Chrysler_300_SRT-8_2010",
+  "76": "Chrysler_Aspen_SUV_2009",
+  "77": "Chrysler_Crossfire_Convertible_2008",
+  "78": "Chrysler_PT_Cruiser_Convertible_2008",
+  "79": "Chrysler_Sebring_Convertible_2010",
+  "80": "Chrysler_Town_and_Country_Minivan_2012",
+  "81": "Daewoo_Nubira_Wagon_2002",
+  "82": "Dodge_Caliber_Wagon_2007",
+  "83": "Dodge_Caliber_Wagon_2012",
+  "84": "Dodge_Caravan_Minivan_1997",
+  "85": "Dodge_Challenger_SRT8_2011",
+  "86": "Dodge_Charger_SRT-8_2009",
+  "87": "Dodge_Charger_Sedan_2012",
+  "88": "Dodge_Dakota_Club_Cab_2007",
+  "89": "Dodge_Dakota_Crew_Cab_2010",
+  "90": "Dodge_Durango_SUV_2007",
+  "91": "Dodge_Durango_SUV_2012",
+  "92": "Dodge_Journey_SUV_2012",
+  "93": "Dodge_Magnum_Wagon_2008",
+  "94": "Dodge_Ram_Pickup_3500_Crew_Cab_2010",
+  "95": "Dodge_Ram_Pickup_3500_Quad_Cab_2009",
+  "96": "Dodge_Sprinter_Cargo_Van_2009",
+  "97": "Eagle_Talon_Hatchback_1998",
+  "98": "FIAT_500_Abarth_2012",
+  "99": "FIAT_500_Convertible_2012",
+  "100": "Ferrari_458_Italia_Convertible_2012",
+  "101": "Ferrari_458_Italia_Coupe_2012",
+  "102": "Ferrari_California_Convertible_2012",
+  "103": "Ferrari_FF_Coupe_2012",
+  "104": "Fisker_Karma_Sedan_2012",
+  "105": "Ford_E-Series_Wagon_Van_2012",
+  "106": "Ford_Edge_SUV_2012",
+  "107": "Ford_Expedition_EL_SUV_2009",
+  "108": "Ford_F-150_Regular_Cab_2007",
+  "109": "Ford_F-150_Regular_Cab_2012",
+  "110": "Ford_F-450_Super_Duty_Crew_Cab_2012",
+  "111": "Ford_Fiesta_Sedan_2012",
+  "112": "Ford_Focus_Sedan_2007",
+  "113": "Ford_Freestar_Minivan_2007",
+  "114": "Ford_GT_Coupe_2006",
+  "115": "Ford_Mustang_Convertible_2007",
+  "116": "Ford_Ranger_SuperCab_2011",
+  "117": "GMC_Acadia_SUV_2012",
+  "118": "GMC_Canyon_Extended_Cab_2012",
+  "119": "GMC_Savana_Van_2012",
+  "120": "GMC_Terrain_SUV_2012",
+  "121": "GMC_Yukon_Hybrid_SUV_2012",
+  "122": "Geo_Metro_Convertible_1993",
+  "123": "HUMMER_H2_SUT_Crew_Cab_2009",
+  "124": "HUMMER_H3T_Crew_Cab_2010",
+  "125": "Honda_Accord_Coupe_2012",
+  "126": "Honda_Accord_Sedan_2012",
+  "127": "Honda_Odyssey_Minivan_2007",
+  "128": "Honda_Odyssey_Minivan_2012",
+  "129": "Hyundai_Accent_Sedan_2012",
+  "130": "Hyundai_Azera_Sedan_2012",
+  "131": "Hyundai_Elantra_Sedan_2007",
+  "132": "Hyundai_Elantra_Touring_Hatchback_2012",
+  "133": "Hyundai_Genesis_Sedan_2012",
+  "134": "Hyundai_Santa_Fe_SUV_2012",
+  "135": "Hyundai_Sonata_Hybrid_Sedan_2012",
+  "136": "Hyundai_Sonata_Sedan_2012",
+  "137": "Hyundai_Tucson_SUV_2012",
+  "138": "Hyundai_Veloster_Hatchback_2012",
+  "139": "Hyundai_Veracruz_SUV_2012",
+  "140": "Infiniti_G_Coupe_IPL_2012",
+  "141": "Infiniti_QX56_SUV_2011",
+  "142": "Isuzu_Ascender_SUV_2008",
+  "143": "Jaguar_XK_XKR_2012",
+  "144": "Jeep_Compass_SUV_2012",
+  "145": "Jeep_Grand_Cherokee_SUV_2012",
+  "146": "Jeep_Liberty_SUV_2012",
+  "147": "Jeep_Patriot_SUV_2012",
+  "148": "Jeep_Wrangler_SUV_2012",
+  "149": "Lamborghini_Aventador_Coupe_2012",
+  "150": "Lamborghini_Diablo_Coupe_2001",
+  "151": "Lamborghini_Gallardo_LP_570-4_Superleggera_2012",
+  "152": "Lamborghini_Reventon_Coupe_2008",
+  "153": "Land_Rover_LR2_SUV_2012",
+  "154": "Land_Rover_Range_Rover_SUV_2012",
+  "155": "Lincoln_Town_Car_Sedan_2011",
+  "156": "MINI_Cooper_Roadster_Convertible_2012",
+  "157": "Maybach_Landaulet_Convertible_2012",
+  "158": "Mazda_Tribute_SUV_2011",
+  "159": "McLaren_MP4-12C_Coupe_2012",
+  "160": "Mercedes-Benz_300-Class_Convertible_1993",
+  "161": "Mercedes-Benz_C-Class_Sedan_2012",
+  "162": "Mercedes-Benz_E-Class_Sedan_2012",
+  "163": "Mercedes-Benz_S-Class_Sedan_2012",
+  "164": "Mercedes-Benz_SL-Class_Coupe_2009",
+  "165": "Mercedes-Benz_Sprinter_Van_2012",
+  "166": "Mitsubishi_Lancer_Sedan_2012",
+  "167": "Nissan_240SX_Coupe_1998",
+  "168": "Nissan_Juke_Hatchback_2012",
+  "169": "Nissan_Leaf_Hatchback_2012",
+  "170": "Nissan_NV_Passenger_Van_2012",
+  "171": "Plymouth_Neon_Coupe_1999",
+  "172": "Porsche_Panamera_Sedan_2012",
+  "173": "Rolls-Royce_Ghost_Sedan_2012",
+  "174": "Rolls-Royce_Phantom_Drophead_Coupe_Convertible_2012",
+  "175": "Rolls-Royce_Phantom_Sedan_2012",
+  "176": "Scion_xD_Hatchback_2012",
+  "177": "Spyker_C8_Convertible_2009",
+  "178": "Spyker_C8_Coupe_2009",
+  "179": "Suzuki_Aerio_Sedan_2007",
+  "180": "Suzuki_Kizashi_Sedan_2012",
+  "181": "Suzuki_SX4_Hatchback_2012",
+  "182": "Suzuki_SX4_Sedan_2012",
+  "183": "Tesla_Model_S_Sedan_2012",
+  "184": "Toyota_4Runner_SUV_2012",
+  "185": "Toyota_Camry_Sedan_2012",
+  "186": "Toyota_Corolla_Sedan_2012",
+  "187": "Toyota_Sequoia_SUV_2012",
+  "188": "Volkswagen_Beetle_Hatchback_2012",
+  "189": "Volkswagen_Golf_Hatchback_1991",
+  "190": "Volkswagen_Golf_Hatchback_2012",
+  "191": "Volvo_240_Sedan_1993",
+  "192": "Volvo_C30_Hatchback_2012",
+  "193": "Volvo_XC90_SUV_2007",
+  "194": "smart_fortwo_Convertible_2012"
+}

train_config.json

@@ -0,0 +1,23 @@
+{
+    "model": "convnext_tiny",
+    "num_classes": 195,
+    "img_size": 224,
+    "batch_size": 32,
+    "seed": 42,
+    "epochs": 30,
+    "lr": 0.0003,
+    "weight_decay": 0.0001,
+    "patience": 5,
+    "lr_patience": 2,
+    "label_smoothing": 0.1,
+    "imagenet_mean": [
+        0.485,
+        0.456,
+        0.406
+    ],
+    "imagenet_std": [
+        0.229,
+        0.224,
+        0.225
+    ]
+}