# For more details on models, see https://pytorch.org/vision/main/models.html

# EfficientNet models: Designed for efficiency with compound scaling of depth, width, and resolution.
# These models balance accuracy and computational efficiency, ideal for resource-constrained environments.
efficientnet_b0:
  metrics:
    Acc@1: 77.692  # Top-1 accuracy on ImageNet
    Acc@5: 93.532  # Top-5 accuracy on ImageNet
    GFLOPS: 0.39   # Computational complexity
    Params: 5.3M   # Number of parameters
  model_fn: models.efficientnet_b0
  resolution: 224  # Input image resolution
  weights: models.EfficientNet_B0_Weights.IMAGENET1K_V1  # Pretrained weights on ImageNet

efficientnet_b1:
  metrics:
    Acc@1: 78.642
    Acc@5: 94.186
    GFLOPS: 0.69
    Params: 7.8M
  model_fn: models.efficientnet_b1
  resolution: 240
  weights: models.EfficientNet_B1_Weights.IMAGENET1K_V1

efficientnet_b2:
  metrics:
    Acc@1: 80.608
    Acc@5: 95.31
    GFLOPS: 1.09
    Params: 9.1M
  model_fn: models.efficientnet_b2
  resolution: 260
  weights: models.EfficientNet_B2_Weights.IMAGENET1K_V1

efficientnet_b3:
  metrics:
    Acc@1: 82.008
    Acc@5: 96.054
    GFLOPS: 1.83
    Params: 12.2M
  model_fn: models.efficientnet_b3
  resolution: 300
  weights: models.EfficientNet_B3_Weights.IMAGENET1K_V1

efficientnet_b4:
  metrics:
    Acc@1: 83.384
    Acc@5: 96.594
    GFLOPS: 4.39
    Params: 19.3M
  model_fn: models.efficientnet_b4
  resolution: 380
  weights: models.EfficientNet_B4_Weights.IMAGENET1K_V1

efficientnet_b5:
  metrics:
    Acc@1: 83.444
    Acc@5: 96.628
    GFLOPS: 10.27
    Params: 30.4M
  model_fn: models.efficientnet_b5
  resolution: 456
  weights: models.EfficientNet_B5_Weights.IMAGENET1K_V1

efficientnet_b6:
  metrics:
    Acc@1: 84.008
    Acc@5: 96.916
    GFLOPS: 19.07
    Params: 43.0M
  model_fn: models.efficientnet_b6
  resolution: 528
  weights: models.EfficientNet_B6_Weights.IMAGENET1K_V1

efficientnet_b7:
  metrics:
    Acc@1: 84.122
    Acc@5: 96.908
    GFLOPS: 37.75
    Params: 66.3M
  model_fn: models.efficientnet_b7
  resolution: 600
  weights: models.EfficientNet_B7_Weights.IMAGENET1K_V1

# EfficientNet V2 models: Improved training efficiency and performance over V1.
# These models use progressive learning and optimized scaling for better accuracy.
efficientnet_v2_l:
  metrics:
    Acc@1: 85.808
    Acc@5: 97.788
    GFLOPS: 56.08
    Params: 118.5M
  model_fn: models.efficientnet_v2_l
  resolution: 480
  weights: models.EfficientNet_V2_L_Weights.IMAGENET1K_V1

efficientnet_v2_m:
  metrics:
    Acc@1: 85.112
    Acc@5: 97.156
    GFLOPS: 24.58
    Params: 54.1M
  model_fn: models.efficientnet_v2_m
  resolution: 480
  weights: models.EfficientNet_V2_M_Weights.IMAGENET1K_V1

efficientnet_v2_s:
  metrics:
    Acc@1: 84.228
    Acc@5: 96.878
    GFLOPS: 8.37
    Params: 21.5M
  model_fn: models.efficientnet_v2_s
  resolution: 384
  weights: models.EfficientNet_V2_S_Weights.IMAGENET1K_V1

# RegNet models: Designed for scalability and efficiency with a focus on network design.
# These models optimize for both accuracy and computational efficiency.
regnet_y_128gf:
  metrics:
    Acc@1: 86.068  # High accuracy but computationally expensive
    Acc@5: 97.844
    GFLOPS: 127.52
    Params: 644.8M
  model_fn: models.regnet_y_128gf
  resolution: 224
  weights: models.RegNet_Y_128GF_Weights.IMAGENET1K_SWAG_LINEAR_V1

regnet_y_16gf:
  metrics:
    Acc@1: 82.886
    Acc@5: 96.328
    GFLOPS: 15.91
    Params: 83.6M
  model_fn: models.regnet_y_16gf
  resolution: 224
  weights: models.RegNet_Y_16GF_Weights.IMAGENET1K_V2

regnet_y_1_6gf:
  metrics:
    Acc@1: 80.876
    Acc@5: 95.444
    GFLOPS: 1.61
    Params: 11.2M
  model_fn: models.regnet_y_1_6gf
  resolution: 224
  weights: models.RegNet_Y_1_6GF_Weights.IMAGENET1K_V2

regnet_y_32gf:
  metrics:
    Acc@1: 83.368
    Acc@5: 96.498
    GFLOPS: 32.28
    Params: 145.0M
  model_fn: models.regnet_y_32gf
  resolution: 224
  weights: models.RegNet_Y_32GF_Weights.IMAGENET1K_V2

regnet_y_3_2gf:
  metrics:
    Acc@1: 81.982
    Acc@5: 95.972
    GFLOPS: 3.18
    Params: 19.4M
  model_fn: models.regnet_y_3_2gf
  resolution: 224
  weights: models.RegNet_Y_3_2GF_Weights.IMAGENET1K_V2

regnet_y_400mf:
  metrics:
    Acc@1: 75.804
    Acc@5: 92.742
    GFLOPS: 0.4
    Params: 4.3M
  model_fn: models.regnet_y_400mf
  resolution: 224
  weights: models.RegNet_Y_400MF_Weights.IMAGENET1K_V2

regnet_y_800mf:
  metrics:
    Acc@1: 78.828
    Acc@5: 94.502
    GFLOPS: 0.83
    Params: 6.4M
  model_fn: models.regnet_y_800mf
  resolution: 224
  weights: models.RegNet_Y_800MF_Weights.IMAGENET1K_V2

regnet_y_8gf:
  metrics:
    Acc@1: 82.828
    Acc@5: 96.33
    GFLOPS: 8.47
    Params: 39.4M
  model_fn: models.regnet_y_8gf
  resolution: 224
  weights: models.RegNet_Y_8GF_Weights.IMAGENET1K_V2

# Vision Transformer (ViT) models: Transformer-based architecture for image classification.
# These models excel in capturing long-range dependencies but require significant compute for larger variants.
vit_b_16:
  metrics:
    Acc@1: 81.072  # Base ViT model with balanced accuracy and efficiency
    Acc@5: 95.318
    GFLOPS: 17.56
    Params: 86.6M
  model_fn: models.vit_b_16
  resolution: 224
  weights: models.ViT_B_16_Weights.IMAGENET1K_V1

vit_b_32:
  metrics:
    Acc@1: 75.912  # Smaller patch size version of ViT, lower accuracy but fewer computations
    Acc@5: 92.466
    GFLOPS: 4.41
    Params: 88.2M
  model_fn: models.vit_b_32
  resolution: 224
  weights: models.ViT_B_32_Weights.IMAGENET1K_V1

vit_h_14:
  metrics:
    Acc@1: 88.552  # High-performance ViT model with very high accuracy and computational cost
    Acc@5: 98.694
    GFLOPS: 1016.72
    Params: 633.5M
  model_fn: models.vit_h_14
  resolution: 224
  weights: models.ViT_H_14_Weights.IMAGENET1K_SWAG_E2E_V1

vit_l_16:
  metrics:
    Acc@1: 79.662  # Larger ViT model with improved accuracy over base models
    Acc@5: 94.638
    GFLOPS: 61.55
    Params: 304.3M
  model_fn: models.vit_l_16
  resolution: 224
  weights: models.ViT_L_16_Weights.IMAGENET1K_V1

vit_l_32:
  metrics:
    Acc@1: 76.972  # Larger ViT with larger patch size, trading accuracy for reduced compute
    Acc@5: 93.07
    GFLOPS: 15.38
    Params: 306.5M
  model_fn: models.vit_l_32
  resolution: 224
  weights: models.ViT_L_32_Weights.IMAGENET1K_V1