Data Augmentation Experiment

I wrote an app to classify an object using transfer learning, and the model was trained on CIFAR-10 dataset, all by myself.

Effect of Data Augmentation on Validation Accuracy

Experiment Setup

We conducted a series of experiments to evaluate the effect of data augmentation on ResNet-18 validation performance. Two augmentation pipelines were tested: a standard augmentation and a more complex augmentation with stronger transformations.

Augmentation Pipelines

1. Standard Augmentation:

train_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.RandomHorizontalFlip(),
    transforms.RandomCrop(224, padding=8),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

2. Complex Augmentation:

train_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.RandomResizedCrop(224, scale=(0.7, 1.0)),
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(
        brightness=0.2,
        contrast=0.2,
        saturation=0.2,
        hue=0.05
    ),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

Validation Results

Analysis and Observations

- Introducing even a simple augmentation pipeline improved validation accuracy from 78.90% to over 82% at optimal N_AUG, highlighting the importance of data diversity in transfer learning.
- Increasing N_AUG beyond 4 did not further improve performance, suggesting that excessive augmentation may introduce redundant or noisy transformations.
- Surprisingly, the more complex augmentation pipeline resulted in slightly lower validation accuracy (~79.90%), indicating that overly aggressive transformations can distort input features beyond the range the pretrained model expects.
- These results confirm a key principle: augmentation should increase useful variability without deviating from the original feature distribution that the pretrained model has learned to interpret.

Conclusion

- Moderate augmentation with carefully tuned N_AUG is optimal for ResNet-18 fine-tuning on CIFAR-10.
- Excessive or complex augmentation may harm performance due to over-perturbation of input features.
- This experiment demonstrates that thoughtful augmentation is a simple yet highly effective lever to improve transfer learning outcomes.

Raw Log for Your Reference

Here's the training log for when there's no data augmentation, meaning train_transform looks like this:

train_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

Here's the corresponding training log:

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 1.3370 | Val Loss: 1.9445 | Train Acc: 55.20% | Val Acc: 54.10%
Epoch [02/20] Train Loss: 0.3165 | Val Loss: 2.0112 | Train Acc: 90.50% | Val Acc: 58.10%
Epoch [03/20] Train Loss: 0.1490 | Val Loss: 1.3208 | Train Acc: 96.30% | Val Acc: 68.60%
Epoch [04/20] Train Loss: 0.0754 | Val Loss: 1.0769 | Train Acc: 98.50% | Val Acc: 71.80%
Epoch [05/20] Train Loss: 0.0518 | Val Loss: 1.0204 | Train Acc: 98.90% | Val Acc: 71.90%
Epoch [06/20] Train Loss: 0.0154 | Val Loss: 0.9471 | Train Acc: 99.90% | Val Acc: 73.90%
Epoch [07/20] Train Loss: 0.0095 | Val Loss: 0.9050 | Train Acc: 100.00% | Val Acc: 75.10%
Epoch [08/20] Train Loss: 0.0071 | Val Loss: 0.8919 | Train Acc: 100.00% | Val Acc: 75.50%
Epoch [09/20] Train Loss: 0.0036 | Val Loss: 0.8561 | Train Acc: 100.00% | Val Acc: 75.80%
Epoch [10/20] Train Loss: 0.0023 | Val Loss: 0.8120 | Train Acc: 100.00% | Val Acc: 76.50%
Epoch [11/20] Train Loss: 0.0028 | Val Loss: 0.7939 | Train Acc: 100.00% | Val Acc: 76.90%
Epoch [12/20] Train Loss: 0.0016 | Val Loss: 0.7637 | Train Acc: 100.00% | Val Acc: 77.50%
Epoch [13/20] Train Loss: 0.0013 | Val Loss: 0.7415 | Train Acc: 100.00% | Val Acc: 78.30%
Epoch [14/20] Train Loss: 0.0012 | Val Loss: 0.7292 | Train Acc: 100.00% | Val Acc: 78.00%
Epoch [15/20] Train Loss: 0.0010 | Val Loss: 0.7215 | Train Acc: 100.00% | Val Acc: 78.50%
Epoch [16/20] Train Loss: 0.0009 | Val Loss: 0.7136 | Train Acc: 100.00% | Val Acc: 78.70%
Epoch [17/20] Train Loss: 0.0007 | Val Loss: 0.7087 | Train Acc: 100.00% | Val Acc: 78.90%
Epoch [18/20] Train Loss: 0.0006 | Val Loss: 0.7132 | Train Acc: 100.00% | Val Acc: 78.60%
Epoch [19/20] Train Loss: 0.0005 | Val Loss: 0.7098 | Train Acc: 100.00% | Val Acc: 78.70%
Epoch [20/20] Train Loss: 0.0006 | Val Loss: 0.7075 | Train Acc: 100.00% | Val Acc: 78.30%

Best results for resnet18:
Train Loss: 0.0007 | Val Loss: 0.7087 | Train Acc: 100.00% | Val Acc: 78.90%
Training Time: 76.41 seconds

From here on out, let's do data augmentation using standard image augmentation techniques, here's the transform code:

train_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.RandomHorizontalFlip(),
    transforms.RandomCrop(224, padding=8),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

Here's the training log for different values of N_AUG.

========== CONFIG ==========
NUM_CLASSES = 10
TRAINING_SAMPLE_PER_CLASS = 100
VALIDATION_SAMPLE_PER_CLASS = 100
BATCH_SIZE = 256
EPOCHS = 20
LR = 0.001
TRAINABLE_LAYERS = 2
EARLY_STOP_PATIENCE = 10
N_AUG = 1
DEVICE = cuda
============================

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 1.3549 | Val Loss: 1.5754 | Train Acc: 54.10% | Val Acc: 62.40%
Epoch [02/20] Train Loss: 0.4315 | Val Loss: 2.0113 | Train Acc: 86.00% | Val Acc: 61.90%
Epoch [03/20] Train Loss: 0.3334 | Val Loss: 1.7052 | Train Acc: 90.00% | Val Acc: 66.10%
Epoch [04/20] Train Loss: 0.2071 | Val Loss: 1.3286 | Train Acc: 93.40% | Val Acc: 69.70%
Epoch [05/20] Train Loss: 0.1246 | Val Loss: 1.1073 | Train Acc: 96.50% | Val Acc: 74.50%
Epoch [06/20] Train Loss: 0.0770 | Val Loss: 1.1168 | Train Acc: 97.80% | Val Acc: 73.80%
Epoch [07/20] Train Loss: 0.0704 | Val Loss: 1.0630 | Train Acc: 97.70% | Val Acc: 75.10%
Epoch [08/20] Train Loss: 0.0363 | Val Loss: 0.9845 | Train Acc: 99.30% | Val Acc: 76.10%
Epoch [09/20] Train Loss: 0.0340 | Val Loss: 0.8410 | Train Acc: 99.30% | Val Acc: 77.90%
Epoch [10/20] Train Loss: 0.0251 | Val Loss: 0.8184 | Train Acc: 99.80% | Val Acc: 78.70%
Epoch [11/20] Train Loss: 0.0250 | Val Loss: 0.8143 | Train Acc: 99.50% | Val Acc: 78.40%
Epoch [12/20] Train Loss: 0.0146 | Val Loss: 0.8361 | Train Acc: 99.80% | Val Acc: 78.10%
Epoch [13/20] Train Loss: 0.0115 | Val Loss: 0.7682 | Train Acc: 100.00% | Val Acc: 79.40%
Epoch [14/20] Train Loss: 0.0064 | Val Loss: 0.7347 | Train Acc: 100.00% | Val Acc: 80.00%
Epoch [15/20] Train Loss: 0.0059 | Val Loss: 0.7145 | Train Acc: 100.00% | Val Acc: 79.10%
Epoch [16/20] Train Loss: 0.0088 | Val Loss: 0.7160 | Train Acc: 99.80% | Val Acc: 80.50%
Epoch [17/20] Train Loss: 0.0044 | Val Loss: 0.7340 | Train Acc: 100.00% | Val Acc: 80.60%
Epoch [18/20] Train Loss: 0.0035 | Val Loss: 0.7265 | Train Acc: 100.00% | Val Acc: 80.10%
Epoch [19/20] Train Loss: 0.0040 | Val Loss: 0.7524 | Train Acc: 99.90% | Val Acc: 79.50%
Epoch [20/20] Train Loss: 0.0026 | Val Loss: 0.7415 | Train Acc: 100.00% | Val Acc: 79.40%

Best results for resnet18:
Train Loss: 0.0044 | Val Loss: 0.7340 | Train Acc: 100.00% | Val Acc: 80.60%
Training Time: 78.92 seconds


========== CONFIG ==========
NUM_CLASSES = 10
TRAINING_SAMPLE_PER_CLASS = 100
VALIDATION_SAMPLE_PER_CLASS = 100
BATCH_SIZE = 256
EPOCHS = 20
LR = 0.001
TRAINABLE_LAYERS = 2
EARLY_STOP_PATIENCE = 10
N_AUG = 3
DEVICE = cuda
============================

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 0.6881 | Val Loss: 1.4389 | Train Acc: 77.50% | Val Acc: 67.70%
Epoch [02/20] Train Loss: 0.1263 | Val Loss: 0.9544 | Train Acc: 96.27% | Val Acc: 77.40%
Epoch [03/20] Train Loss: 0.0448 | Val Loss: 0.8698 | Train Acc: 99.03% | Val Acc: 77.70%
Epoch [04/20] Train Loss: 0.0215 | Val Loss: 0.7644 | Train Acc: 99.40% | Val Acc: 78.60%
Epoch [05/20] Train Loss: 0.0121 | Val Loss: 0.8842 | Train Acc: 99.73% | Val Acc: 78.20%
Epoch [06/20] Train Loss: 0.0132 | Val Loss: 0.9081 | Train Acc: 99.63% | Val Acc: 78.30%
Epoch [07/20] Train Loss: 0.0066 | Val Loss: 0.8199 | Train Acc: 99.90% | Val Acc: 79.00%
Epoch [08/20] Train Loss: 0.0035 | Val Loss: 0.8092 | Train Acc: 99.97% | Val Acc: 78.20%
Epoch [09/20] Train Loss: 0.0032 | Val Loss: 0.7783 | Train Acc: 99.97% | Val Acc: 80.10%
Epoch [10/20] Train Loss: 0.0013 | Val Loss: 0.7581 | Train Acc: 100.00% | Val Acc: 80.50%
Epoch [11/20] Train Loss: 0.0010 | Val Loss: 0.7460 | Train Acc: 100.00% | Val Acc: 80.50%
Epoch [12/20] Train Loss: 0.0006 | Val Loss: 0.7176 | Train Acc: 100.00% | Val Acc: 81.30%
Epoch [13/20] Train Loss: 0.0008 | Val Loss: 0.7047 | Train Acc: 100.00% | Val Acc: 81.50%
Epoch [14/20] Train Loss: 0.0005 | Val Loss: 0.7091 | Train Acc: 100.00% | Val Acc: 81.70%
Epoch [15/20] Train Loss: 0.0005 | Val Loss: 0.7083 | Train Acc: 100.00% | Val Acc: 81.20%
Epoch [16/20] Train Loss: 0.0006 | Val Loss: 0.7202 | Train Acc: 100.00% | Val Acc: 81.90%
Epoch [17/20] Train Loss: 0.0004 | Val Loss: 0.7152 | Train Acc: 100.00% | Val Acc: 81.90%
Epoch [18/20] Train Loss: 0.0004 | Val Loss: 0.7152 | Train Acc: 100.00% | Val Acc: 82.20%
Epoch [19/20] Train Loss: 0.0004 | Val Loss: 0.7131 | Train Acc: 100.00% | Val Acc: 82.40%
Epoch [20/20] Train Loss: 0.0003 | Val Loss: 0.7060 | Train Acc: 100.00% | Val Acc: 82.30%

Best results for resnet18:
Train Loss: 0.0004 | Val Loss: 0.7131 | Train Acc: 100.00% | Val Acc: 82.40%
Training Time: 122.17 seconds


========== CONFIG ==========
NUM_CLASSES = 10
TRAINING_SAMPLE_PER_CLASS = 100
VALIDATION_SAMPLE_PER_CLASS = 100
BATCH_SIZE = 256
EPOCHS = 20
LR = 0.001
TRAINABLE_LAYERS = 2
EARLY_STOP_PATIENCE = 10
N_AUG = 4
DEVICE = cuda
============================

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 0.5575 | Val Loss: 1.0303 | Train Acc: 81.97% | Val Acc: 74.00%
Epoch [02/20] Train Loss: 0.0700 | Val Loss: 0.9473 | Train Acc: 98.08% | Val Acc: 76.50%
Epoch [03/20] Train Loss: 0.0239 | Val Loss: 0.7925 | Train Acc: 99.55% | Val Acc: 78.50%
Epoch [04/20] Train Loss: 0.0103 | Val Loss: 0.7661 | Train Acc: 99.83% | Val Acc: 79.90%
Epoch [05/20] Train Loss: 0.0077 | Val Loss: 0.8783 | Train Acc: 99.83% | Val Acc: 77.70%
Epoch [06/20] Train Loss: 0.0073 | Val Loss: 0.8420 | Train Acc: 99.85% | Val Acc: 79.40%
Epoch [07/20] Train Loss: 0.0032 | Val Loss: 0.7737 | Train Acc: 99.95% | Val Acc: 80.30%
Epoch [08/20] Train Loss: 0.0021 | Val Loss: 0.7466 | Train Acc: 100.00% | Val Acc: 80.20%
Epoch [09/20] Train Loss: 0.0012 | Val Loss: 0.7240 | Train Acc: 100.00% | Val Acc: 81.30%
Epoch [10/20] Train Loss: 0.0007 | Val Loss: 0.7333 | Train Acc: 100.00% | Val Acc: 80.70%
Epoch [11/20] Train Loss: 0.0006 | Val Loss: 0.7475 | Train Acc: 100.00% | Val Acc: 81.40%
Epoch [12/20] Train Loss: 0.0004 | Val Loss: 0.7163 | Train Acc: 100.00% | Val Acc: 81.80%
Epoch [13/20] Train Loss: 0.0004 | Val Loss: 0.7068 | Train Acc: 100.00% | Val Acc: 82.60%
Epoch [14/20] Train Loss: 0.0003 | Val Loss: 0.7128 | Train Acc: 100.00% | Val Acc: 81.90%
Epoch [15/20] Train Loss: 0.0003 | Val Loss: 0.7149 | Train Acc: 100.00% | Val Acc: 81.90%
Epoch [16/20] Train Loss: 0.0003 | Val Loss: 0.7346 | Train Acc: 100.00% | Val Acc: 82.10%
Epoch [17/20] Train Loss: 0.0003 | Val Loss: 0.7218 | Train Acc: 100.00% | Val Acc: 82.30%
Epoch [18/20] Train Loss: 0.0003 | Val Loss: 0.7346 | Train Acc: 100.00% | Val Acc: 82.00%
Epoch [19/20] Train Loss: 0.0003 | Val Loss: 0.7266 | Train Acc: 100.00% | Val Acc: 82.20%
Epoch [20/20] Train Loss: 0.0002 | Val Loss: 0.7311 | Train Acc: 100.00% | Val Acc: 82.20%

Best results for resnet18:
Train Loss: 0.0004 | Val Loss: 0.7068 | Train Acc: 100.00% | Val Acc: 82.60%
Training Time: 144.81 seconds


========== CONFIG ==========
NUM_CLASSES = 10
TRAINING_SAMPLE_PER_CLASS = 100
VALIDATION_SAMPLE_PER_CLASS = 100
BATCH_SIZE = 256
EPOCHS = 20
LR = 0.001
TRAINABLE_LAYERS = 2
EARLY_STOP_PATIENCE = 10
N_AUG = 5
DEVICE = cuda
============================

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 0.4689 | Val Loss: 1.2141 | Train Acc: 84.92% | Val Acc: 72.50%
Epoch [02/20] Train Loss: 0.0470 | Val Loss: 0.9456 | Train Acc: 98.86% | Val Acc: 75.30%
Epoch [03/20] Train Loss: 0.0149 | Val Loss: 0.7469 | Train Acc: 99.76% | Val Acc: 79.80%
Epoch [04/20] Train Loss: 0.0066 | Val Loss: 0.7828 | Train Acc: 99.88% | Val Acc: 78.50%
Epoch [05/20] Train Loss: 0.0033 | Val Loss: 0.8726 | Train Acc: 99.94% | Val Acc: 78.20%
Epoch [06/20] Train Loss: 0.0016 | Val Loss: 0.7470 | Train Acc: 100.00% | Val Acc: 80.60%
Epoch [07/20] Train Loss: 0.0009 | Val Loss: 0.7269 | Train Acc: 100.00% | Val Acc: 81.00%
Epoch [08/20] Train Loss: 0.0010 | Val Loss: 0.8140 | Train Acc: 99.98% | Val Acc: 79.60%
Epoch [09/20] Train Loss: 0.0013 | Val Loss: 0.8110 | Train Acc: 99.98% | Val Acc: 79.70%
Epoch [10/20] Train Loss: 0.0007 | Val Loss: 0.7776 | Train Acc: 100.00% | Val Acc: 80.50%
Epoch [11/20] Train Loss: 0.0005 | Val Loss: 0.7703 | Train Acc: 100.00% | Val Acc: 80.90%
Epoch [12/20] Train Loss: 0.0003 | Val Loss: 0.7382 | Train Acc: 100.00% | Val Acc: 80.80%
Epoch [13/20] Train Loss: 0.0003 | Val Loss: 0.7256 | Train Acc: 100.00% | Val Acc: 81.80%
Epoch [14/20] Train Loss: 0.0002 | Val Loss: 0.7374 | Train Acc: 100.00% | Val Acc: 81.60%
Epoch [15/20] Train Loss: 0.0002 | Val Loss: 0.7457 | Train Acc: 100.00% | Val Acc: 81.00%
Epoch [16/20] Train Loss: 0.0002 | Val Loss: 0.7578 | Train Acc: 100.00% | Val Acc: 81.00%
Epoch [17/20] Train Loss: 0.0002 | Val Loss: 0.7834 | Train Acc: 100.00% | Val Acc: 81.10%
Epoch [18/20] Train Loss: 0.0002 | Val Loss: 0.7569 | Train Acc: 100.00% | Val Acc: 81.70%
Epoch [19/20] Train Loss: 0.0002 | Val Loss: 0.7576 | Train Acc: 100.00% | Val Acc: 80.90%
Epoch [20/20] Train Loss: 0.0002 | Val Loss: 0.7647 | Train Acc: 100.00% | Val Acc: 81.10%

Best results for resnet18:
Train Loss: 0.0003 | Val Loss: 0.7256 | Train Acc: 100.00% | Val Acc: 81.80%
Training Time: 166.25 seconds

Judging by the results, the optimal value for N_AUG is 3 or 4. Let's use more complex transform to see if it improves validation accuracy. Here's the code and log:

train_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.RandomResizedCrop(224, scale=(0.7, 1.0)),
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(
        brightness=0.2,
        contrast=0.2,
        saturation=0.2,
        hue=0.05
    ),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

========== CONFIG ==========
NUM_CLASSES = 10
TRAINING_SAMPLE_PER_CLASS = 100
VALIDATION_SAMPLE_PER_CLASS = 100
BATCH_SIZE = 256
EPOCHS = 20
LR = 0.001
TRAINABLE_LAYERS = 2
EARLY_STOP_PATIENCE = 10
N_AUG = 5
DEVICE = cuda
============================

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 0.5717 | Val Loss: 1.2045 | Train Acc: 81.44% | Val Acc: 71.10%
Epoch [02/20] Train Loss: 0.1210 | Val Loss: 0.8994 | Train Acc: 96.42% | Val Acc: 75.70%
Epoch [03/20] Train Loss: 0.0603 | Val Loss: 0.9678 | Train Acc: 98.34% | Val Acc: 75.20%
Epoch [04/20] Train Loss: 0.0402 | Val Loss: 0.8390 | Train Acc: 98.90% | Val Acc: 78.10%
Epoch [05/20] Train Loss: 0.0242 | Val Loss: 0.7963 | Train Acc: 99.44% | Val Acc: 78.90%
Epoch [06/20] Train Loss: 0.0218 | Val Loss: 0.9132 | Train Acc: 99.40% | Val Acc: 77.00%
Epoch [07/20] Train Loss: 0.0134 | Val Loss: 0.8210 | Train Acc: 99.68% | Val Acc: 78.60%
Epoch [08/20] Train Loss: 0.0095 | Val Loss: 0.8353 | Train Acc: 99.86% | Val Acc: 78.40%
Epoch [09/20] Train Loss: 0.0094 | Val Loss: 0.9591 | Train Acc: 99.74% | Val Acc: 76.90%
Epoch [10/20] Train Loss: 0.0058 | Val Loss: 0.9056 | Train Acc: 99.84% | Val Acc: 79.90%
Epoch [11/20] Train Loss: 0.0114 | Val Loss: 1.0335 | Train Acc: 99.70% | Val Acc: 76.70%
Epoch [12/20] Train Loss: 0.0146 | Val Loss: 1.0974 | Train Acc: 99.50% | Val Acc: 76.20%
Epoch [13/20] Train Loss: 0.0155 | Val Loss: 1.0826 | Train Acc: 99.60% | Val Acc: 75.90%
Epoch [14/20] Train Loss: 0.0179 | Val Loss: 1.0649 | Train Acc: 99.48% | Val Acc: 73.90%
Epoch [15/20] Train Loss: 0.0233 | Val Loss: 1.0058 | Train Acc: 99.32% | Val Acc: 76.80%
Epoch [16/20] Train Loss: 0.0237 | Val Loss: 0.9900 | Train Acc: 99.20% | Val Acc: 75.40%
Epoch [17/20] Train Loss: 0.0301 | Val Loss: 1.0669 | Train Acc: 98.98% | Val Acc: 76.20%
Epoch [18/20] Train Loss: 0.0268 | Val Loss: 1.5781 | Train Acc: 99.14% | Val Acc: 72.00%
Epoch [19/20] Train Loss: 0.0173 | Val Loss: 1.1010 | Train Acc: 99.46% | Val Acc: 74.80%
Epoch [20/20] Train Loss: 0.0140 | Val Loss: 1.0322 | Train Acc: 99.60% | Val Acc: 76.00%
Early stopping triggered at epoch 20

Best results for resnet18:
Train Loss: 0.0058 | Val Loss: 0.9056 | Train Acc: 99.84% | Val Acc: 79.90%
Training Time: 379.45 seconds

Here's another more complex transform function. Here's the code and log:

train_transform = transforms.Compose([
    transforms.Resize(224),
    transforms.RandomResizedCrop(224, scale=(0.6, 1.0)),  # zoom in/out
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.1),
    transforms.RandomGrayscale(p=0.1),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])


========== CONFIG ==========
NUM_CLASSES = 10
TRAINING_SAMPLE_PER_CLASS = 100
VALIDATION_SAMPLE_PER_CLASS = 100
BATCH_SIZE = 256
EPOCHS = 20
LR = 0.001
TRAINABLE_LAYERS = 2
EARLY_STOP_PATIENCE = 10
N_AUG = 3
DEVICE = cuda
============================

==============================
Training model: resnet18
==============================
Epoch [01/20] Train Loss: 1.0339 | Val Loss: 2.3851 | Train Acc: 65.23% | Val Acc: 60.50%
Epoch [02/20] Train Loss: 0.4568 | Val Loss: 0.9064 | Train Acc: 84.13% | Val Acc: 73.70%
Epoch [03/20] Train Loss: 0.3050 | Val Loss: 0.9382 | Train Acc: 90.27% | Val Acc: 74.40%
Epoch [04/20] Train Loss: 0.2116 | Val Loss: 0.7938 | Train Acc: 92.57% | Val Acc: 75.20%
Epoch [05/20] Train Loss: 0.1569 | Val Loss: 0.8974 | Train Acc: 95.10% | Val Acc: 74.20%
Epoch [06/20] Train Loss: 0.1196 | Val Loss: 0.7643 | Train Acc: 96.20% | Val Acc: 77.60%
Epoch [07/20] Train Loss: 0.1148 | Val Loss: 0.7781 | Train Acc: 96.23% | Val Acc: 78.60%
Epoch [08/20] Train Loss: 0.0841 | Val Loss: 0.7664 | Train Acc: 97.23% | Val Acc: 79.80%
Epoch [09/20] Train Loss: 0.0722 | Val Loss: 0.8293 | Train Acc: 97.83% | Val Acc: 76.70%
Epoch [10/20] Train Loss: 0.0686 | Val Loss: 0.8533 | Train Acc: 97.77% | Val Acc: 79.10%
Epoch [11/20] Train Loss: 0.0717 | Val Loss: 0.8412 | Train Acc: 97.93% | Val Acc: 78.60%
Epoch [12/20] Train Loss: 0.0684 | Val Loss: 1.0308 | Train Acc: 98.00% | Val Acc: 74.70%
Epoch [13/20] Train Loss: 0.0532 | Val Loss: 0.8491 | Train Acc: 98.27% | Val Acc: 78.00%
Epoch [14/20] Train Loss: 0.0564 | Val Loss: 0.8526 | Train Acc: 98.37% | Val Acc: 76.70%
Epoch [15/20] Train Loss: 0.0539 | Val Loss: 0.8671 | Train Acc: 98.47% | Val Acc: 77.70%
Epoch [16/20] Train Loss: 0.0532 | Val Loss: 0.9870 | Train Acc: 98.23% | Val Acc: 77.40%
Epoch [17/20] Train Loss: 0.0577 | Val Loss: 0.9720 | Train Acc: 98.50% | Val Acc: 77.90%
Epoch [18/20] Train Loss: 0.0682 | Val Loss: 1.1606 | Train Acc: 97.90% | Val Acc: 73.90%
Early stopping triggered at epoch 18

Best results for resnet18:
Train Loss: 0.0841 | Val Loss: 0.7664 | Train Acc: 97.23% | Val Acc: 79.80%
Training Time: 254.64 seconds

Any comments? Feel free to participate below in the Facebook comment section.