Transfer Learning with EfficientNetV2-S

• Transfer Learning with EfficientNetV2-S is a technique that employs a pre-trained CNN backbone and modular head replacement to adapt to domain-specific tasks.
• It utilizes stage-wise training by initially freezing base layers and then finetuning selected network blocks with targeted data augmentation.
• Benchmark studies reveal high accuracy and eco-efficiency, achieving performance gains with reduced training time and carbon emissions.

Transfer learning with EfficientNetV2-S represents an overview of efficient CNN feature extraction, modular architectural adaptation, and targeted finetuning across diverse vision domains. The approach leverages pre-trained EfficientNetV2-S backbones—originally optimized on large-scale datasets such as ImageNet—and adapts these representations to domain-specific downstream tasks by replacing and re-training terminal classifiers, optionally modifying higher network stages, and deploying data augmentation and regularization strategies crafted for each domain’s data regime and class distribution.

1. EfficientNetV2-S Architecture and Pretraining

EfficientNetV2-S is the “small” member of the EfficientNetV2 family, designed for scaled image recognition with enhanced training efficiency relative to its predecessors. Standard configuration involves approximately 21 million parameters and input resolution of 384×384 px for the S variant. Core architectural elements include a sequence of fused-MBConv and MBConv blocks integrated with squeeze-and-excitation, progressive image-size scheduling, and dynamic regularization—collectively supporting rapid, stable optimization (Fan et al., 2021). Pretraining on ImageNet-1k establishes generic feature extractors suitable for transfer via domain adaptation (Kunwar, 2024, Fan et al., 2021, Farabi et al., 3 Oct 2025).

2. Transfer-Learning Protocols and Customization

Across applications, the dominant transfer-learning strategies include backbone freezing (feature extraction), partial or full unfreezing (finetuning), and head replacement.

• In waste-classification (Garbage Dataset), EfficientNetV2-S was initialized with pretrained weights, its native head removed, and a custom head (GlobalAveragePooling2D → Dense(256) → Dropout (p ≈ 0.3) → Dense(10, softmax)) inserted. Training was staged: base layers frozen for initial epochs (feature extraction), followed by full-network finetuning (Kunwar, 2024). This hybrid protocol mitigates overfitting for limited-size domains while enabling adaptation to task-specific features.
• For breast cancer histopathology, the EfficientNetV2-S convolutional body (stages 1–6) was frozen. Only the final network stages and new classification head (1×1 convolution + BatchNorm + ACON-C activation → GlobalAveragePooling → Dense(softmax)) were trained (Fan et al., 2021). Explicit architectural modification included replacing SiLU activation with ACON-C, a two-parameter, sigmoid-gated “maxout” variant:

$$f_{\mathrm{ACON\text{-}C}}(x) = (p_1 - p_2)x\,\sigma[\beta(p_1 - p_2)x] + p_2 x$$

where $p_1, p_2$ are trainable, channel-wise parameters and $\sigma$ is sigmoid.

• In facial emotion recognition (FER2013), only the lower layers (all fused-MBConvs up through stage 5) were frozen; the upper blocks and new classification head were trained. The classification head consisted of GlobalAveragePooling, Dropout (p not stated), and Dense(7, softmax). Class-weighted cross-entropy loss handled data imbalance, with weights inversely proportional to class frequencies (Farabi et al., 3 Oct 2025).

3. Data Preparation and Augmentation

Domain-specific preprocessing and augmentation are integral to transfer learning with EfficientNetV2-S:

• For garbage classification, the original ~24k multi-class set was balanced by undersampling all classes to ≤1,000 images each; images were resized to the dataset’s mean dimensions and subjected to random horizontal/vertical flips, rotations, random crops, and shifts during training (Kunwar, 2024).
• For breast histopathology, all 7909 RGB patches were resized to 384×384, ImageNet-normalized, with no further color/stain augmentation (Fan et al., 2021).
• FER2013 images were upsampled from 48×48 px grayscale to 224×224 px RGB via channel replication, and normalized with ImageNet statistics. Augmentations included RandomResizedCrop, RandomHorizontalFlip, RandomRotation (±15°), and ColorJitter (±20% for brightness/contrast/saturation) (Farabi et al., 3 Oct 2025).

Augmentation pipelines target both regularization (combatting overfitting due to small or imbalanced sets) and the simulation of real-world intra-class variability.

4. Optimization, Regularization, and Hyperparameter Tuning

All cited studies used Adam optimizer (β₁=0.9, β₂=0.999), with learning rates adapted via either manual grid search or automated hyperparameter tuning (e.g., Optuna selected lr ≈1×10⁻⁴ for waste classification). Weight decay ($\ell_2$ ≈1×10⁻⁵), dropout (p ≈0.3, where stated), and gradient clipping (norm = 1.0) were employed for regularization (Kunwar, 2024). Early stopping on validation loss was standard, with batch sizes ranging from 16 (histopathology) to 64 (FER2013). Cosine-annealing learning schedules were also adopted (Farabi et al., 3 Oct 2025).

For imbalance-aware optimization, class-weighted loss functions:

$$L(\theta) = - \frac{1}{N} \sum_{i=1}^{N} w_{y_i} \log p_{y_i}(x_i; \theta)$$

were utilized, with $w_c$ set inversely to class cardinality $n_c$ (Farabi et al., 3 Oct 2025). This targets improved minority-class sensitivity—crucial for FER and biomedical domains.

5. Evaluation Metrics and Comparative Performance

Evaluation protocols employ standard supervised metrics: overall and per-class Accuracy, Recall, Precision, F1-score, Intersection over Union (IoU), and macro-averaged F1 where class imbalance is present.

Application	Dataset	Acc.	Macro-F1	IoU	Notable Baseline(s)	Carbon Cost (Train)
Garbage classification	Garbage (10-class)	96.41%	0.95	0.957	ResNet50: 94.63%, MobileNet: 68.89%	≈0.04 kg CO₂
Breast cancer diagnosis	BreaKHis (8-class)	84.71%	Not stated	—	EfficientNetV2-S (unmodified): up to -6.4%	Not stated
Facial emotion recognition	FER2013 (7-class)	62.8%	0.590	—	Macro-F1 competitive w/ “conventional CNN baselines”	Not stated

• EfficientNetV2-S set a performance baseline in garbage classification: 96.41% accuracy after tuning; IoU ≈0.96; with training time of 6016.56 s (1.7 h on 2×T4 GPUs) and carbon emissions ≈0.04 kg CO₂ (Kunwar, 2024).
• In histopathology, transfer learning with EfficientNetV2-SA (ACON-C activated head) delivered 84.71% accuracy (average over four magnifications), outperforming the standard EfficientNetV2-S at multiple scales (e.g., +6.4% at 40×). Only ~2 million parameters in higher layers were fine-tuned, reducing overfitting risk (Fan et al., 2021).
• On FER2013, InsideOut (EfficientNetV2-S backbone) attained 62.8% accuracy and 0.590 macro-F1. Highest per-class F1 scores appeared in Happy (0.832) and Surprise (0.755); minority classes (Disgust, Fear) showed improved recall due to class weighting (Farabi et al., 3 Oct 2025).

6. Environmental and Computational Considerations

Evaluation of carbon emissions is a distinctive feature of recent EfficientNetV2-S transfer learning studies. For example, (Kunwar, 2024) logged power draw and estimated emissions for training, validation, and deployment using CodeCarbon. EfficientNetV2-S demonstrated superior eco-efficiency, requiring ~75% of the training time and only ~25% of the carbon footprint compared to EfficientNetV2-M, for nearly the same top-line accuracy. Emissions are quantified via:

$$\text{Emission (kg CO}_2) = \text{EnergyConsumed (kWh)} \times \text{CarbonIntensity (kg CO}_2/\text{kWh)}$$

This frames model selection not only in terms of predictive performance but also sustainability, with EfficientNetV2-S consistently representing the optimal tradeoff among studied backbones.

7. Domain-Specific Adaptation and Practical Guidelines

Empirical findings recommend the following best practices for transfer learning with EfficientNetV2-S:

• Utilize ImageNet-pretrained EfficientNetV2-S as the backbone; replace the head with a domain-appropriate architecture, typically (GAP → Dense → Dropout → Output) (Kunwar, 2024, Farabi et al., 3 Oct 2025).
• Apply staged training: freeze the backbone for initial head training, then progressively unfreeze higher blocks for finetuning.
• Deploy extensive augmentation during training to accommodate domain-specific image distributions and mitigate overfitting, especially when sample sizes are limited (e.g., random flips, rotations, crops).
• Leverage automated hyperparameter optimization (e.g., Optuna) for learning rate, dropout, and weight decay (Kunwar, 2024).
• For imbalanced data, use class-weighted categorical cross-entropy or other loss functions explicitly accounting for class distribution (Farabi et al., 3 Oct 2025).
• Track and report carbon emissions for all stages—data preparation, training, inference—when assessing and selecting models.

In summary, transfer learning with EfficientNetV2-S synthesizes efficient large-scale feature transfer, data and regularization best practices, and sustainability-aware benchmarking. It achieves robust, state-of-the-art accuracy across multiple domains, with environmental and resource considerations increasingly foregrounded in its evaluation (Kunwar, 2024, Fan et al., 2021, Farabi et al., 3 Oct 2025).