EfficientNetV2-L
118M Parameters
Mixed FP16 · A100
60.62% Val Acc
AI Model
EfficientNetV2-L deep learning classifier · Two-stage transfer learning from ImageNet · 8 phytoplankton species · 22 training epochs · confusion matrix · per-class F1 scores. Trained on Google Colab A100 GPU with 12,315 FlowCAM library images.
04
Model Architecture
EfficientNetV2-L · 384×384 input · 118M params · Mixed FP16 · Two-stage transfer learning
Forward Pass — hover each block
Input
384×384×3
→
Augment
Flip·Rot·Zoom
→
Preprocess
ENv2 normalize
→
EfficientNetV2-L
117.7M · frozen S1
→
GAP
1280-dim
→
BN + Drop 0.4
1280-dim
→
Dense Swish
512-dim
→
Drop 0.3
512-dim
→
Logits fp32
8 classes out
Stage 1: Head only · LR 2e-4 cosine · 12 epochs · backbone frozen
Stage 2: Top 5% unfrozen · LR 5e-6 constant · 10 epochs · BN frozen
Loss: SparseCatCE from_logits · AdamW wd 1e-4
Logit-adjusted CE with sqrt-inverse class priors
Batch 32 · Steps 262/epoch · Mixed FP16 · A100
05
Training History
Stage 1 (ep 1–12) + Stage 2 fine-tune (ep 13–22) · 22 total epochs · data from metrics.json
Accuracy — Train vs Validation
Loss — Train vs Validation
Stage 1 — Head Training
Epochs1–12
Best val_acc59.86%
LR scheduleCosine 2e-4 → 2e-5
Trainable params662K (head only)
Loss at ep120.964
Stage 2 — Gentle Fine-Tune
Epochs13–22
Best val_acc60.62%
LRConstant 5e-6
Unfrozen layersTop 5% backbone
BN layersAll frozen
Best Val Accuracy
Epoch 19 · Stage 2
06
Model Evaluation
Confusion matrix · per-class F1 · precision · recall · 2,095 validation samples
Confusion Matrix — hover cells for exact counts
F1 Score Radar
Per-Class Metrics — Precision · Recall · F1 · Support
| Class | Precision | Recall | F1 Score | Support | F1 Bar | Status |
|---|
07
Model Confidence Analysis
Softmax probability distribution across classes on discrete cast inference
Confidence by Class (High-Conf Only)
Discrete vs Library Performance
V9 was trained on clean library images. Discrete cast particles are raw ocean samples — more challenging.
Library val_acc60.62%
Discrete high-conf rate13.8%
Dominant class (discrete)Nanophytoplankton (294)
Rarest class (discrete)Synechococcus (1)
Improvement pathRetrain on Discrete pseudo-labels
Top-3 Accuracy Insight
Top-3 accuracy during library training reached 98%+ — the correct class is almost always in the top 3 predictions.
Val Top-1 acc60.62%
Val Top-3 acc~98.2%
Biggest confusion pairNanophyto ↔ Picocyano
Best performingMineral particles (F1 0.96)
Most confusedCrocosphaera (F1 0.08)
Phytoplankton Classification · Goes Lab (LDEO) · Columbia University · Model v9 · EfficientNetV2-L · TF 2.19
Session Active