vicpantoja2/aedes-egg-weights

Experiments

Real-ESRGAN

Model	Total Iters	Best Iteration	Total Duration	Status / Result
Real-ESRNet x4	1,000,000	824,800	5d 20h 21m	✅ Stable (End of learning)
Real-ESRGAN x4	400,000	260,900	~3.5 days	🔄 GAN Dynamics (Stable)
SR3 (Rounds 1-4)	1,185,000	1,085,000	~3d 19h	🔴 Saturation / Overfitting
SR3 (Round 5)	1,200,000	925,000	~2d	🏆 Best PSNR (17.81 dB)

Model / Phase	Average Time per Epoch	Time per Iteration
Real-ESRNet x4	~13.7 min	0.500s
Real-ESRGAN x4	~23 min	0.816s (avg)
SR3 (Rounds 1-4)	Variable	~0.86s
SR3 (Round 5)	-	~0.46s

---

1. Real-ESRNet x4

• Learning: Exemplary behavior. The minimum loss occurred late (824k), indicating that the model continued refining details for a long time before saturating.
• Verdict: Checkpoint 800k+ is the best for mathematical fidelity (PSNR/SSIM).

---

2. Real-ESRGAN x4

• Convergence: Lowest pixel loss at 260k. After this point, pixel loss increased while Perceptual/GAN losses fluctuated, focusing on textures.
• Verdict: After 260k, the Discriminator prioritizes visual realism over pure metrics (L1).

---

SR3 (Super-Resolution via Iterative Refinement)**

Training Rounds Summary

Round	Interval (Iterations)	Epochs	Total Duration
Round 1	0 — 1,185,000	0 — 968	91h 07m 22s
Round 2	0 — 1,200,000	0 — 83	47h 57m 16s

Training 1 was performed with 411 images until iteration 1,085,000. It then proceeded with 4,932 images for 23h 49m 14s. Training 2 was performed with 4,932 images images

Maximum Performance per Round:

• Best Checkpoint (Round 1): 17.42 dB (Iter 1,085,000, Epoch 916).
• Best Checkpoint (Round 2): 17.81 dB (Iter 925,000, Epoch 57).

📌 Training Configuration

• Data: LR/SR image pairs at 64x64 and 128x128 resolutions.

---

Rounds Detailing

Round 1: Consolidated Training (Incremental)

This round represents the initial development, including dataset transition and hardware optimizations.

• Total Duration: 91h 07m 22s.
• Metrics:
  • Peak (Best PSNR): 17.42 dB.
  • Stability: Showed high initial variability (±2.14 dB), stabilizing after the introduction of Gradient Accumulation.
• Technical Observations:
  • Hardware: Use of Gradient Accumulation (Effective Batch 2) to overcome the 9.7 GB VRAM limit (NVIDIA A100).
  • Overfitting: Identified after iteration 1,085,000, where performance became unstable.
  • Augmentation (End of Round): Expansion to 4,932 images in this phase resulted in performance drop (-0.74 dB), indicating that the model "specialized" in the smaller dataset rejected highly transformed data.

Round 2: Full Dataset Training

Training restarted with the dataset of 4,932 images from the base checkpoint.

• Total Duration: 47h 57m 16s.
• 🏆 Maximum Result: 17.81 dB (Reached at 925,000 iterations).
• Advantage: Converged 14.7% faster than incremental phases.
• Risk: Abrupt collapse behavior after reaching the peak, requiring manual and early selection of checkpoints (early stopping).

SR3 Verdict: Training initiated with the full dataset was superior in absolute metrics, although Round 1 provided fundamental knowledge about memory limitations and the model's sensitivity to augmentation.