mobile-leafdoc

We are working to reproduce the Frontiers in Plant Science (Dec 2023) model that fine-tunes MobileNetV3-Small on PlantVillage and then performs post-training quantization to an ONNX network. The paper (https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2023.1308528/full) reports ~99.5% test accuracy on PlantVillage and a parameter reduction from ~1.5M → 0.93M with no accuracy loss after quantization, highlighting suitability for mobile/edge deployment. However, lab-style PlantVillage (https://github.com/spMohanty/PlantVillage-Dataset/tree/master) images may not reflect real field conditions, prior work cited in the paper shows that performance can degrade on in-the-wild photos with shadows, clutter, and misalignment.

We therefore target generalization and reliability under real-world shift while preserving edge constraints (size/latency) using the PlantDoc dataset (https://github.com/pratikkayal/PlantDoc-Dataset)

Experimental design

PlantVillage: ~54k leaf images, 38 classes, studio-like (plain background). Great for fast training but not realistic. PlantDoc: ~2.6k “in-the-wild” images (phones/internet), 13 species, ~17 classes.

✅ baseline:

PV only → MobilenetV3-Small, paper settings. Measure ID metrics.

Major Gaps:

1. No preprocessing done on the PlantVillage dataset
2. Trained on only the PlantVillage dataset

Runs

Frontiers2023/Run 1

epochs 20 ran on plantvillage dataset, no normalization exact settings as given on the paper

Frontiers2023/Run 2

epochs 200 ran on plantvillage dataset, no normalization exact settings as given on the paper

• got acc ~ 99.6%, slightly higher than the paper
• this is the final checkpoint

• ✅ Done

✅ robustness block:

Experimenting on 3 levels

• ✅ PV only → test on PD (zero-shot)

  train on PlantVillage only, test on PlantDoc.

• ✅ PV → fine-tune on PD

  train on PlantVillage, then fine-tune on PlantDoc.

  Using 'Frontiers2023/Run 2/mobilenetv3small_best' model and fine-tune it with PlantDoc dataset. Fine-tune on PlantDoc dataset, with normalization and data augmentation. (run1 and run2 are not much different) ( https://colab.research.google.com/drive/1bKeXgYZbvrBcp2T4WxQ0mcyTFCStOT7p?usp=sharing )

  Runs

  finetune/Run 1

  epochs 20

  • got best acc ~ 50%

  finetune/Run 2

  epochs 50

  • got best acc ~ 53%
  • evaluate on PlantDoc test data - Baseline (Zero-Shot): 2.54% - Fine-Tuned: 50.42%

• ✅ PV+PD mixed (balanced sampler) with color constancy + RandAugment.

  1. train on PlantVillage+PlantDoc (only overlapping classes)

  mixed:

  epochs 30

  • PV test acc=0.9964 F1=0.9954
  • PD test acc=0.5297 F1=0.5182

  2. using a dataset-balanced sampler and robust augmentations.

  Runs use RandAugment + ColorConstancy (with balance classes)

  mixed_pv_pd/Run 1

  epochs 10

  • got best acc ~ 98%

  mixed_pv_pd/Run 2

  epochs 15

  • got best acc ~ 99.4%
  • evaluate on PlantDoc test data - Mixed: 56.78%

  3. Using weights for the PD data samples (to balance the datasize imbalance)

  mixed_pdweighted:

  epochs 30

  • use RandAugment + ColorConstancy (with balance classes)
  • PV test acc=0.9974 F1=0.9965
  • PD test acc=0.5847 F1=0.5818 (topk_acc=0.86)

  4. try using other augmentation, Asymmetric (Heavy noise for PV, Light for PD), + FDA (Fourier Domain Adaptation) ( https://drive.google.com/file/d/1fZ8GDheW7PDFLvOjPkTO_UOFbwCAJvCC/view?usp=sharing )

  Runs use Asymmetric + FDA (with balance classes + domains (upweights PD 5x))

  mixed_with_fda/Run 1

  epochs 20

  • PV test: 98.6%
  • PD test: 60.17%, (top k): 84.75%

  4. try using other augmentation, Asymmetric (Heavy noise for PV, Light for PD), + FDA (Fourier Domain Adaptation) with ImageNetV3_large ( https://colab.research.google.com/drive/1o_6SrJaObWygqOW__M8XthyJ7GKcUWOd?usp=sharing )

  Runs use ImageNetV3_large model

  mixed_with_imageNewV3_large

  epochs 10

  • PV test: 99.01%
  • PD test: 61.44%

❌ check compression block:

• first perform PTQ-INT8
• Compare FP32 vs INT8 on ID + OOD and report edge metrics.