plantdoc-predictor 1.1.1

A Python library for predicting plant diseases from leaf images across multiple datasets (PlantVillage, Rice, and more) using pre-trained or custom deep learning models.

🌿 PlantDoc-Predictor

---

A Python library for predicting plant diseases from leaf images using pre-trained or custom deep learning models — now with multi-dataset support (PlantVillage + Rice, and more on the way), selectable through a single dataset argument.

---

🚀 Features

• ✅ Pretrained models included — Ready-to-use architectures like InceptionV3, ResNet50, MobileNetV2, ViT, Swin, DeiT3 and more.
• 🌍 Multi-dataset support (1.1.0+) — Detect diseases across multiple datasets (PlantVillage 38-class + Rice 6-class, more coming) via a single dataset argument. Pick a dataset and the best model is chosen for you.
• 🧠 Unified API — One interface for both built-in and custom .h5 / .pth models from Keras and PyTorch.
• 🧩 Custom model support — Load your own model and label mapping JSON.
• 🌱 Extensible — Add a whole new dataset by dropping in one <dataset>_registry.json — no code changes.
• 🧰 Visualization support — Displays prediction confidence and leaf images.
• ⚙️ Cross-platform — Works seamlessly on Windows, macOS, and Linux.

🔬 Advanced Features (NEW 🚀) Added in release 1.0.0 and above.

• 🧠 Full Model Access — Retrieve the complete loaded model for fine-tuning and experimentation.
• ⚙️ Weights Extraction — Access model weights for analysis, comparison, and research.
• 🧩 Layer Introspection — List all layers and inspect architecture programmatically.
• 🏆 Top-K Predictions — Get ranked predictions, not just the top-1 label (1.0.2+).
• 🖼️ PIL Image Input — Pass PIL.Image objects directly — ideal for Streamlit / FastAPI (1.0.2+).
• 📦 Batch Prediction — BatchPredictor over folders/lists with CSV & JSON export (1.0.2+).
• 💻 Command-Line Interface — plantdoc models / plantdoc predict (1.0.2+).
• 🛡️ Guarded Prediction — GuardedPredictor rejects non-leaf images with a CLIP-based guard (1.0.3+).
• 🔥 Grad-CAM Explainability — ExplainablePredictor overlays a heatmap showing which leaf regions drove the prediction (1.0.3+).
• 🌍 Multi-Dataset Support — Select a dataset with Predictor(dataset="rice"); each dataset lives in its own auto-merged registry file. Ships 7 PyTorch rice models (1.1.0+).

---

🔥 Feature Extraction (Research-Grade 🚀)

• 📊 Extract intermediate representations from any layer
• 🧪 Enables feature-space SMOTE, clustering, and embedding analysis
• 🧠 Works across all models in the model zoo
• 🔍 Supports custom layer selection

👉 This transforms PlantDoc-Predictor into a feature extraction + research framework, not just an inference tool.

---

🧠 Why Use PlantDoc-Predictor

PlantDoc-Predictor was created to reduce repetitive work in plant-disease research.
Researchers and agritech developers often train from scratch — this tool eliminates that friction by providing:

• Pretrained baselines for benchmarking new models.
• Standardized label sets and metadata for reproducibility.
• Plug-and-play inference for agricultural image datasets.
• A unified interface for rapid experimentation and deployment.

Whether you’re a researcher, startup, or precision-agriculture developer, PlantDoc-Predictor simplifies your workflow and lets you focus on innovation — not setup.

---

🧬 Supported Models

PlantDoc-Predictor includes a diverse model zoo of pretrained CNN and transformer architectures across multiple datasets. Models are automatically downloaded from the remote registry when first used, making them ready for plug-and-play inference.

Use any model by name via Predictor(model_name=...), or pick a whole dataset with Predictor(dataset=...) to get its default model.

🌱 PlantVillage — 38-class dataset

The original model zoo, fine-tuned on the PlantVillage 38-class dataset (0.2.1+). Select with Predictor(dataset="plantvillage") (also the library default). Default model: rec_add_attention_v1.

Model Name	Input Size	Accuracy	Framework	Description
Recursive Additive Attention v1	224×224	99.7%	Keras	Custom CNN with Recursive Additive Attention (research model)
ConvNeXt Base v1	224×224	99.1%	Keras	Modern ConvNet architecture inspired by transformers
ConvNeXt Small v1	224×224	99.5%	Keras	Lightweight ConvNeXt variant with high accuracy
ConvNeXt Tiny v1	224×224	99.23%	Keras	Efficient ConvNeXt model for faster inference
Swin Base Patch4 Window7	224×224	99.1%	PyTorch	Hierarchical Vision Transformer with shifted windows
Swin Tiny Patch4 Window7	224×224	99.1%	PyTorch	Lightweight Swin Transformer for efficient inference
ViT Base 16 v1	224×224	99.1%	PyTorch	Vision Transformer base model
ViT Large 16 v1	224×224	99.1%	PyTorch	Larger ViT model with higher capacity
ViT Small 16 v1	224×224	99.1%	PyTorch	Smaller ViT model for faster inference
ViT Tiny 16 v1	224×224	99.1%	PyTorch	Lightweight Vision Transformer
RegNetX-160 v1 (NEW)	224×224	99.1%	PyTorch	RegNetX-160 fine-tuned on PlantVillage dataset
RegNetY-160 v1 (NEW)	224×224	99.1%	PyTorch	RegNetY-160 fine-tuned on PlantVillage dataset
RegNetY-320 v1 (NEW)	224×224	99.1%	PyTorch	RegNetY-320 fine-tuned on PlantVillage dataset
DenseNet169 v1	224×224	99.68%	Keras	Best performing DenseNet variant
VGG19 v1	224×224	98.98%	Keras	Deep VGG architecture with strong performance
DenseNet121 v1	224×224	98.68%	Keras	Dense connectivity for efficient feature reuse
InceptionV3 v1	299×299	98.2%	Keras	Inception architecture with multi-scale feature extraction
ResNet50 v1	224×224	97.8%	Keras	Residual network for deep feature learning
EfficientNetB50 v1	224×224	97.8%	Keras	Efficient scaling of CNN architecture
DenseNet210 v1	224×224	97.0%	Keras	Very deep DenseNet variant
MobileNetV2 v1	224×224	96.8%	Keras	Mobile-friendly lightweight architecture
VGG16 v1	224×224	96.8%	Keras	Classic deep CNN architecture
AlexNet v1	224×224	92.8%	Keras	Early CNN baseline model

🌾 Rice Disease Detection — 6-class dataset (1.1.0+)

7 PyTorch (timm) models fine-tuned on a 6-class rice-leaf dataset — Bacterial Leaf Blight, Brown Spot, Healthy Rice Leaf, Leaf Blast, Leaf scald, Sheath Blight. Select with Predictor(dataset="rice"). Default model: rice_swin_base_v1. Labels come straight from each checkpoint's embedded class_to_idx.

Model Name	model_name	Input Size	Accuracy	Framework	Description
Swin-Base (default)	rice_swin_base_v1	224×224	—	PyTorch	Swin-Base (patch4, window7) fine-tuned on Rice 6-class
Swin-Tiny	rice_swin_tiny_v1	224×224	—	PyTorch	Lightweight Swin Transformer
DeiT3-Base	rice_deit3_base_v1	224×224	96.74%	PyTorch	Data-efficient image transformer (patch16)
ViT-Small	rice_vit_small_v1	224×224	—	PyTorch	Vision Transformer small (patch16)
DenseNet-121	rice_densenet121_v1	224×224	94.78%	PyTorch	Dense connectivity CNN
ResNet-101	rice_resnet101_v1	224×224	—	PyTorch	Deep residual network
RegNetY-160	rice_regnety_160_v1	224×224	—	PyTorch	RegNetY design-space CNN

See the Multi-Dataset Support section above for a runnable example that compares all 7 models on one image. Accuracies shown are the measured validation figures; — means not benchmarked yet.

---

🧪 Research Models (Published Work)

PlantDoc-Predictor also includes models derived from peer-reviewed research papers, enabling reproducibility and direct comparison with published work.

These models represent novel architectures and contributions to the field of plant disease classification.

Model	Paper	Authors	Accuracy	Description
Recursive Additive Attention v1	IEEE Paper	Subham Divakar, Rojalina Priyadarshini	99.70%	Custom CNN with Recursive Additive Attention mechanism for enhanced feature interaction and classification performance

---

🚀 Upcoming Models

PlantDoc-Predictor is actively expanding its Model Zoo.
Vision Transformers (ViT), ConvNeXt, Swin Transformers, and RegNet are already included (see the Model Zoo above). Future releases aim to add:

• EfficientNetV2 family
• Hybrid CNN–Transformer models
• Multimodal models for plant disease detection (presented at several conferences)

These additions will let researchers and developers experiment with state-of-the-art deep learning architectures for plant disease classification.

Stay tuned for future releases as the PlantDoc Model Zoo continues to grow. 🌿

---

🌍 Multi-Dataset Support

PlantDoc-Predictor now ships models across multiple datasets, selectable through a single dataset argument. When you don't specify anything, it defaults to PlantVillage (backward-compatible).

Dataset	Classes	Images	Crops	Notes
PlantVillage	38	~54K	14 crops	✅ Available — lab-condition images (23 models)
Rice Leaf Disease	6	~3.8K	Rice	✅ Available — 7 PyTorch models (bacterial leaf blight, brown spot, healthy, leaf blast, leaf scald, sheath blight)
PlantDoc	27	~2.6K	13 crops	🔜 Real-world field images, more challenging
Cassava Leaf Disease	5	~21K	Cassava	🔜 Kaggle competition dataset, field images
Citrus Disease	5	~759	Citrus	🔜 Blackspot, canker, greening, healthy
Wheat Disease	3+	varies	Wheat	🔜 Rust, septoria, yellow rust
Corn / Maize Disease	4	~3K	Maize	🔜 Blight, common rust, gray leaf spot

Discover everything available at runtime:

from plantdoc_predictor import list_datasets
list_datasets()                        # datasets + their default models

plantdoc datasets                      # same, from the CLI

📌 Each dataset lives in its own <dataset>_registry.json under plantdoc_predictor/models/, auto-discovered and merged at load time. Adding a new dataset is a pure data change — no code edits.

---

🌱 PlantVillage (default dataset)

The original 38-class lab-condition dataset across 14 crops, with 23 models spanning CNN and transformer families (InceptionV3, ResNet50, EfficientNet, MobileNetV2, DenseNet121/169/210, VGG16/19, AlexNet, ConvNeXt, ViT, Swin, RegNet, and the custom Recursive Additive Attention model). Default model: rec_add_attention_v1.

from plantdoc_predictor import Predictor

# Nothing specified -> PlantVillage default (fully backward-compatible)
Predictor().predict("leaf.jpg")

# Equivalent, explicit
Predictor(dataset="plantvillage").predict("leaf.jpg")

# Or pick any PlantVillage model by name
Predictor(model_name="convnext_small_v1").predict("leaf.jpg")

plantdoc models --dataset plantvillage      # list the 23 PlantVillage models
plantdoc predict leaf.jpg                    # uses the PlantVillage default

---

🌾 Rice Disease Detection (1.1.0+)

A 6-class rice-leaf dataset (Bacterial Leaf Blight, Brown Spot, Healthy Rice Leaf, Leaf Blast, Leaf scald, Sheath Blight) with 7 PyTorch (timm) models. Default model: rice_swin_base_v1. Labels are read directly from each checkpoint's embedded class_to_idx, and weights are downloaded from HuggingFace on first use (cached under ~/.plantdoc/).

Registry name	Architecture	Val Acc
rice_swin_base_v1 (default)	Swin-Base (patch4, window7, 224)	—
rice_swin_tiny_v1	Swin-Tiny (patch4, window7, 224)	—
rice_deit3_base_v1	DeiT3-Base (patch16, 224)	96.74%
rice_vit_small_v1	ViT-Small (patch16, 224)	—
rice_densenet121_v1	DenseNet-121	94.78%
rice_resnet101_v1	ResNet-101	—
rice_regnety_160_v1	RegNetY-160	—

from plantdoc_predictor import Predictor

# Auto-picks the rice default model (Swin-Base)
Predictor(dataset="rice").predict("rice_leaf.jpg")

# Pick a specific rice model
Predictor(model_name="rice_deit3_base_v1").predict("rice_leaf.jpg")

plantdoc models --dataset rice
plantdoc predict rice_leaf.jpg --dataset rice

Compare all 7 rice models on one image:

from plantdoc_predictor import Predictor, list_available_models

img = "rice_leaf.jpg"

for name in list_available_models(dataset="rice"):
    r = Predictor(model_name=name).predict(img)
    print(f"{name:22} -> {r['label']:22} {r['confidence']*100:5.1f}%")

Example output on a Brown Spot leaf — all models agree:

rice_swin_base_v1      -> Brown Spot              99.8%
rice_swin_tiny_v1      -> Brown Spot             100.0%
rice_deit3_base_v1     -> Brown Spot             100.0%
rice_vit_small_v1      -> Brown Spot             100.0%
rice_densenet121_v1    -> Brown Spot             100.0%
rice_resnet101_v1      -> Brown Spot             100.0%
rice_regnety_160_v1    -> Brown Spot             100.0%

💡 list_available_models(dataset="rice") also prints a table as a side effect. To stay quiet, pull names from the registry instead:

from plantdoc_predictor.predictor import load_model_registry
rice_models = [m["name"] for m in load_model_registry() if m.get("dataset") == "rice"]

⚠️ Grad-CAM (plantdoc explain) is not available for rice models yet — they are all PyTorch, and Grad-CAM is currently Keras-only.

---

🌍 Community & Contact

PlantDoc-Predictor is growing into a widely used open-source machine learning library for plant disease classification.
Community participation is essential to help expand the PlantDoc Model Zoo and improve the ecosystem.

📈 Current Usage

The library is actively used by researchers, developers, and agritech enthusiasts worldwide.

These downloads reflect the growing interest in AI-powered plant disease detection and encourage further development of the library.

---

🤝 Contribute to the Project

We warmly welcome contributions from the community to help improve and expand PlantDoc-Predictor.

You can contribute by:

• 🧠 Adding new pretrained models to the PlantDoc Model Zoo
• ⚙️ Improving preprocessing or inference pipelines
• 📊 Benchmarking new architectures
• 🧪 Adding test cases and improving reproducibility
• 📚 Improving documentation and tutorials
• 🐛 Fixing bugs and optimizing performance

If you are interested in contributing to the development of this widely used ML library, feel free to open a pull request or start a discussion.

---

📫 Contact the Author

If you have ideas, questions, research collaboration proposals, or model suggestions, feel free to reach out.

Author: Subham Divakar
📧 Email: [e-mail]
🐙 GitHub: https://github.com/shubham10divakar

We welcome researchers, ML engineers, and contributors who want to help make PlantDoc-Predictor the largest open-source model zoo for plant disease detection. 🌿

---

---

📊 Model Zoo Comparison

The following table compares the pretrained models included in PlantDoc-Predictor .

Model	Parameters	Input Size	Accuracy	Framework	Best Use Case
Recursive Additive Attention v1	~12–15M*	224×224	99.7%	Keras	Research-grade model with attention (best overall)
ConvNeXt Base v1	~89M	224×224	99.1%	Keras	Modern CNN alternative to transformers
ConvNeXt Small v1	~50M	224×224	99.5%	Keras	High accuracy with better efficiency
ConvNeXt Tiny v1	~28M	224×224	99.23%	Keras	Efficient modern CNN
Swin Base Patch4 Window7	~88M	224×224	99.1%	PyTorch	Hierarchical Vision Transformer
Swin Tiny Patch4 Window7	~28M	224×224	99.1%	PyTorch	Lightweight transformer
ViT Large 16 v1	~307M	224×224	99.1%	PyTorch	Maximum capacity transformer
ViT Base 16 v1	~86M	224×224	99.1%	PyTorch	Standard transformer baseline
ViT Small 16 v1	~48M	224×224	99.1%	PyTorch	Balanced transformer
ViT Tiny 16 v1	~22M	224×224	99.1%	PyTorch	Lightweight transformer
RegNetX-160 v1 (NEW)	~83M	224×224	99.1%	PyTorch	Efficient CNN with quantizable design
RegNetY-160 v1 (NEW)	~84M	224×224	99.1%	PyTorch	RegNet + SE attention for better accuracy
RegNetY-320 v1 (NEW)	~145M	224×224	99.1%	PyTorch	Largest RegNet variant, highest capacity
DenseNet169 v1	~14M	224×224	99.68%	Keras	Best classical CNN
VGG19 v1	~144M	224×224	98.98%	Keras	High-capacity CNN
DenseNet121 v1	~8M	224×224	98.68%	Keras	Efficient deep CNN
InceptionV3 v1	~23M	299×299	98.2%	Keras	Multi-scale feature extraction
ResNet50 v1	~25M	224×224	97.8%	Keras	Deep residual learning
EfficientNetB50 v1	~30M	224×224	97.8%	Keras	Accuracy-efficiency tradeoff
DenseNet210 v1	~20M	224×224	97.0%	Keras	Very deep dense architecture
MobileNetV2 v1	~3.5M	224×224	96.8%	Keras	Mobile / edge devices
VGG16 v1	~138M	224×224	96.8%	Keras	Benchmark model
AlexNet v1	~60M	224×224	92.8%	Keras	Historical baseline

---

🏆 Model Performance Leaderboard

Top performing models on the PlantVillage 38-class dataset.

Rank	Model	Accuracy	Framework
🥇	Recursive Additive Attention v1	99.70%	Keras
🥈	DenseNet169 v1	99.68%	Keras
🥉	ConvNeXt Small v1	99.50%	Keras
4	ConvNeXt Tiny v1	99.23%	Keras
5	ConvNeXt Base v1	99.10%	Keras
6	Swin Base Patch4 Window7	99.10%	PyTorch
7	Swin Tiny Patch4 Window7	99.10%	PyTorch
8	ViT Base 16 v1	99.10%	PyTorch
9	ViT Large 16 v1	99.10%	PyTorch
10	ViT Small 16 v1	99.10%	PyTorch
11	ViT Tiny 16 v1	99.10%	PyTorch
12	RegNetX-160 v1 (NEW)	99.10%	PyTorch
13	RegNetY-160 v1 (NEW)	99.10%	PyTorch
14	RegNetY-320 v1 (NEW)	99.10%	PyTorch
15	VGG19 v1	98.98%	Keras
16	DenseNet121 v1	98.68%	Keras
17	InceptionV3 v1	98.20%	Keras
18	ResNet50 v1	97.80%	Keras
19	EfficientNetB50 v1	97.80%	Keras
20	DenseNet210 v1	97.00%	Keras
21	MobileNetV2 v1	96.80%	Keras
22	VGG16 v1	96.80%	Keras
23	AlexNet v1	92.80%	Keras

---

📦 Installation

Install directly via pip:

pip install plantdoc-predictor

🚀 How to Use plantdoc_predictor

plantdoc_predictor provides an easy-to-use interface for plant disease prediction using multiple pretrained deep learning models trained on the PlantVillage 38-class dataset.

---

1️⃣ List Available Models

You can view all available pretrained models using:

from plantdoc_predictor import predictor, Predictor

predictor.list_available_models()

Output:-

Available Models:
-----------------
- rec_add_attention_v1      | Input: [224, 224] | Acc: 99.70% | Recursive Additive Attention CNN (best overall model)
- convnext_base_v1          | Input: [224, 224] | Acc: 99.10% | ConvNeXt Base fine-tuned on PlantVillage dataset
- convnext_small_v1         | Input: [224, 224] | Acc: 99.50% | ConvNeXt Small (high accuracy + efficient)
- convnext_tiny_v1          | Input: [224, 224] | Acc: 99.23% | Lightweight ConvNeXt for fast inference
- swin_base_patch4_window7  | Input: [224, 224] | Acc: 99.10% | Swin Transformer (hierarchical vision transformer)
- swin_tiny_patch4_window7  | Input: [224, 224] | Acc: 99.10% | Lightweight Swin Transformer
- vit_base_16_v1            | Input: [224, 224] | Acc: 99.10% | Vision Transformer (base)
- vit_large_16_v1           | Input: [224, 224] | Acc: 99.10% | Vision Transformer (large)
- vit_small_16_v1           | Input: [224, 224] | Acc: 99.10% | Vision Transformer (small)
- vit_tiny_16_v1            | Input: [224, 224] | Acc: 99.10% | Vision Transformer (tiny)
- regnetx_160_v1            | Input: [224, 224] | Acc: 99.10% | RegNetX-160 fine-tuned on PlantVillage dataset
- regnety_160_v1            | Input: [224, 224] | Acc: 99.10% | RegNetY-160 fine-tuned on PlantVillage dataset
- regnety_320_v1            | Input: [224, 224] | Acc: 99.10% | RegNetY-320 fine-tuned on PlantVillage dataset
- densenet169_v1            | Input: [224, 224] | Acc: 99.68% | Best-performing classical CNN
- vgg19_v1                  | Input: [224, 224] | Acc: 98.98% | Deep VGG architecture
- densenet121_v1            | Input: [224, 224] | Acc: 98.68% | Efficient DenseNet variant
- inceptionv3_v1            | Input: [299, 299] | Acc: 98.20% | InceptionV3 with multi-scale feature extraction
- resnet50_v1               | Input: [224, 224] | Acc: 97.80% | Residual learning-based CNN
- efficientnetb50_v1        | Input: [224, 224] | Acc: 97.80% | EfficientNet balancing accuracy & efficiency
- densenet210_v1            | Input: [224, 224] | Acc: 97.00% | Deep DenseNet architecture
- mobilenetv2_v1            | Input: [224, 224] | Acc: 96.80% | Mobile/edge optimized model
- vgg16_v1                  | Input: [224, 224] | Acc: 96.80% | Classic CNN baseline
- alexnet_v1                | Input: [224, 224] | Acc: 92.80% | Early CNN baseline

2. Choose a model from the available ones and use it below.

from plantdoc_predictor import predictor, Predictor
#predictor.list_available_models()

predictor = Predictor(model_name="efficientnetb50_v1", verbose=False)
result = predictor.predict("D:/D/my docs/my docs/projects/plant disease detection on streamlit cloud/streamlit plant disease detection/data/plantvillagedataset/train/color/Blueberry___healthy/0a3f8b2f-9bb1-4da9-85a1-fb5a52c059e2___RS_HL 2478.JPG")
print(result)

Output:-
{
 'model': 'efficientnetb50_v1',
 'label': 'Blueberry___healthy',
 'confidence': 0.9999573230743408
}

when verbose=True in Predictor(model_name="efficientnetb50_v1", verbose=True)

from plantdoc_predictor import predictor, Predictor
#predictor.list_available_models()

predictor = Predictor(model_name="efficientnetb50_v1", verbose=True)
result = predictor.predict("D:/D/my docs/my docs/projects/plant disease detection on streamlit cloud/streamlit plant disease detection/data/plantvillagedataset/train/color/Blueberry___healthy/0a3f8b2f-9bb1-4da9-85a1-fb5a52c059e2___RS_HL 2478.JPG")
print(result)

OUTPUT:-

✔ Using cached file: efficientnetb50_v1.h5
✔ Using cached file: efficientnetb50_v1_labels.json
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
1/1 ━━━━━━━━━━━━━━━━━━━━ 2s 2s/step

================= Prediction Result =================
📂 Image Path     : D:/D/my docs/my docs/projects/plant disease detection on streamlit cloud/streamlit plant disease detection/data/plantvillagedataset/train/color/Blueberry___healthy/0a3f8b2f-9bb1-4da9-85a1-fb5a52c059e2___RS_HL 2478.JPG
🧩 Model Used     : efficientnetb50_v1
✅ Predicted Class: Blueberry___healthy
🔢 Confidence     : 100.00%
{'model': 'efficientnetb50_v1', 'label': 'Blueberry___healthy', 'confidence': 0.9999573230743408}

🧠 Access Full Model now

You can retrieve the full Keras model for advanced use cases like fine-tuning or inspection.

from plantdoc_predictor import Predictor

predictor = Predictor(model_name="resnet50_v1")

model = predictor.get_model()

print(type(model))

Output:
%runfile 'D:/D/my docs/my docs/projects/plantdoc-predictor/plantdoc-predictor/plantdoc_predictor/Test_v1.py' --wdir
2026-03-21 12:39:04.100272: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2026-03-21 12:39:06.296357: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.
2026-03-21 12:39:06.914439: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.
To enable the following instructions: SSE3 SSE4.1 SSE4.2 AVX AVX2 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
<class 'keras.src.models.functional.Functional'>

⚙️ Extract Model Weights

Access raw model weights for research and experimentation.

from plantdoc_predictor import Predictor

predictor = Predictor(model_name="resnet50_v1")

model = predictor.get_model()
#print(type(model))
weights = predictor.get_weights()

print("Total weight tensors:", len(weights))
print("First tensor shape:", weights[0].shape)

Output:
%runfile 'D:/D/my docs/my docs/projects/plantdoc-predictor/plantdoc-predictor/plantdoc_predictor/Test_v1.py' --wdir
Reloaded modules: recursive_additive_attention_v1, predictor
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
Total weight tensors: 320
First tensor shape: (7, 7, 3, 64)

🧩 Layer-wise Weight Info

Get structured information about model weights.

from plantdoc_predictor import Predictor
predictor = Predictor(model_name="resnet50_v1")

model = predictor.get_model()
#print(type(model))
weights = predictor.get_weights()

weights_info = predictor.get_weights_info()

for layer, shapes in weights_info.items():
    print(layer, shapes)

Output:
%runfile 'D:/D/my docs/my docs/projects/plantdoc-predictor/plantdoc-predictor/plantdoc_predictor/Test_v1.py' --wdir
Reloaded modules: recursive_additive_attention_v1, predictor
WARNING:absl:Compiled the loaded model, but the compiled metrics have yet to be built. `model.compile_metrics` will be empty until you train or evaluate the model.
conv1_conv [(7, 7, 3, 64), (64,)]
conv1_bn [(64,), (64,), (64,), (64,)]
conv2_block1_1_conv [(1, 1, 64, 64), (64,)]
conv2_block1_1_bn [(64,), (64,), (64,), (64,)]
conv2_block1_2_conv [(3, 3, 64, 64), (64,)]
conv2_block1_2_bn [(64,), (64,), (64,), (64,)]
conv2_block1_0_conv [(1, 1, 64, 256), (256,)]
conv2_block1_3_conv [(1, 1, 64, 256), (256,)]
conv2_block1_0_bn [(256,), (256,), (256,), (256,)]
conv2_block1_3_bn [(256,), (256,), (256,), (256,)]
conv2_block2_1_conv [(1, 1, 256, 64), (64,)]
conv2_block2_1_bn [(64,), (64,), (64,), (64,)]
conv2_block2_2_conv [(3, 3, 64, 64), (64,)]
conv2_block2_2_bn [(64,), (64,), (64,), (64,)]
conv2_block2_3_conv [(1, 1, 64, 256), (256,)]
conv2_block2_3_bn [(256,), (256,), (256,), (256,)]
conv2_block3_1_conv [(1, 1, 256, 64), (64,)]
conv2_block3_1_bn [(64,), (64,), (64,), (64,)]
conv2_block3_2_conv [(3, 3, 64, 64), (64,)]
conv2_block3_2_bn [(64,), (64,), (64,), (64,)]
conv2_block3_3_conv [(1, 1, 64, 256), (256,)]
conv2_block3_3_bn [(256,), (256,), (256,), (256,)]
conv3_block1_1_conv [(1, 1, 256, 128), (128,)]
conv3_block1_1_bn [(128,), (128,), (128,), (128,)]
conv3_block1_2_conv [(3, 3, 128, 128), (128,)]
conv3_block1_2_bn [(128,), (128,), (128,), (128,)]
conv3_block1_0_conv [(1, 1, 256, 512), (512,)]
conv3_block1_3_conv [(1, 1, 128, 512), (512,)]
conv3_block1_0_bn [(512,), (512,), (512,), (512,)]
conv3_block1_3_bn [(512,), (512,), (512,), (512,)]
conv3_block2_1_conv [(1, 1, 512, 128), (128,)]
conv3_block2_1_bn [(128,), (128,), (128,), (128,)]
conv3_block2_2_conv [(3, 3, 128, 128), (128,)]
conv3_block2_2_bn [(128,), (128,), (128,), (128,)]
conv3_block2_3_conv [(1, 1, 128, 512), (512,)]
conv3_block2_3_bn [(512,), (512,), (512,), (512,)]
conv3_block3_1_conv [(1, 1, 512, 128), (128,)]
conv3_block3_1_bn [(128,), (128,), (128,), (128,)]
conv3_block3_2_conv [(3, 3, 128, 128), (128,)]
conv3_block3_2_bn [(128,), (128,), (128,), (128,)]
conv3_block3_3_conv [(1, 1, 128, 512), (512,)]
conv3_block3_3_bn [(512,), (512,), (512,), (512,)]
conv3_block4_1_conv [(1, 1, 512, 128), (128,)]
conv3_block4_1_bn [(128,), (128,), (128,), (128,)]
conv3_block4_2_conv [(3, 3, 128, 128), (128,)]
conv3_block4_2_bn [(128,), (128,), (128,), (128,)]
conv3_block4_3_conv [(1, 1, 128, 512), (512,)]
conv3_block4_3_bn [(512,), (512,), (512,), (512,)]
conv4_block1_1_conv [(1, 1, 512, 256), (256,)]
conv4_block1_1_bn [(256,), (256,), (256,), (256,)]
conv4_block1_2_conv [(3, 3, 256, 256), (256,)]
conv4_block1_2_bn [(256,), (256,), (256,), (256,)]
conv4_block1_0_conv [(1, 1, 512, 1024), (1024,)]
conv4_block1_3_conv [(1, 1, 256, 1024), (1024,)]
conv4_block1_0_bn [(1024,), (1024,), (1024,), (1024,)]
conv4_block1_3_bn [(1024,), (1024,), (1024,), (1024,)]
conv4_block2_1_conv [(1, 1, 1024, 256), (256,)]
conv4_block2_1_bn [(256,), (256,), (256,), (256,)]
conv4_block2_2_conv [(3, 3, 256, 256), (256,)]
conv4_block2_2_bn [(256,), (256,), (256,), (256,)]
conv4_block2_3_conv [(1, 1, 256, 1024), (1024,)]
conv4_block2_3_bn [(1024,), (1024,), (1024,), (1024,)]
conv4_block3_1_conv [(1, 1, 1024, 256), (256,)]
conv4_block3_1_bn [(256,), (256,), (256,), (256,)]
conv4_block3_2_conv [(3, 3, 256, 256), (256,)]
conv4_block3_2_bn [(256,), (256,), (256,), (256,)]
conv4_block3_3_conv [(1, 1, 256, 1024), (1024,)]
conv4_block3_3_bn [(1024,), (1024,), (1024,), (1024,)]
conv4_block4_1_conv [(1, 1, 1024, 256), (256,)]
conv4_block4_1_bn [(256,), (256,), (256,), (256,)]
conv4_block4_2_conv [(3, 3, 256, 256), (256,)]
conv4_block4_2_bn [(256,), (256,), (256,), (256,)]
conv4_block4_3_conv [(1, 1, 256, 1024), (1024,)]
conv4_block4_3_bn [(1024,), (1024,), (1024,), (1024,)]
conv4_block5_1_conv [(1, 1, 1024, 256), (256,)]
conv4_block5_1_bn [(256,), (256,), (256,), (256,)]
conv4_block5_2_conv [(3, 3, 256, 256), (256,)]
conv4_block5_2_bn [(256,), (256,), (256,), (256,)]
conv4_block5_3_conv [(1, 1, 256, 1024), (1024,)]
conv4_block5_3_bn [(1024,), (1024,), (1024,), (1024,)]
conv4_block6_1_conv [(1, 1, 1024, 256), (256,)]
conv4_block6_1_bn [(256,), (256,), (256,), (256,)]
conv4_block6_2_conv [(3, 3, 256, 256), (256,)]
conv4_block6_2_bn [(256,), (256,), (256,), (256,)]
conv4_block6_3_conv [(1, 1, 256, 1024), (1024,)]
conv4_block6_3_bn [(1024,), (1024,), (1024,), (1024,)]
conv5_block1_1_conv [(1, 1, 1024, 512), (512,)]
conv5_block1_1_bn [(512,), (512,), (512,), (512,)]
conv5_block1_2_conv [(3, 3, 512, 512), (512,)]
conv5_block1_2_bn [(512,), (512,), (512,), (512,)]
conv5_block1_0_conv [(1, 1, 1024, 2048), (2048,)]
conv5_block1_3_conv [(1, 1, 512, 2048), (2048,)]
conv5_block1_0_bn [(2048,), (2048,), (2048,), (2048,)]
conv5_block1_3_bn [(2048,), (2048,), (2048,), (2048,)]
conv5_block2_1_conv [(1, 1, 2048, 512), (512,)]
conv5_block2_1_bn [(512,), (512,), (512,), (512,)]
conv5_block2_2_conv [(3, 3, 512, 512), (512,)]
conv5_block2_2_bn [(512,), (512,), (512,), (512,)]
conv5_block2_3_conv [(1, 1, 512, 2048), (2048,)]
conv5_block2_3_bn [(2048,), (2048,), (2048,), (2048,)]
conv5_block3_1_conv [(1, 1, 2048, 512), (512,)]
conv5_block3_1_bn [(512,), (512,), (512,), (512,)]
conv5_block3_2_conv [(3, 3, 512, 512), (512,)]
conv5_block3_2_bn [(512,), (512,), (512,), (512,)]
conv5_block3_3_conv [(1, 1, 512, 2048), (2048,)]
conv5_block3_3_bn [(2048,), (2048,), (2048,), (2048,)]
dense [(2048, 38), (38,)]

🏆 Top-K Predictions

Return the top N ranked predictions instead of just the best one. Works on both Keras and PyTorch models.

from plantdoc_predictor import Predictor

predictor = Predictor(model_name="densenet169_v1")
result = predictor.predict("leaf.jpg", top_k=3)
print(result)

# {
#   'model': 'densenet169_v1',
#   'label': 'Apple___Apple_scab',
#   'confidence': 0.984,
#   'top_k': [
#       {'label': 'Apple___Apple_scab',   'confidence': 0.984},
#       {'label': 'Apple___Black_rot',    'confidence': 0.011},
#       {'label': 'Apple___Cedar_apple_rust', 'confidence': 0.003}
#   ]
# }

---

🖼️ PIL Image Input

predict() accepts a PIL.Image directly — no need to save to disk first. Perfect for Streamlit, FastAPI, and in-memory pipelines.

from PIL import Image
from plantdoc_predictor import Predictor

predictor = Predictor(model_name="densenet169_v1")
result = predictor.predict(Image.open("leaf.jpg"))

---

🔥 Feature Extraction (Research-Grade)

Extract intermediate representations from any layer for feature-space SMOTE, clustering, or embedding analysis. (Keras models)

from plantdoc_predictor import Predictor

predictor = Predictor(model_name="densenet169_v1")

# Default: second-to-last layer (embedding); or pass layer_name="..."
features = predictor.extract_features("leaf.jpg")
print(features.shape)   # e.g. (1, 1664)

# Inspect available layers to target a specific one
print(predictor.list_layers())
features = predictor.extract_features("leaf.jpg", layer_name="conv5_block16_concat")

---

📦 Batch Prediction

Run inference over a list of paths or PIL.Image objects and export results.

from plantdoc_predictor import BatchPredictor

bp = BatchPredictor(model_name="densenet169_v1")
results = bp.run(["img1.jpg", "img2.jpg", "img3.jpg"], top_k=3)

bp.export_csv(results, "results.csv")
bp.export_json(results, "results.json")
bp.summary(results)

---

🛡️ GuardedPredictor — Reject Non-Leaf Images

The disease model is a closed-world 38-class classifier — feed it a dog photo and it still returns a confident disease label. GuardedPredictor adds a two-layer guard:

1. CLIP leaf guard — openai/clip-vit-base-patch32 scores how leaf-like the image is.
2. Confidence floor (optional) — rejects low-confidence disease predictions.

from plantdoc_predictor import GuardedPredictor

gp = GuardedPredictor(model_name="densenet169_v1", guard_threshold=0.5)

# Non-leaf image → rejected
gp.predict("dog.jpg")
# {'model': 'densenet169_v1', 'is_leaf': False, 'guard_score': 0.13,
#  'label': 'unknown', 'confidence': None, 'crop': None, 'disease': None, 'is_healthy': None}

# Leaf image → passes guard, returns parsed result
gp.predict("apple_scab.jpg")
# {'model': 'densenet169_v1', 'is_leaf': True, 'guard_score': 0.88,
#  'label': 'Apple___Apple_scab', 'confidence': 0.98,
#  'crop': 'Apple', 'disease': 'Apple scab', 'is_healthy': False}

Parameter	Default	Description
guard_threshold	0.5	CLIP leaf-score cutoff. Below this → rejected as non-leaf.
min_confidence	0.0 (off)	Disease-model confidence floor. Below this → "unknown".

ℹ️ CLIP lazy-loads on first predict() (~400 MB one-time download, cached in ~/.cache/huggingface/). Importing GuardedPredictor is instant.

---

🔥 ExplainablePredictor — Grad-CAM Heatmaps (NEW 🚀)

See why a model made its prediction. ExplainablePredictor produces a Grad-CAM heatmap highlighting the leaf regions that drove the result — ideal for research papers, debugging, and user trust. (Keras models; PyTorch ViT/Swin support is on the roadmap.)

Example — Apple scab leaf: the heatmap concentrates on the scab lesions, exactly the regions a plant pathologist would inspect.

Input leaf	Grad-CAM heatmap

from plantdoc_predictor import ExplainablePredictor

ep = ExplainablePredictor(model_name="densenet169_v1")

# Predict + save an overlaid heatmap to disk
result = ep.explain("leaf.JPG", save_to="heatmap.jpg")
print(result)
# {
#   'model': 'densenet169_v1',
#   'label': 'Apple___Apple_scab',
#   'confidence': 0.98,
#   'crop': 'Apple', 'disease': 'Apple scab', 'is_healthy': False,
#   'layer_name': 'conv5_block16_concat',   # auto-detected last conv layer
#   'heatmap_path': 'heatmap.jpg'
# }

Options:

# Target a specific layer (defaults to the last conv feature map)
ep.explain("leaf.jpg", save_to="cam.jpg", layer_name="conv5_block16_concat")

# Explain a class other than the predicted one
ep.explain("leaf.jpg", save_to="cam.jpg", class_index=12)

# Control overlay strength (0–1)
ep.explain("leaf.jpg", save_to="cam.jpg", alpha=0.6)

# Get the raw arrays back instead of (or in addition to) saving
result = ep.explain("leaf.jpg", return_heatmap=True)
heatmap = result["heatmap"]   # HxW float array in [0, 1]
overlay = result["overlay"]   # HxWx3 uint8 RGB image

Parameter	Default	Description
save_to	None	Path to write the overlaid heatmap (.jpg/.png).
layer_name	auto	Target conv layer; defaults to the last 4D feature-map layer.
class_index	predicted	Which class to explain.
alpha	0.4	Heatmap blend strength.
return_heatmap	False	Also return the raw heatmap + overlay arrays.

---

💻 Command-Line Interface

After installing, the plantdoc command is available in your terminal.

# List all available models
plantdoc models

# Predict a single image
plantdoc predict leaf.jpg

# Choose a model + top-3 predictions
plantdoc predict leaf.jpg --model densenet169_v1 --top-k 3

# Machine-readable JSON output
plantdoc predict leaf.jpg --json

# Guarded prediction — reject non-leaf images via the CLIP guard
plantdoc predict leaf.jpg --guard --guard-threshold 0.5
plantdoc predict leaf.jpg --guard --min-confidence 0.6

# Batch a whole folder → export results
plantdoc predict ./images_folder/ --output results.csv
plantdoc predict ./images_folder/ --output results.json

# Grad-CAM explanation — write a heatmap overlay
plantdoc explain leaf.jpg --save-to cam.jpg
plantdoc explain leaf.jpg --model densenet169_v1 --layer conv5_block16_concat --alpha 0.6
plantdoc explain leaf.jpg --json

---

📝 License

This project is licensed under the MIT License.

You are free to:

• ✅ Use the library for both commercial and academic purposes
• 🔧 Modify, distribute, or integrate it into your own software
• 🌍 Reference and extend it in research or production projects

Just ensure you include the original copyright notice and this license file.

See the full text in the LICENSE file.

---

🌍 Contributing

Contributions are warmly welcomed! 🌱

We value community participation to make PlantDoc-Predictor more robust, accurate, and user-friendly.

You can contribute by:

• 🧠 Adding new pretrained or fine-tuned models
• ⚙️ Improving preprocessing or postprocessing modules
• 🧩 Extending model registry or dataset support
• 🧪 Writing new test cases for reproducibility
• 🐛 Fixing bugs and optimizing performance
• 📚 Improving documentation and adding usage examples

🔧 Steps to Contribute

1. Fork this repository
2. Create your feature branch

   git checkout -b feature/your-feature
   

📫 Contact

Author: Subham Divakar
Email: [e-mail]
GitHub: shubham10divakar
PyPI: https://pypi.org/project/plantdoc-predictor/
LinkedIn: linkedin.com/in/subhamdivakar
Project Website: My Site

If you have any questions, collaboration ideas, or model suggestions — feel free to reach out!
You can also open an issue or submit a pull request in the GitHub repository.

---

❤️ Acknowledgements

A heartfelt thank you to all the open-source contributors and researchers who made this project possible:

• 🌿 PlantVillage Dataset — for providing an invaluable resource for agricultural disease detection.
• 🤖 TensorFlow/Keras — for powering deep learning model training and inference.
• 🧩 Python Open Source Community — for libraries that make AI tools easier to build.
• 🧪 Researchers & Reviewers — for advancing the field of AI in agriculture.
• 💻 Contributors — for helping test, improve, and document PlantDoc-Predictor.

Your support continues to make plant disease prediction accessible and impactful across the world.

---

“Empowering agriculture with AI — one leaf at a time.” 🌾
— Subham Divakar, Creator of PlantDoc-Predictor