How tqwei05/mvpdr works
MVPDR is a research benchmark and few-shot multimodal plant-disease recognition baseline. The documented workflow is to select a dataset/configuration, train the baseline, and evaluate recognition performance. The graph preserves the three supported datasets and the text-prompt and CLIP branches; sampled excerpts do not expose all of main.py’s setup and dispatch logic, so uncertain wiring is omitted.
Main entry points: main.py Experiment runner · __init__.py Dataset registry · utils.py Model utilities · plantwild.py PlantWild loader
Components 21
Experiment workflow
- Experiment runner
main.py - Dataset configs
configs
Datasets and prompts
- Dataset registry
datasets/__init__.py - PlantWild loader
datasets/plantwild.py - PlantDoc loader
datasets/plantdoc.py - PlantVillage loader
datasets/plantvillage.py - Dataset utilities
datasets/utils.py - Disease prompts
gpt_files_plt44
Multimodal modeling
- Model utilities
utils.py - BPE tokenizer
clip/simple_tokenizer.py - CLIP API
clip/clip.py - CLIP encoders
clip/model.py - Visual adapter
main.py - Prompt adapter
main.py
Training and evaluation
- Baseline training
main.py - Recognition metrics
utils.py - Adapter checkpoints
main.py - Evaluation results
main.py
Other components
- Experimenter
- Plant image data
- CLIP weights
Connections 25
- Experimenter → Experiment runner: runs
- Experiment runner → Dataset configs: reads config
- Experiment runner → Dataset registry: builds dataset
- Dataset registry → PlantWild loader: dispatches
- Dataset registry → PlantDoc loader: dispatches
- Dataset registry → PlantVillage loader: dispatches
- PlantWild loader → Dataset utilities: uses
- PlantDoc loader → Dataset utilities: uses
- PlantVillage loader → Dataset utilities: uses
- Plant image data → PlantWild loader: provides images
- Plant image data → PlantDoc loader: provides images
- Plant image data → PlantVillage loader: provides images
- Experiment runner → Model utilities: uses
- Disease prompts → Model utilities: provides prompts
- Model utilities → BPE tokenizer: tokenizes text
- Model utilities → CLIP encoders: encodes prompts
- Experiment runner → CLIP API: loads CLIP
- CLIP API → CLIP encoders: builds model
- CLIP weights → CLIP API: provides weights
- Experiment runner → Baseline training: trains and evaluates
- Baseline training → Visual adapter: updates
- Baseline training → Prompt adapter: updates
- Baseline training → Recognition metrics: scores predictions
- Baseline training → Adapter checkpoints: saves and loads
- Baseline training → Evaluation results: reports metrics