NavdeepSinghNegi999/DeepVisionIntelligence
🧠 DeepVisionIntelligence — An end-to-end multimodal AI system that transforms raw visual inputs into structured language representations using deep CNN feature extraction and transformer-based decoding.
What's novel
🧠 DeepVisionIntelligence — An end-to-end multimodal AI system that transforms raw visual inputs into structured language representations using deep CNN feature extraction and transformer-based decoding.
Code Analysis
20 files read · 4 roundsA Transformer-based image captioning system that uses a frozen EfficientNetB0 encoder and a single-layer Transformer decoder to generate word-by-word captions from images, deployed via FastAPI on Google Cloud Run.
Strengths
Clean separation of concerns across data, model, training, inference, and evaluation modules; the training pipeline is well-structured with MLflow tracking, warmup LR scheduling, and proper masked loss computation.
Weaknesses
No automated tests exist (only a notebook), the Transformer is a single encoder block and single decoder block (not a true multi-layer stack), the FastAPI app has no input validation or error handling, and the generation loop has a subtle bug where it indexes predictions by loop counter rather than sequence position.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
22
Contributors
1
Files
51
Active weeks
3
Repository
Language
Python
Stars
1
Forks
0
License
—