Course Objective & Scope
The transformer architecture has become standard in computer vision too, but whether to choose a CNN or a ViT depends on data volume, resolution and hardware. The programme explains patch embedding, positional encoding and self-attention in an image context and compares ViT, DeiT, Swin Transformer and DINOv2.
Fine-tuning with Hugging Face Transformers and timm, parameter-efficient adaptation with LoRA, attention map visualisation and optimising inference cost are covered hands-on. Participants run a CNN-versus-ViT comparison experiment for their own task.
Who Should Attend?
- Computer vision engineers who have worked with CNN-based models
- AI specialists responsible for model architecture choices
- Data scientists in research and product development teams
- Prerequisite: PyTorch and deep learning fundamentals
Key Takeaways
- Explain the logic of self-attention and patch embedding
- Choose between ViT, Swin and DINOv2 per task
- Adapt a ViT model through fine-tuning with Hugging Face
- Run parameter-efficient training on limited resources with LoRA
- Interpret attention maps to analyse model behaviour
Curriculum
Transformer Logic for Images
- Patch embedding and positional encoding
- Multi-head self-attention and the CLS token
The ViT Family and Hierarchical Models
- ViT, DeiT and the distillation token
- Swin Transformer and shifted window attention
Self-Supervised ViTs
- Pre-training with DINOv2 and MAE
- Feature extraction and linear probing
Fine-Tuning and Parameter-Efficient Adaptation
- Training with Hugging Face Transformers and timm
- LoRA and layer-wise learning rates
Comparison with CNNs and Production Decisions
- Trade-offs in data volume, resolution and latency
- Attention map analysis and inference optimisation
Other Courses in This Category
-
Document preprocessing, text recognition, layout analysis and table/form extraction with PaddleOCR, Surya, LayoutLMv3 and Donut, for technical teams turning invoices, contracts and operational documents into structured data.
- Live Course
- 5 modules
-
RTSP video streams, GStreamer/FFmpeg, real-time detection, multi-object tracking with ByteTrack/BoT-SORT, speed/crowd analytics and NVIDIA DeepStream, for teams extracting live operational insights from camera networks.
- Live Course
- 5 modules
-
Human action, ergonomics and anomaly detection with 3D CNNs (SlowFast, X3D), VideoMAE/TimeSformer and skeleton-based models (MMPose, ST-GCN), for teams building workplace safety, assembly-line and behaviour analysis systems.
- Live Course
- 5 modules
-
Shared embedding spaces with CLIP/SigLIP, Qwen2.5-VL, LLaVA and InternVL architectures, visual QA, multimodal RAG and QLoRA fine-tuning, for engineers building systems that reason across images and text.
- Live Course
- 5 modules
-
Surface defect detection on production lines, unsupervised anomaly detection with anomalib (PatchCore, PaDiM, EfficientAd), industrial camera/lighting design and PLC integration, for quality and manufacturing engineering teams.
- Live Course
- 5 modules
-
Inpatient fall prevention, ICU monitoring, hand-hygiene and PPE compliance, operating-room instrument counting and privacy-preserving edge anonymisation, for hospital digital transformation and healthtech teams.
- Live Course
- 5 modules
Get in touch
