xujiashuai/cv_vitb16_classification_vision-efficient-tuning-prompt
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

ViT-B/16 Vision Efficient Tuning - Prompt (cv_vitb16_classification_vision-efficient-tuning-prompt)

ViT-B/16 with Prompt Tuning for image classification. Prompt tokens are prepended to patch embeddings at each transformer block.

  • Model ID: iic/cv_vitb16_classification_vision-efficient-tuning-prompt
  • Task: Image Classification (CIFAR-100, 100 classes)
  • Framework: PyTorch
  • Backbone: ViT-B/16 (patch_size=16, embed_dim=768, depth=12, num_heads=12)
  • Prompt Length: 10 tokens per block

Verification Environment

ItemValue
NPUAscend910B4
CANN8.5.1
PyTorch2.9.0+cpu
torch_npu2.9.0.post1
Python3.11.14

Accuracy Evaluation

CPU vs NPU cosine similarity:

MetricValue
Cosine Similarity0.999992

Inference Output

Output shape: [1, 100]
Logits sample: [-1.027, -3.788, 0.819, -1.239, -1.530]
Predicted class: 61
SUCCESS

Reproduction

python3 inference.py
python3 evaluate.py

Materials

  • inference.py - NPU inference script
  • evaluate.py - CPU vs NPU cosine similarity evaluation