segment-anything-ncnn

a ncnn example of segment-anything

PS:

the image_embeddings maybe take a long time, because of some MultiHeadAttention ops isn't fused.
maybe we should use pnnx to optimize this.

model support:

ViT-B SAM model

models are available in Baidu Pan and Google Drive

Run

mkdir -p build
cd build
cmake ..
make
./ncnn_sam

time profile

op type         avg time(ms)    %
MatMul          2268.23         21.14%
Reshape         2199.36         20.51%
InnerProduct    1899.4          17.71%
GELU            1809.85         16.87%
BinaryOp        1351.21         12.61%
Softmax         513.15          4.78%
Permute         442.24          4.12%
Crop            106.28          0.99%
LayerNorm       65.11           0.61%
Padding         35.7            0.33%
Convolution     34.43           0.32%
MemoryData      2.76            0.03%
Split           0.00            0%
total time:     10727.72