Morph MORPH
SIGN IN SIGN UP
dhruvhead / vision.cpp UNCLAIMED
0 0 55 C++

COMMITS

main
June 6, 2025
A
birefnet: add sigmoid at end, fix various tests
Acly committed
7dae1c1
May 26, 2025
A
sam: fix tests, use WHCN conv_transpose_2d
Acly committed
2ca942a
May 16, 2025
A
birefnet: implement decoder & squeeze block
Acly committed
317bc5e
A
birefnet: implement deformable conv, aspp, and basic decoder block
Acly committed
5b7fae1
May 14, 2025
A
birefnet: finish encoder implementation
Acly committed
000c0cc
May 9, 2025
A
ggml: add some utility to manage common usage of ggml context/buffer/backend/graph/allocr
Acly committed
8dbc4b7
A
birefnet: test for entire swin_transformer
Acly committed
bae78b2
May 8, 2025
A
birefnet: dont dump the full relative position bias for window attention - it's really big after all (800MB for full model) - compute it from bias table and window index like in pytorch
Acly committed
fd4ee24
May 6, 2025
A
birefnet: implement patch merging and attention mask
Acly committed
ce5faf1
May 5, 2025
A
birefnet: implement swin_block, support mask & shift for window attention
Acly committed
cbbd8eb
May 4, 2025
A
birefnet: implement patch_embed and window_attention
Acly committed
d2bd56f
April 30, 2025
A
ggml: add test for WHCN conv_transpose_2d
Acly committed
116fdec
April 28, 2025
A
Rebase to ggml master; add tests for window_partition
Acly committed
fa6734f
April 14, 2025
A
sam: make conv_transpose_2d use same layout as regular conv_2d and add benchmark
Acly committed
2ecfd02
April 11, 2025
A
sam: add NHWC conv-2d and conv-transpose-2d for vulkan - now runs 100% on gpu - image embeds: 120ms - mask decode: 16ms
Acly committed
d175274
A
run tests for vulkan backend
Acly committed
9f7b1e6
April 1, 2025
A
sam: support vulkan backend (not optimized)
Acly committed
2d11d14
March 29, 2025
A
sam: conv-2d changes and depwthwise benchmark
Acly committed
4e563fb
March 18, 2025
A
sam: switch mask decoder to use NHWC image embeddings directly - there are still some permutes for conv_2d_transpose
Acly committed
52ff9ee
March 12, 2025
A
sam: implement box prompt
Acly committed
7177bdf
March 11, 2025
A
sam: move to cpp/hpp library
Acly committed
56db2d9
March 10, 2025
A
ggml: move convolution and batch norm to ggml kernels
Acly committed
b8167de
March 8, 2025
A
sam: replace batch norm with built-in
Acly committed
91afdc1
March 7, 2025
A
sam: switch to NHWC layout - fastaaa - needs fix: depthwise conv2d accumulation is not thread-safe
Acly committed
5461e61
March 3, 2025
A
cmake: add benchmark executable and enable debug symbols for ggml libs
Acly committed
93e6a6d
A
ggml: add a version of conv_2d which works on channels-most-contiguous memory layout
Acly committed
16ae20f
A
Make workbench use ggml backends with blas
Acly committed
ddf9cea
A
sam: remove unused dilation parameters from conv_2d
Acly committed
6668af9
March 1, 2025
A
Add benchmark executable for profiling
Acly committed
4b967aa
A
sam: Precompute sqare root for batch norm
Acly committed
1a61d8a