huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-15 18:28:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

FastViT fixes. (#2452)

Browse Source 
* correct optional SE layer dimensions.
 * head_dim instead of num_heads is 32.
 * update test example output.
This commit is contained in:
Jani Monoses
2024-08-28 12:20:09 +03:00
committed by GitHub
parent aafa24ed93
commit 29e25c458d
3 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@ -241,7 +241,7 @@ If you have an addition to this list, please submit a pull request.
- Parler-TTS, text-to-speech model. - Parler-TTS, text-to-speech model.
- Computer Vision Models. - Computer Vision Models.
- DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT, - DINOv2, ConvMixer, EfficientNet, ResNet, ViT, VGG, RepVGG, ConvNeXT,
ConvNeXTv2, MobileOne, EfficientVit (MSRA), MobileNetv4, Hiera. ConvNeXTv2, MobileOne, EfficientVit (MSRA), MobileNetv4, Hiera, FastViT.
- yolo-v3, yolo-v8. - yolo-v3, yolo-v8.
- Segment-Anything Model (SAM). - Segment-Anything Model (SAM).
- SegFormer. - SegFormer.

10
candle-examples/examples/fastvit/README.md
View File

@ -12,9 +12,9 @@ $ cargo run --example fastvit --release -- --image candle-examples/examples/yolo
loaded image Tensor[dims 3, 256, 256; f32] loaded image Tensor[dims 3, 256, 256; f32]
model built model built
mountain bike, all-terrain bike, off-roader: 43.45% mountain bike, all-terrain bike, off-roader: 52.67%
bicycle-built-for-two, tandem bicycle, tandem: 14.16% bicycle-built-for-two, tandem bicycle, tandem: 7.93%
unicycle, monocycle : 4.12% unicycle, monocycle : 3.46%
crash helmet : 2.26% maillot : 1.32%
alp : 1.40% crash helmet : 1.28%
``` ```

6
candle-transformers/src/models/fastvit.rs
View File

@ -339,8 +339,8 @@ fn positional_encoding(dim: usize, vb: VarBuilder) -> Result<Func<'static>> {
fn attention(dim: usize, vb: VarBuilder) -> Result<Func<'static>> { fn attention(dim: usize, vb: VarBuilder) -> Result<Func<'static>> {
let qkv = linear_no_bias(dim, dim * 3, vb.pp("qkv"))?; let qkv = linear_no_bias(dim, dim * 3, vb.pp("qkv"))?;
let proj = linear(dim, dim, vb.pp("proj"))?; let proj = linear(dim, dim, vb.pp("proj"))?;
let num_heads = 32; let head_dim = 32;
let head_dim = dim / num_heads; let num_heads = dim / head_dim;
let scale = (head_dim as f64).powf(-0.5); let scale = (head_dim as f64).powf(-0.5);
Ok(Func::new(move |xs| { Ok(Func::new(move |xs| {
@ -434,7 +434,7 @@ fn fastvit_patch_embed(
) -> Result<Func<'static>> { ) -> Result<Func<'static>> {
let lk = conv_norm(in_channels, out_channels, 7, 2, vb.pp("proj.0.large_conv"))?; let lk = conv_norm(in_channels, out_channels, 7, 2, vb.pp("proj.0.large_conv"))?;
let sk = conv_norm(in_channels, out_channels, 3, 2, vb.pp("proj.0.small_conv"))?; let sk = conv_norm(in_channels, out_channels, 3, 2, vb.pp("proj.0.small_conv"))?;
let se = squeeze_and_excitation(out_channels, out_channels / 16, vb.pp("proj.0.se")); let se = squeeze_and_excitation(out_channels, out_channels / 4, vb.pp("proj.0.se"));
let mb = mobileone_block(out_channels, out_channels, 1, 1, 0, true, vb.pp("proj.1"))?; let mb = mobileone_block(out_channels, out_channels, 1, 1, 0, true, vb.pp("proj.1"))?;
Ok(Func::new(move |xs| { Ok(Func::new(move |xs| {