mirror of
https://github.com/huggingface/candle.git
synced 2025-06-16 18:48:51 +00:00
6056fd5c90
* onnx: fix pad, unsqueeze both implementations have off-by-one errors: - Pad 'reflect' cycle for eg `dim==3` is `[0,1,2,1]` which has length of 4 (or `dim*2 - 2`) not 5 (current code `dim*2 - 1`) - Unsqueeze(-1) for tensor with `dim==3` should be 3 (ie `dim+index+1`) not 2 (ie currently `dim+index`) in addition, Pad is incorrectly calculating the starting padding. If we want to pad out 2 elements to the start, and we have this cycle of indices of length 6, then we should skip 4 elements, but currently we skip 2. A more visual representation of what's going on is below: ``` pad_start: 2 data: [a,b,c,d] indices: [0, 1, 2, 3, 2, 1, 0, 1, 2, 3, 2, 1, 0, ..] // zigzag between 0..4 actual: skip [ c d| c b a b] expected: ~ skip ~ [ c b| a b c d] ``` The values between `[` and `|` are padding and the values between `|` and `]` in the example should match the original data being padded. * Fix clippy lints. --------- Co-authored-by: Laurent <laurent.mazare@gmail.com>
83 lines
2.6 KiB
Rust
83 lines
2.6 KiB
Rust
//! EVA-02: Explore the limits of Visual representation at scAle
|
|
//! https://github.com/baaivision/EVA
|
|
|
|
#[cfg(feature = "mkl")]
|
|
extern crate intel_mkl_src;
|
|
|
|
#[cfg(feature = "accelerate")]
|
|
extern crate accelerate_src;
|
|
|
|
use clap::Parser;
|
|
|
|
use candle::{DType, Device, IndexOp, Result, Tensor, D};
|
|
use candle_nn::{Module, VarBuilder};
|
|
use candle_transformers::models::eva2;
|
|
|
|
/// Loads an image from disk using the image crate, this returns a tensor with shape
|
|
/// (3, 448, 448). OpenAI normalization is applied.
|
|
pub fn load_image448_openai_norm<P: AsRef<std::path::Path>>(p: P) -> Result<Tensor> {
|
|
let img = image::ImageReader::open(p)?
|
|
.decode()
|
|
.map_err(candle::Error::wrap)?
|
|
.resize_to_fill(448, 448, image::imageops::FilterType::Triangle);
|
|
let img = img.to_rgb8();
|
|
let data = img.into_raw();
|
|
let data = Tensor::from_vec(data, (448, 448, 3), &Device::Cpu)?.permute((2, 0, 1))?;
|
|
let mean =
|
|
Tensor::new(&[0.48145466f32, 0.4578275, 0.40821073], &Device::Cpu)?.reshape((3, 1, 1))?;
|
|
let std = Tensor::new(&[0.26862954f32, 0.261_302_6, 0.275_777_1], &Device::Cpu)?
|
|
.reshape((3, 1, 1))?;
|
|
(data.to_dtype(candle::DType::F32)? / 255.)?
|
|
.broadcast_sub(&mean)?
|
|
.broadcast_div(&std)
|
|
}
|
|
|
|
#[derive(Parser)]
|
|
struct Args {
|
|
#[arg(long)]
|
|
model: Option<String>,
|
|
|
|
#[arg(long)]
|
|
image: String,
|
|
|
|
/// Run on CPU rather than on GPU.
|
|
#[arg(long)]
|
|
cpu: bool,
|
|
}
|
|
|
|
pub fn main() -> anyhow::Result<()> {
|
|
let args = Args::parse();
|
|
|
|
let device = candle_examples::device(args.cpu)?;
|
|
|
|
let image = load_image448_openai_norm(args.image)?.to_device(&device)?;
|
|
println!("loaded image {image:?}");
|
|
|
|
let model_file = match args.model {
|
|
None => {
|
|
let api = hf_hub::api::sync::Api::new()?;
|
|
let api = api.model("vincent-espitalier/candle-eva2".into());
|
|
api.get("eva02_base_patch14_448.mim_in22k_ft_in22k_in1k_adapted.safetensors")?
|
|
}
|
|
Some(model) => model.into(),
|
|
};
|
|
let vb = unsafe { VarBuilder::from_mmaped_safetensors(&[model_file], DType::F32, &device)? };
|
|
|
|
let model = eva2::vit_base(vb)?;
|
|
println!("model built");
|
|
let logits = model.forward(&image.unsqueeze(0)?)?;
|
|
let prs = candle_nn::ops::softmax(&logits, D::Minus1)?
|
|
.i(0)?
|
|
.to_vec1::<f32>()?;
|
|
let mut prs = prs.iter().enumerate().collect::<Vec<_>>();
|
|
prs.sort_by(|(_, p1), (_, p2)| p2.total_cmp(p1));
|
|
for &(category_idx, pr) in prs.iter().take(5) {
|
|
println!(
|
|
"{:24}: {:.2}%",
|
|
candle_examples::imagenet::CLASSES[category_idx],
|
|
100. * pr
|
|
);
|
|
}
|
|
Ok(())
|
|
}
|