🚨Fixed wrong padding value in OWLv2 #41938

gjamesgoenawan · 2025-10-29T14:54:23Z

What does this PR do?

This PR proposes changing the default padding value from 0.5 to 0.0 in OWLv2. While OWLv1 originally used a padding value of 0.5 (gray) as described in its paper [1], OWLv2 adopts 0.0 instead [2], consistent with its official implementation [3]. Using the incorrect padding value (0.5) leads to degraded performance on the LVIS dataset.

Implementation	LVIS mAP
Scenic	43.9
Transformers (0.5 padding)	43.4
Transformers (0.0 padding)	44.0

Reproducing the results

Testing scripts:
The following scripts explicitly resized and pad the image beforehand so no padding will be done in the processor.

import os
import re
import torch
import argparse
import warnings
import numpy as np
import torch.distributed as dist
from torch.utils.data import Dataset, DataLoader, DistributedSampler
from transformers import Owlv2Processor, Owlv2ForObjectDetection
from PIL import Image
from lvis import LVIS, LVISResults, LVISEval
from tqdm import tqdm


warnings.filterwarnings("ignore")

NOT_PROMPTABLE_MARKER = '#'
PROMPT_TEMPLATES = [
    'itap of a {}.',
    'a bad photo of the {}.',
    'a origami {}.',
    'a photo of the large {}.',
    'a {} in a video game.',
    'art of the {}.',
    'a photo of the small {}.',
]

def _canonicalize_string(string: str) -> str:
    string = string.lower()
    string = re.sub(f'[^a-z0-9-{NOT_PROMPTABLE_MARKER} ]', ' ', string)
    string = re.sub(r'\s+', ' ', string)
    string = re.sub(r'-+', '-', string)
    string = string.strip()
    string = re.sub(f'([^^]){NOT_PROMPTABLE_MARKER}+', r'\1', string)
    return string

class LVISDataset(Dataset):
    def __init__(self, ann_file, img_dir, processor, pad_value):
        self.lvis = LVIS(ann_file)
        self.img_ids = sorted(self.lvis.imgs.keys())
        self.img_dir = img_dir
        self.processor = processor
        self.img_size = self.processor.image_processor.size['height']
        self.pad_value = pad_value

    def __len__(self):
        return len(self.img_ids)

    def __getitem__(self, idx):
        img_id = self.img_ids[idx]
        img_info = self.lvis.imgs[img_id]
        img_path = os.path.join(self.img_dir, os.path.basename(img_info['coco_url']))
        # Load image
        image = Image.open(img_path).convert("RGB")
        image = np.array(image).astype(np.float32) / 255.0  # scale to [0,1]

        # Determine square size
        max_side = max(image.shape[1], image.shape[0])

        # Create padded square with floating-point pad value
        pad_value = np.array(self.pad_value, dtype=np.float32)  # e.g., [0.5,0.5,0.5]
        padded_image = np.ones((max_side, max_side, 3), dtype=np.float32) * pad_value

        # Paste original image at top-left
        padded_image[:image.shape[0], :image.shape[1], :] = image

        # Convert back to PIL for resizing
        padded_image = Image.fromarray((padded_image * 255).astype(np.uint8))

        # Resize to target size
        resized_image = padded_image.resize((self.img_size, self.img_size), Image.Resampling.BILINEAR)

        # Process image
        pixel_values = self.processor.image_processor(
            images=resized_image,
            return_tensors="pt"
        )['pixel_values']
        return img_id, image, img_info['width'], img_info['height'], pixel_values


def collate_fn(batch):
    img_ids, images, widths, heights, pixel_values = zip(*batch)
    return list(img_ids), list(images), list(widths), list(heights), torch.cat(list(pixel_values), axis=0)


def main():
    parser = argparse.ArgumentParser(description="Evaluate OWLv2 on LVIS dataset")
    parser.add_argument("--dataset_dir", default="/path/to/lvis/dataset")
    parser.add_argument("--pad_value", type=float, default=0.5)
    parser.add_argument("--local_rank", default=int(os.getenv('LOCAL_RANK', -1)), type=int)
    parser.add_argument("--topk", type=int, default=300)
    parser.add_argument("--num_workers", type=int, default=4)
    args = parser.parse_args()

    torch.cuda.set_device(args.local_rank)

    dist.init_process_group(
        backend="nccl",
        init_method="env://",
        world_size=int(os.getenv("WORLD_SIZE", 1)),
        rank=int(os.getenv("RANK", 0)),
        device_id=torch.device(f'cuda:{args.local_rank}'),
    )
    rank = dist.get_rank()
    world_size = dist.get_world_size()

    print(f'Using Pad Value : {args.pad_value}')
    device = torch.device(f"cuda:{args.local_rank}" if args.local_rank >= 0 else "cuda")
    if rank == 0:
        print(f"Running evaluation on {world_size} GPUs, device={device}")

    processor = Owlv2Processor.from_pretrained("google/owlv2-base-patch16-ensemble",  use_fast=True)
    model = Owlv2ForObjectDetection.from_pretrained("google/owlv2-base-patch16-ensemble").to(device).eval()

    ann_file = os.path.join(args.dataset_dir, "lvis_v1_val.json")
    img_dir = os.path.join(args.dataset_dir, "val2017")

    dataset = LVISDataset(ann_file, img_dir, processor=processor, pad_value=args.pad_value)
    sampler = DistributedSampler(dataset, num_replicas=world_size, rank=rank, shuffle=False)
    dataloader = DataLoader(
        dataset,
        batch_size=1,
        sampler=sampler,
        collate_fn=collate_fn,
        num_workers=args.num_workers,
        pin_memory=True,
        persistent_workers=(args.num_workers > 0)
    )

    lvis_gt = dataset.lvis
    cats = sorted(lvis_gt.cats.items(), key=lambda x: x[0])
    class_names = [cat['name'] for _, cat in cats]

    texts_ens = []
    for template in PROMPT_TEMPLATES:
        texts_ens += [_canonicalize_string(template.format(name)) for name in class_names]

    with torch.no_grad():
        text_inputs = processor.tokenizer(
            texts_ens, padding=True, truncation=True, max_length=16, return_tensors="pt"
        ).to(device)
        text_outputs = model.owlv2.text_model(**text_inputs)
        text_embeds = model.owlv2.text_projection(text_outputs[1])
        text_embeds = text_embeds / torch.linalg.norm(text_embeds, ord=2, dim=-1, keepdim=True)
        input_ids = text_inputs['input_ids'].reshape(1, -1, text_inputs['input_ids'].shape[-1])
        query_mask = input_ids[..., 0] > 0

    print(f'RANK {rank}, Ready!')
    dist.barrier()

    raw_predictions = []
    progress_bar = tqdm(dataloader, desc="Evaluating") if rank == 0 else dataloader
    
    
    for n, batch in enumerate(progress_bar):
        img_ids, images, widths, heights, pixel_values = batch
        with torch.no_grad():
            num_patches_height = model.num_patches_height
            num_patches_width = model.num_patches_width

            vision_outputs = model.owlv2.vision_model(pixel_values=pixel_values.to(device))
            last_hidden_state = vision_outputs[0]
            image_embeds = model.owlv2.vision_model.post_layernorm(last_hidden_state)
            class_token_out = torch.broadcast_to(image_embeds[:, :1, :], image_embeds[:, :-1].shape)
            image_embeds = image_embeds[:, 1:, :] * class_token_out
            image_embeds = model.layer_norm(image_embeds)
            image_embeds = image_embeds.reshape(
                image_embeds.shape[0], num_patches_height, num_patches_width, image_embeds.shape[-1]
            )

            image_feats = image_embeds.view(image_embeds.shape[0], -1, image_embeds.shape[-1])
            (pred_logits, _) = model.class_predictor(image_feats, text_embeds, query_mask)
            pred_boxes = model.box_predictor(image_feats, image_embeds, False)

            num_templates = len(PROMPT_TEMPLATES)
            num_classes = len(class_names)
            scores = pred_logits.reshape(1, -1, num_templates, num_classes).mean(2)

            bsz, num_patches, num_classes = scores.shape
            k = min(args.topk, num_patches * num_classes)
            scores_flat = scores.view(bsz, -1)
            topk_scores, topk_inds = torch.topk(scores_flat, k, dim=1)

            patch_inds = topk_inds // num_classes
            label_inds = topk_inds % num_classes
            batch_idx = torch.arange(bsz, device=pred_boxes.device).unsqueeze(-1)
            selected_boxes = pred_boxes[batch_idx, patch_inds]

            raw_predictions.append([
                img_ids, widths, heights,
                topk_scores.cpu(), label_inds.cpu(), selected_boxes.cpu()
            ])
        
    torch.cuda.synchronize()

    predictions = []
    for img_ids, widths, heights, topk_scores_cpu, label_inds_cpu, selected_boxes_cpu in raw_predictions:
        image_id = img_ids[0]
        w, h = float(widths[0]), float(heights[0])
        scale = max(w, h)
        scores_np = topk_scores_cpu[0].numpy()
        labels_np = label_inds_cpu[0].numpy()
        boxes_np = selected_boxes_cpu[0].numpy()

        cx, cy, bw, bh = boxes_np[:, 0], boxes_np[:, 1], boxes_np[:, 2], boxes_np[:, 3]
        x, y = (cx - bw / 2) * scale, (cy - bh / 2) * scale
        width, height = bw * scale, bh * scale

        preds_img = [
            {
                "image_id": image_id,
                "category_id": cats[label][0],
                "bbox": [float(x[i]), float(y[i]), float(width[i]), float(height[i])],
                "score": float(scores_np[i]),
            }
            for i, label in enumerate(labels_np)
        ]
        predictions.extend(preds_img)

    print(f'RANK {rank}, Done!')
    all_predictions = [None] * world_size
    dist.all_gather_object(all_predictions, predictions)

    if rank == 0:
        full_predictions = [p for sublist in all_predictions for p in sublist]
        lvis_dt = LVISResults(lvis_gt, full_predictions)
        lvis_eval = LVISEval(lvis_gt, lvis_dt, iou_type='bbox')
        lvis_eval.evaluate()
        lvis_eval.accumulate()
        lvis_eval.summarize()
        lvis_eval.print_results()

    dist.destroy_process_group()


if __name__ == "__main__":
    main()

Commands:

# 0.5 padding:
torchrun --nproc-per-node=NUM_GPUS myscript.py --pad_value 0.5 --dataset_dir /path/to/lvis/

# 0.0 padding:
torchrun --nproc-per-node=NUM_GPUS myscript.py --pad_value 0.0 --dataset_dir /path/to/lvis/

Please prepare LVIS dataset beforehand with the following structure:

/path/to/lvis/
      ├── val2017
      │ ├── 000000062833.jpg
      │ └── ...
      └── lvis_v1_val.json

After Running the scripts, the following logs should be printed:

0.5 padding

Using Pad Value : 0.5
Running evaluation on 1 GPUs, device=cuda:0
RANK 0, Ready!
RANK 0, Done!
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=all] = 0.434
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=300 catIds=all] = 0.600
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=300 catIds=all] = 0.473
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=     s | maxDets=300 catIds=all] = 0.330
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=     m | maxDets=300 catIds=all] = 0.533
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=     l | maxDets=300 catIds=all] = 0.652
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=  r] = 0.403
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=  c] = 0.430
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=  f] = 0.451
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=all] = 0.563
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=     s | maxDets=300 catIds=all] = 0.406
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=     m | maxDets=300 catIds=all] = 0.672
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=     l | maxDets=300 catIds=all] = 0.805

0.0 padding

Using Pad Value : 0.0
Running evaluation on 1 GPUs, device=cuda:0
RANK 0, Ready!
RANK 0, Done!
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=all] = 0.440
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=300 catIds=all] = 0.602
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=300 catIds=all] = 0.482
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=     s | maxDets=300 catIds=all] = 0.333
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=     m | maxDets=300 catIds=all] = 0.540
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=     l | maxDets=300 catIds=all] = 0.664
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=  r] = 0.406
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=  c] = 0.438
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=  f] = 0.458
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=300 catIds=all] = 0.570
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=     s | maxDets=300 catIds=all] = 0.411
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=     m | maxDets=300 catIds=all] = 0.678
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=     l | maxDets=300 catIds=all] = 0.815

Reference:
[1] OWLv1 (Figure A4.)
[2] OWLv2 (Figure A3),
[3] OWLv2 original implementation, which is changed with this PR (scenic/projects/owl_vit/evaluator.py, line 158).

fixed padding value

Rocketknight1 · 2025-10-30T11:54:24Z

cc @yonigozlan

yonigozlan

Hello @gjamesgoenawan ! Thanks a lot for investigating this and providing sources! Ok to merge for me, just waiting on the CI and the slow CI to pass, as this might change some integration tests. If it does I'll push the new (correct) results directly on this PR

yonigozlan · 2025-11-03T21:28:59Z

run-slow: owlv2

github-actions · 2025-11-03T21:32:43Z

[For maintainers] Suggested jobs to run (before merge)

run-slow: owlv2

HuggingFaceDocBuilderDev · 2025-11-03T21:38:19Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

gjamesgoenawan · 2025-11-03T21:59:47Z

@yonigozlan Thanks for reviewing!

I think it would be helpful to mention some inference optimizations from the jax implementation (text embedding caching and prompt ensembling) as these aren’t immediately obvious from the current documentation. Including brief references or examples for these techniques would make it clearer for users.

yonigozlan · 2025-11-03T23:38:34Z

I think it would be helpful to mention some inference optimizations from the jax implementation (text embedding caching and prompt ensembling) as these aren’t immediately obvious from the current documentation. Including brief references or examples for these techniques would make it clearer for users.

Indeed, at least for text embedding caching, it seems there's no easy way to use it with the current API. I'm working on a refactor of vision models, and I'll add this to the list of to-dos.
Not sure about prompt ensembling though, this seems to be a technique used for training? Or am I misunderstanding something

gjamesgoenawan · 2025-11-04T02:45:01Z

Prompt Ensembling is explored in CLIP[1] in Section 3.1.4 & Figure 4.

TLDR; It is an inference technique that significantly improve zero-shot performance by averaging logits over multiple text-prompts.

You can find traces of this in the original JAX evaluator code and the code snippet I shared.

I define the number of prompt templates as n_prompts, which in this case has a value of 7. text_embeds has a shape of (n_prompts*n_classes,16,hidden_dim) and hence pred_logits will have a shape of (bs, n_objects, n_prompts*n_classes). This is then reshaped and averaged by:

scores = pred_logits.reshape(1, -1, num_templates, num_classes).mean(2)
# scores [bs, n_objects, n_classes]

Effectively averaging logits across all prompt templates.

Additionally, the top-k operations is a bit different from Transformers' implementation as well. Specifically this line in postprocessing which strictly enforce maximum of 1 detection per object proposal.

My implementation follow the original closely which doesnt have this restriction. Instead, it took the top-k from all logits, meaning 1 object proposal can result in more than 1 detection.

Without these two methods, evaluation performance differs significantly from the original.

[1] CLIP

yonigozlan · 2025-11-04T15:40:38Z

Thanks for the explanation @gjamesgoenawan, it does sound like some elements were overlooked when adding this model. Feel free to open PRs to add/fix these two features, I'll make sure to review them quickly.

* Update image_processing_owlv2_fast.py fixed padding value * fixed padding value * Change padding constant value from 0.5 to 0.0 * Fixed missed padding value in modular_owlv2.py --------- Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>

* remove attributes and add all missing sub processors to their auto classes * remove all mentions of .attributes * cleanup * fix processor tests * fix modular * remove last attributes * fixup * fixes after merge * fix wrong tokenizer in auto florence2 * fix missing audio_processor + nits * Override __init__ in NewProcessor and change hf-internal-testing-repo (temporarily) * fix auto tokenizer test * add init to markup_lm * update CustomProcessor in custom_processing * remove print * nit * fix test modeling owlv2 * fix test_processing_layoutxlm * Fix owlv2, wav2vec2, markuplm, voxtral issues * add support for loading and saving multiple tokenizer natively * remove exclude_attributes from save_pretrained * Run slow v2 (#41914) * Super * Super * Super * Super --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Fix `detectron2` installation in docker files (#41975) * detectron2 - part 1 * detectron2 - part 2 --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Fix `autoawq[kernels]` installation in quantization docker file (#41978) fix autoawq[kernels] Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * add support for saving encoder only so any parakeet model can be loaded for inference (#41969) * add support for saving encoder only so any decoder model can be loaded Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> * use convolution_bias * convert modular * convolution_bias in convertion script --------- Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Co-authored-by: Eustache Le Bihan <eulebihan@gmail.com> Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com> * Use indices as position_ids in modernebert (#41789) * Use indices as position_ids in modernebert * Move position_ids init to the branch * test tensor parallel: make tests for dense model more robust (#41968) * make test forward and backward more robust * refactor compile part of test tensor parallel * linting * pass rank around instead of calling it over and over * Run slow v2 (#41914) * Super * Super * Super * Super --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Fix `detectron2` installation in docker files (#41975) * detectron2 - part 1 * detectron2 - part 2 --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Fix `autoawq[kernels]` installation in quantization docker file (#41978) fix autoawq[kernels] Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * add support for saving encoder only so any parakeet model can be loaded for inference (#41969) * add support for saving encoder only so any decoder model can be loaded Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> * use convolution_bias * convert modular * convolution_bias in convertion script --------- Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Co-authored-by: Eustache Le Bihan <eulebihan@gmail.com> Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com> --------- Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Co-authored-by: Yih-Dar <2521628+ydshieh@users.noreply.github.com> Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> Co-authored-by: Nithin Rao <nithinrao.koluguri@gmail.com> Co-authored-by: Eustache Le Bihan <eulebihan@gmail.com> Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com> * fix: dict[RopeParameters] to dict[str, RopeParameters] (#41963) * docs: add continuous batching page (#41847) * docs: add continuous batching page * docs(cb): add `generate_batch` example * docs(cb): add `opentelemtry` and `serving` section * feat: add `TODO` note about opentelemetry dependency * docs(cb): add supported features * docs(cb): add unsupported features * docs(cb): add `ContinuousBatchingManager` example * docs(cb): x reference CB in optimizing inference * Fix `torchcodec` version in quantization docker file (#41988) check Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * [kernels] Add Tests & CI for kernels (#41765) * first commit * add tests * add kernel config * add more tests * add ci * small fix * change branch name * update tests * nit * change test name * revert jobs * addressing review * reenable all jobs * address second review * Move the Mi355 to regular docker (#41989) * Move the Mi355 to regular docker * Disable gfx950 compilation for FA on AMD * More data in benchmarking (#41848) * Reduce scope of cross-generate * Rm generate_sall configs * Workflow benchmarks more * Prevent crash when FA is not installed * fix (CI): Refactor SSH runners (#41991) * Change ssh runner type * Add wait step to SSH runner workflow * Rename wait step to wait2 in ssh-runner.yml * Remove wait step from ssh-runner.yml Removed the wait step from the SSH runner workflow. * Update runner type for single GPU A10 instance * Update SSH runner version to 1.90.3 * Add sha256sum to ssh-runner workflow * Update runner type and remove unused steps * fix 3 failed test cases for video_llama_3 model on Intel XPU (#41931) * fix 3 failed test cases for video_llama_3 model on Intel XPU Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com> * update Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com> * adjust format Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com> * update code Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com> --------- Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com> * Integrate colqwen2.5 using colqwen2 modelling code (#40600) * adding option for 2.5 * minor - arg in conversion script * getting started on modelling.py * minor - shouldve been using modular * adressing comments + fixing datatype/device _get method * minor * commiting suggestion Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com> * docs + first test * ruff fix * minor fix * ruff fix * model fix * model fix * fine-grained check, with a hardcoded score from the original Hf implementation. * minor ruff * update tests values with CI hardware * adding 2.5 to conversion script * Apply style fixes --------- Co-authored-by: Sahil Kabir <sahilkabir@Sahils-MacBook-Pro.local> Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com> Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> * Fixed wrong padding value in OWLv2 (#41938) * Update image_processing_owlv2_fast.py fixed padding value * fixed padding value * Change padding constant value from 0.5 to 0.0 * Fixed missed padding value in modular_owlv2.py --------- Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com> * Fix `run slow v2`: empty report when there is only one model (#42002) fix Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * [kernels] change import time in KernelConfig (#42004) * change import time * style * DOC Fix typo in argument name: pseudoquant (#41994) The correct argument name is pseudoquantization. Since there is no error on passing wrong arguments name (which is arguably an anti-pattern), this is difficult for users to debug. * Fix `torch+deepspeed` docker file (#41985) * fix * delete --------- Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> * Correct syntax error in trainer.md (#42001) A comma is missing between two parameters in the signature of compute_loss function. * Reduce the number of benchmark in the CI (#42008) Changed how benchmark cfgs are chosen * Fix continuous batching tests (#42012) * Fix continuous batching tests * make fixup * add back `logging_dir` (#42013) * add back * Apply style fixes --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> * Fix issue with from pretrained and kwargs in image processors (#41997) * accept kwargs in image proc from_pretrained * only use kwargs that are in cls.valid_kwargs * remove specific logic for _from_auto * add image_seq_length to Images_kwargs for backward compatibility * fix missing image kwargs in pix2struct * Fix default image_rows and image_cols initialization in Idefics3 and SmolVLM processors (#41871) * Fix default image_rows and image_cols initialization in Idefics3 and SmolVLM processors * Fix default initialization of image_rows and image_cols in Idefics3 and SmolVLM processors * Add GLPNImageProcessorFast (#41725) * Add GLPNImageProcessorFast for torch backend * Address review feedback - Simplified to_dict() method - Keep tensors as torch instead of converting to numpy for heterogeneous shapes - Removed unnecessary shape guards in post_process_depth_estimation - Improved variable names (tgt -> target_size, d -> resized) - Removed unnecessary GLPNImageProcessorKwargs class * Address review feedback - Simplified to_dict() method - Keep tensors as torch instead of converting to numpy for heterogeneous shapes - Removed unnecessary shape guards in post_process_depth_estimation - Improved variable names (tgt -> target_size, d -> resized) - Removed unnecessary GLPNImageProcessorKwargs class * commits after 2nd review * Address all review feedback and add explicit batched test - Simplified to_dict() with descriptive variable names (d->output_dict) - Fixed resize operation: changed from crop to proper resize with interpolation - Added padding for heterogeneous batch shapes in both slow and fast processors - Fused rescale and normalize operations for efficiency - Improved all variable names (tgt->target_size, d->depth_4d->resized) - Added GLPNImageProcessorKwargs class in slow processor and imported in fast - Renamed test_equivalence_slow_fast to test_slow_fast_equivalence - Added explicit test_slow_fast_equivalence_batched test - All 20 tests passing * using padding from utils * simplify glpn image processor fast * fix docstring --------- Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co> Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com> * add fuyu fast image processors (#41817) * added fast processor for fuyu (#36978) * updated docs for fuyu model (#36978) * updated test_image_processing and image_processing_fuyu_fast * updated fuyu.md and image_processing_fuyu_fast (#36978) * updated test_image_processing_fuyu (#36978) * formatted image_processing_fuyu_fast and test_image_processing_fuyu (#36978) * updated tests and fuyu fast image processing (#36978) * Merge branch 'fuyu-fast-image-processors' of https://github.com/DeXtAr47-oss/transformers into fuyu-fast-image-processors * fixed format (#36978) * formatted files (#36978) * formatted files * revert unnecessary changes * clean up and process by group --------- Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co> * [kernels] Fix XPU layernorm kernel (#41583) * fix * add comment * better fix * style * Update src/transformers/modeling_utils.py Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com> --------- Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com> * [v5] Deprecate Text2Text and related pipelines (#41996) * Deprecate Text2Text and related pipelines * Try a restructure * make fixup * logging -> logger * [FPQuant] MXFP8 and MXFP4 backwards support (#41897) * FP-Quant backwards * fp-quant v0.3.0 docker * availability version bump * fp_quant==0.3.1 * fp_quant v0.3.2 * add working auto_docstring for processors * add auto_docstring to processors first part * add auto_docstring to processors part 2 * modifs after review * fully working auto_docstring and check_docstring with placeholder docstrings * Working check_docstrings for Typed dicts * Add recurring processor args to auto_docstring and add support for removing redundant docstring and placeholders * replace placeholders with real docstrings * fix copies * fixup * remove unwanted changes * fix unprotected imports * Fix unprotected imports * fix unprotected imports * Add __call__ to all docs of processors * nits docs --------- Signed-off-by: nithinraok <nithinrao.koluguri@gmail.com> Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com> Co-authored-by: Yih-Dar <2521628+ydshieh@users.noreply.github.com> Co-authored-by: ydshieh <ydshieh@users.noreply.github.com> Co-authored-by: Nithin Rao <nithinrao.koluguri@gmail.com> Co-authored-by: Eustache Le Bihan <eulebihan@gmail.com> Co-authored-by: eustlb <94853470+eustlb@users.noreply.github.com> Co-authored-by: Rémi Ouazan <83456801+remi-or@users.noreply.github.com> Co-authored-by: Ferdinand Mom <47445085+3outeille@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> Co-authored-by: Luc Georges <McPatate@users.noreply.github.com> Co-authored-by: Mohamed Mekkouri <93391238+MekkCyber@users.noreply.github.com> Co-authored-by: Guillaume LEGENDRE <glegendre01@gmail.com> Co-authored-by: kaixuanliu <kaixuan.liu@intel.com> Co-authored-by: Sahil Kabir <66221472+sahil-kabir@users.noreply.github.com> Co-authored-by: Sahil Kabir <sahilkabir@Sahils-MacBook-Pro.local> Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com> Co-authored-by: James <67161633+gjamesgoenawan@users.noreply.github.com> Co-authored-by: Benjamin Bossan <BenjaminBossan@users.noreply.github.com> Co-authored-by: Yacklin Wong <139425274+Yacklin@users.noreply.github.com> Co-authored-by: Matt <Rocketknight1@users.noreply.github.com> Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com> Co-authored-by: MilkClouds <claude@maum.ai> Co-authored-by: ARAVINDHAN T <arvindhant01@gmail.com> Co-authored-by: Pritam Das <79273068+DeXtAr47-oss@users.noreply.github.com> Co-authored-by: Andrei Panferov <andrei@panferov.org>

gjamesgoenawan added 4 commits October 29, 2025 22:01

Update image_processing_owlv2_fast.py

d1427f2

fixed padding value

fixed padding value

1a2cccf

Change padding constant value from 0.5 to 0.0

ff413a7

Fixed missed padding value in modular_owlv2.py

0eade78

gjamesgoenawan changed the title ~~Fixed padding value of OWLv2~~ Fixed missed padding value in modular_owlv2 Oct 29, 2025

gjamesgoenawan changed the title ~~Fixed missed padding value in modular_owlv2~~ Fixed wrong padding value in OWLv2 Oct 29, 2025

Merge branch 'main' into main

d310be4

yonigozlan approved these changes Nov 3, 2025

View reviewed changes

yonigozlan added the run-slow label Nov 3, 2025

Merge branch 'main' into main

e3481db

yonigozlan merged commit 64397a8 into huggingface:main Nov 3, 2025
15 checks passed

yonigozlan changed the title ~~Fixed wrong padding value in OWLv2~~ 🚨Fixed wrong padding value in OWLv2 Nov 4, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

🚨Fixed wrong padding value in OWLv2 #41938

🚨Fixed wrong padding value in OWLv2 #41938

gjamesgoenawan commented Oct 29, 2025 •

edited

Loading

Uh oh!

Rocketknight1 commented Oct 30, 2025

Uh oh!

yonigozlan left a comment

Uh oh!

yonigozlan commented Nov 3, 2025

Uh oh!

github-actions bot commented Nov 3, 2025

Uh oh!

HuggingFaceDocBuilderDev commented Nov 3, 2025

Uh oh!

gjamesgoenawan commented Nov 3, 2025 •

edited

Loading

Uh oh!

yonigozlan commented Nov 3, 2025

Uh oh!

Uh oh!

gjamesgoenawan commented Nov 4, 2025

Uh oh!

yonigozlan commented Nov 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

🚨Fixed wrong padding value in OWLv2 #41938

🚨Fixed wrong padding value in OWLv2 #41938

Conversation

gjamesgoenawan commented Oct 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Reproducing the results

0.5 padding

0.0 padding

Uh oh!

Rocketknight1 commented Oct 30, 2025

Uh oh!

yonigozlan left a comment

Choose a reason for hiding this comment

Uh oh!

yonigozlan commented Nov 3, 2025

Uh oh!

github-actions bot commented Nov 3, 2025

Uh oh!

HuggingFaceDocBuilderDev commented Nov 3, 2025

Uh oh!

gjamesgoenawan commented Nov 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

yonigozlan commented Nov 3, 2025

Uh oh!

Uh oh!

gjamesgoenawan commented Nov 4, 2025

Uh oh!

yonigozlan commented Nov 4, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

gjamesgoenawan commented Oct 29, 2025 •

edited

Loading

gjamesgoenawan commented Nov 3, 2025 •

edited

Loading