SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

[Conditional, Deformable DETR] Add postprocessing methods (#19709)

Browse Source 
* Add postprocessing methods

* Update docs

* Add fix

* Add test

* Add test for deformable detr postprocessing

* Add post processing methods for segmentation

* Update code examples

* Add post_process to make the pipeline work

* Apply updates

Co-authored-by: Niels Rogge <nielsrogge@Nielss-MacBook-Pro.local>
This commit is contained in:
NielsRogge
2022-10-31 08:28:44 +01:00
committed by GitHub
parent 2e35bac4e7
commit 0b294c2334
11 changed files with 513 additions and 535 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
docs/source/en/model_doc/conditional_detr.mdx
View File

@@ -37,9 +37,10 @@ This model was contributed by [DepuMeng](https://huggingface.co/DepuMeng). The o
[[autodoc]] ConditionalDetrFeatureExtractor
- __call__
- pad_and_create_pixel_mask
- post_process
- post_process_segmentation
- post_process_panoptic
- post_process_object_detection
- post_process_instance_segmentation
- post_process_semantic_segmentation
- post_process_panoptic_segmentation
## ConditionalDetrModel

4
docs/source/en/model_doc/deformable_detr.mdx
View File

@@ -38,9 +38,7 @@ This model was contributed by [nielsr](https://huggingface.co/nielsr). The origi
[[autodoc]] DeformableDetrFeatureExtractor
- __call__
- pad_and_create_pixel_mask
- post_process
- post_process_segmentation
- post_process_panoptic
- post_process_object_detection
## DeformableDetrConfig

638
src/transformers/models/conditional_detr/feature_extraction_conditional_detr.py
View File

@@ -14,11 +14,9 @@
# limitations under the License.
"""Feature extractor class for Conditional DETR."""
import io
import pathlib
import warnings
from collections import defaultdict
from typing import Dict, List, Optional, Union
from typing import Dict, List, Optional, Set, Tuple, Union
import numpy as np
from PIL import Image
@@ -104,21 +102,157 @@ def rgb_to_id(color):
return int(color[0] + 256 * color[1] + 256 * 256 * color[2])
# Copied from transformers.models.detr.feature_extraction_detr.id_to_rgb
def id_to_rgb(id_map):
if isinstance(id_map, np.ndarray):
id_map_copy = id_map.copy()
rgb_shape = tuple(list(id_map.shape) + [3])
rgb_map = np.zeros(rgb_shape, dtype=np.uint8)
for i in range(3):
rgb_map[..., i] = id_map_copy % 256
id_map_copy //= 256
return rgb_map
color = []
for _ in range(3):
color.append(id_map % 256)
id_map //= 256
return color
# Copied from transformers.models.detr.feature_extraction_detr.binary_mask_to_rle
def binary_mask_to_rle(mask):
"""
Args:
Converts given binary mask of shape (height, width) to the run-length encoding (RLE) format.
mask (`torch.Tensor` or `numpy.array`):
A binary mask tensor of shape `(height, width)` where 0 denotes background and 1 denotes the target
segment_id or class_id.
Returns:
`List`: Run-length encoded list of the binary mask. Refer to COCO API for more information about the RLE
format.
"""
if is_torch_tensor(mask):
mask = mask.numpy()
pixels = mask.flatten()
pixels = np.concatenate([[0], pixels, [0]])
runs = np.where(pixels[1:] != pixels[:-1])[0] + 1
runs[1::2] -= runs[::2]
return [x for x in runs]
# Copied from transformers.models.detr.feature_extraction_detr.convert_segmentation_to_rle
def convert_segmentation_to_rle(segmentation):
"""
Converts given segmentation map of shape (height, width) to the run-length encoding (RLE) format.
Args:
segmentation (`torch.Tensor` or `numpy.array`):
A segmentation map of shape `(height, width)` where each value denotes a segment or class id.
Returns:
`List[List]`: A list of lists, where each list is the run-length encoding of a segment / class id.
"""
segment_ids = torch.unique(segmentation)
run_length_encodings = []
for idx in segment_ids:
mask = torch.where(segmentation == idx, 1, 0)
rle = binary_mask_to_rle(mask)
run_length_encodings.append(rle)
return run_length_encodings
# Copied from transformers.models.detr.feature_extraction_detr.remove_low_and_no_objects
def remove_low_and_no_objects(masks, scores, labels, object_mask_threshold, num_labels):
"""
Binarize the given masks using `object_mask_threshold`, it returns the associated values of `masks`, `scores` and
`labels`.
Args:
masks (`torch.Tensor`):
A tensor of shape `(num_queries, height, width)`.
scores (`torch.Tensor`):
A tensor of shape `(num_queries)`.
labels (`torch.Tensor`):
A tensor of shape `(num_queries)`.
object_mask_threshold (`float`):
A number between 0 and 1 used to binarize the masks.
Raises:
`ValueError`: Raised when the first dimension doesn't match in all input tensors.
Returns:
`Tuple[`torch.Tensor`, `torch.Tensor`, `torch.Tensor`]`: The `masks`, `scores` and `labels` without the region
< `object_mask_threshold`.
"""
if not (masks.shape[0] == scores.shape[0] == labels.shape[0]):
raise ValueError("mask, scores and labels must have the same shape!")
to_keep = labels.ne(num_labels) & (scores > object_mask_threshold)
return masks[to_keep], scores[to_keep], labels[to_keep]
# Copied from transformers.models.detr.feature_extraction_detr.check_segment_validity
def check_segment_validity(mask_labels, mask_probs, k, mask_threshold=0.5, overlap_mask_area_threshold=0.8):
# Get the mask associated with the k class
mask_k = mask_labels == k
mask_k_area = mask_k.sum()
# Compute the area of all the stuff in query k
original_area = (mask_probs[k] >= mask_threshold).sum()
mask_exists = mask_k_area > 0 and original_area > 0
# Eliminate disconnected tiny segments
if mask_exists:
area_ratio = mask_k_area / original_area
if not area_ratio.item() > overlap_mask_area_threshold:
mask_exists = False
return mask_exists, mask_k
# Copied from transformers.models.detr.feature_extraction_detr.compute_segments
def compute_segments(
mask_probs,
pred_scores,
pred_labels,
mask_threshold: float = 0.5,
overlap_mask_area_threshold: float = 0.8,
label_ids_to_fuse: Optional[Set[int]] = None,
target_size: Tuple[int, int] = None,
):
height = mask_probs.shape[1] if target_size is None else target_size[0]
width = mask_probs.shape[2] if target_size is None else target_size[1]
segmentation = torch.zeros((height, width), dtype=torch.int32, device=mask_probs.device)
segments: List[Dict] = []
if target_size is not None:
mask_probs = nn.functional.interpolate(
mask_probs.unsqueeze(0), size=target_size, mode="bilinear", align_corners=False
)[0]
current_segment_id = 0
# Weigh each mask by its prediction score
mask_probs *= pred_scores.view(-1, 1, 1)
mask_labels = mask_probs.argmax(0) # [height, width]
# Keep track of instances of each class
stuff_memory_list: Dict[str, int] = {}
for k in range(pred_labels.shape[0]):
pred_class = pred_labels[k].item()
should_fuse = pred_class in label_ids_to_fuse
# Check if mask exists and large enough to be a segment
mask_exists, mask_k = check_segment_validity(
mask_labels, mask_probs, k, mask_threshold, overlap_mask_area_threshold
)
if mask_exists:
if pred_class in stuff_memory_list:
current_segment_id = stuff_memory_list[pred_class]
else:
current_segment_id += 1
# Add current object segment to final segmentation map
segmentation[mask_k] = current_segment_id
segment_score = round(pred_scores[k].item(), 6)
segments.append(
{
"id": current_segment_id,
"label_id": pred_class,
"was_fused": should_fuse,
"score": segment_score,
}
)
if should_fuse:
stuff_memory_list[pred_class] = current_segment_id
return segmentation, segments
class ConditionalDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
@@ -128,7 +262,6 @@ class ConditionalDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtrac
This feature extractor inherits from [`FeatureExtractionMixin`] which contains most of the main methods. Users
should refer to this superclass for more information regarding those methods.
Args:
format (`str`, *optional*, defaults to `"coco_detection"`):
Data format of the annotations. One of "coco_detection" or "coco_panoptic".
@@ -691,25 +824,27 @@ class ConditionalDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtrac
return encoded_inputs
# POSTPROCESSING METHODS
# inspired by https://github.com/Atten4Vis/conditionalDETR/blob/master/models/conditional_detr.py#L258
def post_process(self, outputs, target_sizes):
"""
Args:
Converts the output of [`ConditionalDetrForObjectDetection`] into the format expected by the COCO api. Only
supports PyTorch.
Args:
outputs ([`ConditionalDetrObjectDetectionOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
image size (before any data augmentation). For visualization, this should be the image size after data
augment, but before padding.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_object_detection`",
FutureWarning,
)
out_logits, out_bbox = outputs.logits, outputs.pred_boxes
if len(out_logits) != len(target_sizes):
@@ -734,240 +869,283 @@ class ConditionalDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtrac
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_segmentation with Detr->ConditionalDetr
def post_process_segmentation(self, outputs, target_sizes, threshold=0.9, mask_threshold=0.5):
# Copied from transformers.models.deformable_detr.feature_extraction_deformable_detr.DeformableDetrFeatureExtractor.post_process_object_detection with DeformableDetr->ConditionalDetr
def post_process_object_detection(
self, outputs, threshold: float = 0.5, target_sizes: Union[TensorType, List[Tuple]] = None
):
"""
Converts the output of [`ConditionalDetrForSegmentation`] into image segmentation predictions. Only supports
PyTorch.
Parameters:
outputs ([`ConditionalDetrSegmentationOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction.
threshold (`float`, *optional*, defaults to 0.9):
Threshold to use to filter out queries.
mask_threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, and masks for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process_segmentation` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_semantic_segmentation`.",
FutureWarning,
)
out_logits, raw_masks = outputs.logits, outputs.pred_masks
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, size in zip(out_logits, raw_masks, target_sizes):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_masks = (cur_masks.sigmoid() > mask_threshold) * 1
predictions = {"scores": cur_scores, "labels": cur_classes, "masks": cur_masks}
preds.append(predictions)
return preds
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_instance with Detr->ConditionalDetr
def post_process_instance(self, results, outputs, orig_target_sizes, max_target_sizes, threshold=0.5):
"""
Converts the output of [`ConditionalDetrForSegmentation`] into actual instance segmentation predictions. Only
Converts the output of [`ConditionalDetrForObjectDetection`] into the format expected by the COCO api. Only
supports PyTorch.
Args:
results (`List[Dict]`):
Results list obtained by [`~ConditionalDetrFeatureExtractor.post_process`], to which "masks" results
will be added.
outputs ([`ConditionalDetrSegmentationOutput`]):
outputs ([`DetrObjectDetectionOutput`]):
Raw outputs of the model.
orig_target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
image size (before any data augmentation).
max_target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the maximum size (h, w) of each image of the batch. For evaluation, this must be the
original image size (before any data augmentation).
threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
threshold (`float`, *optional*):
Score threshold to keep object detection predictions.
target_sizes (`torch.Tensor` or `List[Tuple[int, int]]`, *optional*, defaults to `None`):
Tensor of shape `(batch_size, 2)` or list of tuples (`Tuple[int, int]`) containing the target size
(height, width) of each image in the batch. If left to None, predictions will not be resized.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, boxes and masks for an
image in the batch as predicted by the model.
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process_instance` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_instance_segmentation`.",
FutureWarning,
out_logits, out_bbox = outputs.logits, outputs.pred_boxes
if target_sizes is not None:
if len(out_logits) != len(target_sizes):
raise ValueError(
"Make sure that you pass in as many target sizes as the batch dimension of the logits"
)
if len(orig_target_sizes) != len(max_target_sizes):
raise ValueError("Make sure to pass in as many orig_target_sizes as max_target_sizes")
max_h, max_w = max_target_sizes.max(0)[0].tolist()
outputs_masks = outputs.pred_masks.squeeze(2)
outputs_masks = nn.functional.interpolate(
outputs_masks, size=(max_h, max_w), mode="bilinear", align_corners=False
)
outputs_masks = (outputs_masks.sigmoid() > threshold).cpu()
prob = out_logits.sigmoid()
topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), 100, dim=1)
scores = topk_values
topk_boxes = topk_indexes // out_logits.shape[2]
labels = topk_indexes % out_logits.shape[2]
boxes = center_to_corners_format(out_bbox)
boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
for i, (cur_mask, t, tt) in enumerate(zip(outputs_masks, max_target_sizes, orig_target_sizes)):
img_h, img_w = t[0], t[1]
results[i]["masks"] = cur_mask[:, :img_h, :img_w].unsqueeze(1)
results[i]["masks"] = nn.functional.interpolate(
results[i]["masks"].float(), size=tuple(tt.tolist()), mode="nearest"
).byte()
# and from relative [0, 1] to absolute [0, height] coordinates
if isinstance(target_sizes, List):
img_h = torch.Tensor([i[0] for i in target_sizes])
img_w = torch.Tensor([i[1] for i in target_sizes])
else:
img_h, img_w = target_sizes.unbind(1)
scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
boxes = boxes * scale_fct[:, None, :]
results = []
for s, l, b in zip(scores, labels, boxes):
score = s[s > threshold]
label = l[s > threshold]
box = b[s > threshold]
results.append({"scores": score, "labels": label, "boxes": box})
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_panoptic with Detr->ConditionalDetr
def post_process_panoptic(self, outputs, processed_sizes, target_sizes=None, is_thing_map=None, threshold=0.85):
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_semantic_segmentation with Detr->ConditionalDetr
def post_process_semantic_segmentation(self, outputs, target_sizes: List[Tuple[int, int]] = None):
"""
Converts the output of [`ConditionalDetrForSegmentation`] into actual panoptic predictions. Only supports
Converts the output of [`ConditionalDetrForSegmentation`] into semantic segmentation maps. Only supports
PyTorch.
Parameters:
outputs ([`ConditionalDetrSegmentationOutput`]):
Args:
outputs ([`ConditionalDetrForSegmentation`]):
Raw outputs of the model.
processed_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) containing the size (h, w) of each image of the batch, i.e. the size after data
augmentation but before batching.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`, *optional*):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction. If left to
None, it will default to the `processed_sizes`.
is_thing_map (`torch.Tensor` of shape `(batch_size, 2)`, *optional*):
Dictionary mapping class indices to either True or False, depending on whether or not they are a thing.
If not set, defaults to the `is_thing_map` of COCO panoptic.
threshold (`float`, *optional*, defaults to 0.85):
Threshold to use to filter out queries.
target_sizes (`List[Tuple[int, int]]`, *optional*, defaults to `None`):
A list of tuples (`Tuple[int, int]`) containing the target size (height, width) of each image in the
batch. If left to None, predictions will not be resized.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing a PNG string and segments_info values for
an image in the batch as predicted by the model.
`List[torch.Tensor]`:
A list of length `batch_size`, where each item is a semantic segmentation map of shape (height, width)
corresponding to the target_sizes entry (if `target_sizes` is specified). Each entry of each
`torch.Tensor` correspond to a semantic class id.
"""
warnings.warn(
"`post_process_panoptic is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_panoptic_segmentation`.",
FutureWarning,
)
if target_sizes is None:
target_sizes = processed_sizes
if len(processed_sizes) != len(target_sizes):
raise ValueError("Make sure to pass in as many processed_sizes as target_sizes")
class_queries_logits = outputs.logits # [batch_size, num_queries, num_classes+1]
masks_queries_logits = outputs.pred_masks # [batch_size, num_queries, height, width]
if is_thing_map is None:
# default to is_thing_map of COCO panoptic
is_thing_map = {i: i <= 90 for i in range(201)}
# Remove the null class `[..., :-1]`
masks_classes = class_queries_logits.softmax(dim=-1)[..., :-1]
masks_probs = masks_queries_logits.sigmoid() # [batch_size, num_queries, height, width]
out_logits, raw_masks, raw_boxes = outputs.logits, outputs.pred_masks, outputs.pred_boxes
if not len(out_logits) == len(raw_masks) == len(target_sizes):
# Semantic segmentation logits of shape (batch_size, num_classes, height, width)
segmentation = torch.einsum("bqc, bqhw -> bchw", masks_classes, masks_probs)
batch_size = class_queries_logits.shape[0]
# Resize logits and compute semantic segmentation maps
if target_sizes is not None:
if batch_size != len(target_sizes):
raise ValueError(
"Make sure that you pass in as many target sizes as the batch dimension of the logits and masks"
"Make sure that you pass in as many target sizes as the batch dimension of the logits"
)
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, cur_boxes, size, target_size in zip(
out_logits, raw_masks, raw_boxes, processed_sizes, target_sizes
):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_boxes = center_to_corners_format(cur_boxes[keep])
h, w = cur_masks.shape[-2:]
if len(cur_boxes) != len(cur_classes):
raise ValueError("Not as many boxes as there are classes")
# It may be that we have several predicted masks for the same stuff class.
# In the following, we track the list of masks ids for each stuff class (they are merged later on)
cur_masks = cur_masks.flatten(1)
stuff_equiv_classes = defaultdict(lambda: [])
for k, label in enumerate(cur_classes):
if not is_thing_map[label.item()]:
stuff_equiv_classes[label.item()].append(k)
def get_ids_area(masks, scores, dedup=False):
# This helper function creates the final panoptic segmentation image
# It also returns the area of the masks that appears on the image
m_id = masks.transpose(0, 1).softmax(-1)
if m_id.shape[-1] == 0:
# We didn't detect any mask :(
m_id = torch.zeros((h, w), dtype=torch.long, device=m_id.device)
else:
m_id = m_id.argmax(-1).view(h, w)
if dedup:
# Merge the masks corresponding to the same stuff class
for equiv in stuff_equiv_classes.values():
if len(equiv) > 1:
for eq_id in equiv:
m_id.masked_fill_(m_id.eq(eq_id), equiv[0])
final_h, final_w = to_tuple(target_size)
seg_img = Image.fromarray(id_to_rgb(m_id.view(h, w).cpu().numpy()))
seg_img = seg_img.resize(size=(final_w, final_h), resample=Image.NEAREST)
np_seg_img = torch.ByteTensor(torch.ByteStorage.from_buffer(seg_img.tobytes()))
np_seg_img = np_seg_img.view(final_h, final_w, 3)
np_seg_img = np_seg_img.numpy()
m_id = torch.from_numpy(rgb_to_id(np_seg_img))
area = []
for i in range(len(scores)):
area.append(m_id.eq(i).sum().item())
return area, seg_img
area, seg_img = get_ids_area(cur_masks, cur_scores, dedup=True)
if cur_classes.numel() > 0:
# We know filter empty masks as long as we find some
while True:
filtered_small = torch.as_tensor(
[area[i] <= 4 for i, c in enumerate(cur_classes)], dtype=torch.bool, device=keep.device
semantic_segmentation = []
for idx in range(batch_size):
resized_logits = nn.functional.interpolate(
segmentation[idx].unsqueeze(dim=0), size=target_sizes[idx], mode="bilinear", align_corners=False
)
if filtered_small.any().item():
cur_scores = cur_scores[~filtered_small]
cur_classes = cur_classes[~filtered_small]
cur_masks = cur_masks[~filtered_small]
area, seg_img = get_ids_area(cur_masks, cur_scores)
semantic_map = resized_logits[0].argmax(dim=0)
semantic_segmentation.append(semantic_map)
else:
break
semantic_segmentation = segmentation.argmax(dim=1)
semantic_segmentation = [semantic_segmentation[i] for i in range(semantic_segmentation.shape[0])]
else:
cur_classes = torch.ones(1, dtype=torch.long, device=cur_classes.device)
return semantic_segmentation
segments_info = []
for i, a in enumerate(area):
cat = cur_classes[i].item()
segments_info.append({"id": i, "isthing": is_thing_map[cat], "category_id": cat, "area": a})
del cur_classes
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_instance_segmentation with Detr->ConditionalDetr
def post_process_instance_segmentation(
self,
outputs,
threshold: float = 0.5,
mask_threshold: float = 0.5,
overlap_mask_area_threshold: float = 0.8,
target_sizes: Optional[List[Tuple[int, int]]] = None,
return_coco_annotation: Optional[bool] = False,
) -> List[Dict]:
"""
Converts the output of [`ConditionalDetrForSegmentation`] into instance segmentation predictions. Only supports
PyTorch.
with io.BytesIO() as out:
seg_img.save(out, format="PNG")
predictions = {"png_string": out.getvalue(), "segments_info": segments_info}
preds.append(predictions)
return preds
Args:
outputs ([`ConditionalDetrForSegmentation`]):
Raw outputs of the model.
threshold (`float`, *optional*, defaults to 0.5):
The probability score threshold to keep predicted instance masks.
mask_threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
overlap_mask_area_threshold (`float`, *optional*, defaults to 0.8):
The overlap mask area threshold to merge or discard small disconnected parts within each binary
instance mask.
target_sizes (`List[Tuple]`, *optional*):
List of length (batch_size), where each list item (`Tuple[int, int]]`) corresponds to the requested
final size (height, width) of each prediction. If left to None, predictions will not be resized.
return_coco_annotation (`bool`, *optional*):
Defaults to `False`. If set to `True`, segmentation maps are returned in COCO run-length encoding (RLE)
format.
Returns:
`List[Dict]`: A list of dictionaries, one per image, each dictionary containing two keys:
- **segmentation** -- A tensor of shape `(height, width)` where each pixel represents a `segment_id` or
`List[List]` run-length encoding (RLE) of the segmentation map if return_coco_annotation is set to
`True`. Set to `None` if no mask if found above `threshold`.
- **segments_info** -- A dictionary that contains additional information on each segment.
- **id** -- An integer representing the `segment_id`.
- **label_id** -- An integer representing the label / semantic class id corresponding to `segment_id`.
- **score** -- Prediction score of segment with `segment_id`.
"""
class_queries_logits = outputs.logits # [batch_size, num_queries, num_classes+1]
masks_queries_logits = outputs.pred_masks # [batch_size, num_queries, height, width]
batch_size = class_queries_logits.shape[0]
num_labels = class_queries_logits.shape[-1] - 1
mask_probs = masks_queries_logits.sigmoid() # [batch_size, num_queries, height, width]
# Predicted label and score of each query (batch_size, num_queries)
pred_scores, pred_labels = nn.functional.softmax(class_queries_logits, dim=-1).max(-1)
# Loop over items in batch size
results: List[Dict[str, TensorType]] = []
for i in range(batch_size):
mask_probs_item, pred_scores_item, pred_labels_item = remove_low_and_no_objects(
mask_probs[i], pred_scores[i], pred_labels[i], threshold, num_labels
)
# No mask found
if mask_probs_item.shape[0] <= 0:
height, width = target_sizes[i] if target_sizes is not None else mask_probs_item.shape[1:]
segmentation = torch.zeros((height, width)) - 1
results.append({"segmentation": segmentation, "segments_info": []})
continue
# Get segmentation map and segment information of batch item
target_size = target_sizes[i] if target_sizes is not None else None
segmentation, segments = compute_segments(
mask_probs_item,
pred_scores_item,
pred_labels_item,
mask_threshold,
overlap_mask_area_threshold,
target_size,
)
# Return segmentation map in run-length encoding (RLE) format
if return_coco_annotation:
segmentation = convert_segmentation_to_rle(segmentation)
results.append({"segmentation": segmentation, "segments_info": segments})
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_panoptic_segmentation with Detr->ConditionalDetr
def post_process_panoptic_segmentation(
self,
outputs,
threshold: float = 0.5,
mask_threshold: float = 0.5,
overlap_mask_area_threshold: float = 0.8,
label_ids_to_fuse: Optional[Set[int]] = None,
target_sizes: Optional[List[Tuple[int, int]]] = None,
) -> List[Dict]:
"""
Converts the output of [`ConditionalDetrForSegmentation`] into image panoptic segmentation predictions. Only
supports PyTorch.
Args:
outputs ([`ConditionalDetrForSegmentation`]):
The outputs from [`ConditionalDetrForSegmentation`].
threshold (`float`, *optional*, defaults to 0.5):
The probability score threshold to keep predicted instance masks.
mask_threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
overlap_mask_area_threshold (`float`, *optional*, defaults to 0.8):
The overlap mask area threshold to merge or discard small disconnected parts within each binary
instance mask.
label_ids_to_fuse (`Set[int]`, *optional*):
The labels in this state will have all their instances be fused together. For instance we could say
there can only be one sky in an image, but several persons, so the label ID for sky would be in that
set, but not the one for person.
target_sizes (`List[Tuple]`, *optional*):
List of length (batch_size), where each list item (`Tuple[int, int]]`) corresponds to the requested
final size (height, width) of each prediction in batch. If left to None, predictions will not be
resized.
Returns:
`List[Dict]`: A list of dictionaries, one per image, each dictionary containing two keys:
- **segmentation** -- a tensor of shape `(height, width)` where each pixel represents a `segment_id` or
`None` if no mask if found above `threshold`. If `target_sizes` is specified, segmentation is resized to
the corresponding `target_sizes` entry.
- **segments_info** -- A dictionary that contains additional information on each segment.
- **id** -- an integer representing the `segment_id`.
- **label_id** -- An integer representing the label / semantic class id corresponding to `segment_id`.
- **was_fused** -- a boolean, `True` if `label_id` was in `label_ids_to_fuse`, `False` otherwise.
Multiple instances of the same class / label were fused and assigned a single `segment_id`.
- **score** -- Prediction score of segment with `segment_id`.
"""
if label_ids_to_fuse is None:
warnings.warn("`label_ids_to_fuse` unset. No instance will be fused.")
label_ids_to_fuse = set()
class_queries_logits = outputs.logits # [batch_size, num_queries, num_classes+1]
masks_queries_logits = outputs.pred_masks # [batch_size, num_queries, height, width]
batch_size = class_queries_logits.shape[0]
num_labels = class_queries_logits.shape[-1] - 1
mask_probs = masks_queries_logits.sigmoid() # [batch_size, num_queries, height, width]
# Predicted label and score of each query (batch_size, num_queries)
pred_scores, pred_labels = nn.functional.softmax(class_queries_logits, dim=-1).max(-1)
# Loop over items in batch size
results: List[Dict[str, TensorType]] = []
for i in range(batch_size):
mask_probs_item, pred_scores_item, pred_labels_item = remove_low_and_no_objects(
mask_probs[i], pred_scores[i], pred_labels[i], threshold, num_labels
)
# No mask found
if mask_probs_item.shape[0] <= 0:
height, width = target_sizes[i] if target_sizes is not None else mask_probs_item.shape[1:]
segmentation = torch.zeros((height, width)) - 1
results.append({"segmentation": segmentation, "segments_info": []})
continue
# Get segmentation map and segment information of batch item
target_size = target_sizes[i] if target_sizes is not None else None
segmentation, segments = compute_segments(
mask_probs_item,
pred_scores_item,
pred_labels_item,
mask_threshold,
overlap_mask_area_threshold,
label_ids_to_fuse,
target_size,
)
results.append({"segmentation": segmentation, "segments_info": segments})
return results

24
src/transformers/models/conditional_detr/modeling_conditional_detr.py
View File

@@ -1699,11 +1699,11 @@ class ConditionalDetrForObjectDetection(ConditionalDetrPreTrainedModel):
>>> # convert outputs (bounding boxes and class logits) to COCO API
>>> target_sizes = torch.tensor([image.size[::-1]])
>>> results = feature_extractor.post_process(outputs, target_sizes=target_sizes)[0]
>>> results = feature_extractor.post_process_object_detection(
... outputs, threshold=0.5, target_sizes=target_sizes
... )[0]
>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
... box = [round(i, 2) for i in box.tolist()]
... # let's only keep detections with score > 0.5
... if score > 0.5:
... print(
... f"Detected {model.config.id2label[label.item()]} with confidence "
... f"{round(score.item(), 3)} at location {box}"
@@ -1897,17 +1897,13 @@ class ConditionalDetrForSegmentation(ConditionalDetrPreTrainedModel):
>>> # forward pass
>>> outputs = model(**inputs)
>>> # use the `post_process_panoptic` method of `ConditionalDetrFeatureExtractor` to convert to COCO format
>>> processed_sizes = torch.as_tensor(inputs["pixel_values"].shape[-2:]).unsqueeze(0)
>>> result = feature_extractor.post_process_panoptic(outputs, processed_sizes)[0]
>>> # the segmentation is stored in a special-format png
>>> panoptic_seg = Image.open(io.BytesIO(result["png_string"]))
>>> panoptic_seg = numpy.array(panoptic_seg, dtype=numpy.uint8)
>>> # retrieve the ids corresponding to each mask
>>> panoptic_seg_id = rgb_to_id(panoptic_seg)
>>> panoptic_seg_id.shape
(800, 1066)
>>> # Use the `post_process_panoptic_segmentation` method of `ConditionalDetrFeatureExtractor` to retrieve post-processed panoptic segmentation maps
>>> # Segmentation results are returned as a list of dictionaries
>>> result = feature_extractor.post_process_panoptic_segmentation(outputs, target_sizes=[(300, 500)])
>>> # A tensor of shape (height, width) where each value denotes a segment id, filled with -1 if no segment is found
>>> panoptic_seg = result[0]["segmentation"]
>>> # Get prediction score and segment_id to class_id mapping of each segment
>>> panoptic_segments_info = result[0]["segments_info"]
```"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict

274
src/transformers/models/deformable_detr/feature_extraction_deformable_detr.py
View File

@@ -14,11 +14,9 @@
# limitations under the License.
"""Feature extractor class for Deformable DETR."""
import io
import pathlib
import warnings
from collections import defaultdict
from typing import Dict, List, Optional, Union
from typing import Dict, List, Optional, Tuple, Union
import numpy as np
from PIL import Image
@@ -707,6 +705,12 @@ class DeformableDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtract
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_object_detection`.",
FutureWarning,
)
out_logits, out_bbox = outputs.logits, outputs.pred_boxes
if len(out_logits) != len(target_sizes):
@@ -731,240 +735,56 @@ class DeformableDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtract
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_segmentation with Detr->DeformableDetr
def post_process_segmentation(self, outputs, target_sizes, threshold=0.9, mask_threshold=0.5):
def post_process_object_detection(
self, outputs, threshold: float = 0.5, target_sizes: Union[TensorType, List[Tuple]] = None
):
"""
Converts the output of [`DeformableDetrForSegmentation`] into image segmentation predictions. Only supports
PyTorch.
Parameters:
outputs ([`DeformableDetrSegmentationOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction.
threshold (`float`, *optional*, defaults to 0.9):
Threshold to use to filter out queries.
mask_threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, and masks for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process_segmentation` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_semantic_segmentation`.",
FutureWarning,
)
out_logits, raw_masks = outputs.logits, outputs.pred_masks
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, size in zip(out_logits, raw_masks, target_sizes):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_masks = (cur_masks.sigmoid() > mask_threshold) * 1
predictions = {"scores": cur_scores, "labels": cur_classes, "masks": cur_masks}
preds.append(predictions)
return preds
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_instance with Detr->DeformableDetr
def post_process_instance(self, results, outputs, orig_target_sizes, max_target_sizes, threshold=0.5):
"""
Converts the output of [`DeformableDetrForSegmentation`] into actual instance segmentation predictions. Only
Converts the output of [`DeformableDetrForObjectDetection`] into the format expected by the COCO api. Only
supports PyTorch.
Args:
results (`List[Dict]`):
Results list obtained by [`~DeformableDetrFeatureExtractor.post_process`], to which "masks" results
will be added.
outputs ([`DeformableDetrSegmentationOutput`]):
outputs ([`DetrObjectDetectionOutput`]):
Raw outputs of the model.
orig_target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the size (h, w) of each image of the batch. For evaluation, this must be the original
image size (before any data augmentation).
max_target_sizes (`torch.Tensor` of shape `(batch_size, 2)`):
Tensor containing the maximum size (h, w) of each image of the batch. For evaluation, this must be the
original image size (before any data augmentation).
threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
threshold (`float`, *optional*):
Score threshold to keep object detection predictions.
target_sizes (`torch.Tensor` or `List[Tuple[int, int]]`, *optional*, defaults to `None`):
Tensor of shape `(batch_size, 2)` or list of tuples (`Tuple[int, int]`) containing the target size
(height, width) of each image in the batch. If left to None, predictions will not be resized.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, boxes and masks for an
image in the batch as predicted by the model.
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels and boxes for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process_instance` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_instance_segmentation`.",
FutureWarning,
out_logits, out_bbox = outputs.logits, outputs.pred_boxes
if target_sizes is not None:
if len(out_logits) != len(target_sizes):
raise ValueError(
"Make sure that you pass in as many target sizes as the batch dimension of the logits"
)
if len(orig_target_sizes) != len(max_target_sizes):
raise ValueError("Make sure to pass in as many orig_target_sizes as max_target_sizes")
max_h, max_w = max_target_sizes.max(0)[0].tolist()
outputs_masks = outputs.pred_masks.squeeze(2)
outputs_masks = nn.functional.interpolate(
outputs_masks, size=(max_h, max_w), mode="bilinear", align_corners=False
)
outputs_masks = (outputs_masks.sigmoid() > threshold).cpu()
prob = out_logits.sigmoid()
topk_values, topk_indexes = torch.topk(prob.view(out_logits.shape[0], -1), 100, dim=1)
scores = topk_values
topk_boxes = topk_indexes // out_logits.shape[2]
labels = topk_indexes % out_logits.shape[2]
boxes = center_to_corners_format(out_bbox)
boxes = torch.gather(boxes, 1, topk_boxes.unsqueeze(-1).repeat(1, 1, 4))
for i, (cur_mask, t, tt) in enumerate(zip(outputs_masks, max_target_sizes, orig_target_sizes)):
img_h, img_w = t[0], t[1]
results[i]["masks"] = cur_mask[:, :img_h, :img_w].unsqueeze(1)
results[i]["masks"] = nn.functional.interpolate(
results[i]["masks"].float(), size=tuple(tt.tolist()), mode="nearest"
).byte()
# and from relative [0, 1] to absolute [0, height] coordinates
if isinstance(target_sizes, List):
img_h = torch.Tensor([i[0] for i in target_sizes])
img_w = torch.Tensor([i[1] for i in target_sizes])
else:
img_h, img_w = target_sizes.unbind(1)
scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1)
boxes = boxes * scale_fct[:, None, :]
results = []
for s, l, b in zip(scores, labels, boxes):
score = s[s > threshold]
label = l[s > threshold]
box = b[s > threshold]
results.append({"scores": score, "labels": label, "boxes": box})
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_panoptic with Detr->DeformableDetr
def post_process_panoptic(self, outputs, processed_sizes, target_sizes=None, is_thing_map=None, threshold=0.85):
"""
Converts the output of [`DeformableDetrForSegmentation`] into actual panoptic predictions. Only supports
PyTorch.
Parameters:
outputs ([`DeformableDetrSegmentationOutput`]):
Raw outputs of the model.
processed_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) containing the size (h, w) of each image of the batch, i.e. the size after data
augmentation but before batching.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`, *optional*):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction. If left to
None, it will default to the `processed_sizes`.
is_thing_map (`torch.Tensor` of shape `(batch_size, 2)`, *optional*):
Dictionary mapping class indices to either True or False, depending on whether or not they are a thing.
If not set, defaults to the `is_thing_map` of COCO panoptic.
threshold (`float`, *optional*, defaults to 0.85):
Threshold to use to filter out queries.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing a PNG string and segments_info values for
an image in the batch as predicted by the model.
"""
warnings.warn(
"`post_process_panoptic is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_panoptic_segmentation`.",
FutureWarning,
)
if target_sizes is None:
target_sizes = processed_sizes
if len(processed_sizes) != len(target_sizes):
raise ValueError("Make sure to pass in as many processed_sizes as target_sizes")
if is_thing_map is None:
# default to is_thing_map of COCO panoptic
is_thing_map = {i: i <= 90 for i in range(201)}
out_logits, raw_masks, raw_boxes = outputs.logits, outputs.pred_masks, outputs.pred_boxes
if not len(out_logits) == len(raw_masks) == len(target_sizes):
raise ValueError(
"Make sure that you pass in as many target sizes as the batch dimension of the logits and masks"
)
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, cur_boxes, size, target_size in zip(
out_logits, raw_masks, raw_boxes, processed_sizes, target_sizes
):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_boxes = center_to_corners_format(cur_boxes[keep])
h, w = cur_masks.shape[-2:]
if len(cur_boxes) != len(cur_classes):
raise ValueError("Not as many boxes as there are classes")
# It may be that we have several predicted masks for the same stuff class.
# In the following, we track the list of masks ids for each stuff class (they are merged later on)
cur_masks = cur_masks.flatten(1)
stuff_equiv_classes = defaultdict(lambda: [])
for k, label in enumerate(cur_classes):
if not is_thing_map[label.item()]:
stuff_equiv_classes[label.item()].append(k)
def get_ids_area(masks, scores, dedup=False):
# This helper function creates the final panoptic segmentation image
# It also returns the area of the masks that appears on the image
m_id = masks.transpose(0, 1).softmax(-1)
if m_id.shape[-1] == 0:
# We didn't detect any mask :(
m_id = torch.zeros((h, w), dtype=torch.long, device=m_id.device)
else:
m_id = m_id.argmax(-1).view(h, w)
if dedup:
# Merge the masks corresponding to the same stuff class
for equiv in stuff_equiv_classes.values():
if len(equiv) > 1:
for eq_id in equiv:
m_id.masked_fill_(m_id.eq(eq_id), equiv[0])
final_h, final_w = to_tuple(target_size)
seg_img = Image.fromarray(id_to_rgb(m_id.view(h, w).cpu().numpy()))
seg_img = seg_img.resize(size=(final_w, final_h), resample=Image.NEAREST)
np_seg_img = torch.ByteTensor(torch.ByteStorage.from_buffer(seg_img.tobytes()))
np_seg_img = np_seg_img.view(final_h, final_w, 3)
np_seg_img = np_seg_img.numpy()
m_id = torch.from_numpy(rgb_to_id(np_seg_img))
area = []
for i in range(len(scores)):
area.append(m_id.eq(i).sum().item())
return area, seg_img
area, seg_img = get_ids_area(cur_masks, cur_scores, dedup=True)
if cur_classes.numel() > 0:
# We know filter empty masks as long as we find some
while True:
filtered_small = torch.as_tensor(
[area[i] <= 4 for i, c in enumerate(cur_classes)], dtype=torch.bool, device=keep.device
)
if filtered_small.any().item():
cur_scores = cur_scores[~filtered_small]
cur_classes = cur_classes[~filtered_small]
cur_masks = cur_masks[~filtered_small]
area, seg_img = get_ids_area(cur_masks, cur_scores)
else:
break
else:
cur_classes = torch.ones(1, dtype=torch.long, device=cur_classes.device)
segments_info = []
for i, a in enumerate(area):
cat = cur_classes[i].item()
segments_info.append({"id": i, "isthing": is_thing_map[cat], "category_id": cat, "area": a})
del cur_classes
with io.BytesIO() as out:
seg_img.save(out, format="PNG")
predictions = {"png_string": out.getvalue(), "segments_info": segments_info}
preds.append(predictions)
return preds

10
src/transformers/models/deformable_detr/modeling_deformable_detr.py
View File

@@ -236,8 +236,8 @@ class DeformableDetrObjectDetectionOutput(ModelOutput):
pred_boxes (`torch.FloatTensor` of shape `(batch_size, num_queries, 4)`):
Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These
values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding
possible padding). You can use [`~AutoFeatureExtractor.post_process`] to retrieve the unnormalized bounding
boxes.
possible padding). You can use [`~AutoFeatureExtractor.post_process_object_detection`] to retrieve the
unnormalized bounding boxes.
auxiliary_outputs (`list[Dict]`, *optional*):
Optional, only returned when auxilary losses are activated (i.e. `config.auxiliary_loss` is set to `True`)
and labels are provided. It is a list of dictionaries containing the two above keys (`logits` and
@@ -1878,11 +1878,11 @@ class DeformableDetrForObjectDetection(DeformableDetrPreTrainedModel):
>>> # convert outputs (bounding boxes and class logits) to COCO API
>>> target_sizes = torch.tensor([image.size[::-1]])
>>> results = feature_extractor.post_process(outputs, target_sizes=target_sizes)[0]
>>> results = feature_extractor.post_process_object_detection(
... outputs, threshold=0.5, target_sizes=target_sizes
... )[0]
>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
... box = [round(i, 2) for i in box.tolist()]
... # let's only keep detections with score > 0.5
... if score > 0.5:
... print(
... f"Detected {model.config.id2label[label.item()]} with confidence "
... f"{round(score.item(), 3)} at location {box}"

16
src/transformers/models/detr/feature_extraction_detr.py
View File

@@ -878,7 +878,7 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
"""
Converts the output of [`DetrForSegmentation`] into image segmentation predictions. Only supports PyTorch.
Parameters:
Args:
outputs ([`DetrSegmentationOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
@@ -974,7 +974,7 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
"""
Converts the output of [`DetrForSegmentation`] into actual panoptic predictions. Only supports PyTorch.
Parameters:
Args:
outputs ([`DetrSegmentationOutput`]):
Raw outputs of the model.
processed_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
@@ -1165,13 +1165,15 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
def post_process_semantic_segmentation(self, outputs, target_sizes: List[Tuple[int, int]] = None):
"""
Args:
Converts the output of [`DetrForSegmentation`] into semantic segmentation maps. Only supports PyTorch.
Args:
outputs ([`DetrForSegmentation`]):
Raw outputs of the model.
target_sizes (`List[Tuple[int, int]]`, *optional*, defaults to `None`):
A list of tuples (`Tuple[int, int]`) containing the target size (height, width) of each image in the
batch. If left to None, predictions will not be resized.
Returns:
`List[torch.Tensor]`:
A list of length `batch_size`, where each item is a semantic segmentation map of shape (height, width)
@@ -1219,8 +1221,9 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
return_coco_annotation: Optional[bool] = False,
) -> List[Dict]:
"""
Args:
Converts the output of [`DetrForSegmentation`] into instance segmentation predictions. Only supports PyTorch.
Args:
outputs ([`DetrForSegmentation`]):
Raw outputs of the model.
threshold (`float`, *optional*, defaults to 0.5):
@@ -1236,6 +1239,7 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
return_coco_annotation (`bool`, *optional*):
Defaults to `False`. If set to `True`, segmentation maps are returned in COCO run-length encoding (RLE)
format.
Returns:
`List[Dict]`: A list of dictionaries, one per image, each dictionary containing two keys:
- **segmentation** -- A tensor of shape `(height, width)` where each pixel represents a `segment_id` or
@@ -1301,9 +1305,10 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
target_sizes: Optional[List[Tuple[int, int]]] = None,
) -> List[Dict]:
"""
Args:
Converts the output of [`DetrForSegmentation`] into image panoptic segmentation predictions. Only supports
PyTorch.
Args:
outputs ([`DetrForSegmentation`]):
The outputs from [`DetrForSegmentation`].
threshold (`float`, *optional*, defaults to 0.5):
@@ -1321,6 +1326,7 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin):
List of length (batch_size), where each list item (`Tuple[int, int]]`) corresponds to the requested
final size (height, width) of each prediction in batch. If left to None, predictions will not be
resized.
Returns:
`List[Dict]`: A list of dictionaries, one per image, each dictionary containing two keys:
- **segmentation** -- a tensor of shape `(height, width)` where each pixel represents a `segment_id` or

47
src/transformers/models/yolos/feature_extraction_yolos.py
View File

@@ -699,53 +699,6 @@ class YolosFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin)
results = [{"scores": s, "labels": l, "boxes": b} for s, l, b in zip(scores, labels, boxes)]
return results
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_segmentation
def post_process_segmentation(self, outputs, target_sizes, threshold=0.9, mask_threshold=0.5):
"""
Converts the output of [`DetrForSegmentation`] into image segmentation predictions. Only supports PyTorch.
Parameters:
outputs ([`DetrSegmentationOutput`]):
Raw outputs of the model.
target_sizes (`torch.Tensor` of shape `(batch_size, 2)` or `List[Tuple]` of length `batch_size`):
Torch Tensor (or list) corresponding to the requested final size (h, w) of each prediction.
threshold (`float`, *optional*, defaults to 0.9):
Threshold to use to filter out queries.
mask_threshold (`float`, *optional*, defaults to 0.5):
Threshold to use when turning the predicted masks into binary values.
Returns:
`List[Dict]`: A list of dictionaries, each dictionary containing the scores, labels, and masks for an image
in the batch as predicted by the model.
"""
warnings.warn(
"`post_process_segmentation` is deprecated and will be removed in v5 of Transformers, please use"
" `post_process_semantic_segmentation`.",
FutureWarning,
)
out_logits, raw_masks = outputs.logits, outputs.pred_masks
preds = []
def to_tuple(tup):
if isinstance(tup, tuple):
return tup
return tuple(tup.cpu().tolist())
for cur_logits, cur_masks, size in zip(out_logits, raw_masks, target_sizes):
# we filter empty queries and detection below threshold
scores, labels = cur_logits.softmax(-1).max(-1)
keep = labels.ne(outputs.logits.shape[-1] - 1) & (scores > threshold)
cur_scores, cur_classes = cur_logits.softmax(-1).max(-1)
cur_scores = cur_scores[keep]
cur_classes = cur_classes[keep]
cur_masks = cur_masks[keep]
cur_masks = nn.functional.interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1)
cur_masks = (cur_masks.sigmoid() > mask_threshold) * 1
predictions = {"scores": cur_scores, "labels": cur_classes, "masks": cur_masks}
preds.append(predictions)
return preds
# Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor.post_process_object_detection with Detr->Yolos
def post_process_object_detection(
self, outputs, threshold: float = 0.5, target_sizes: Union[TensorType, List[Tuple]] = None

1
tests/models/conditional_detr/test_feature_extraction_conditional_detr.py
View File

@@ -303,7 +303,6 @@ class ConditionalDetrFeatureExtractionTest(FeatureExtractionSavingTestMixin, uni
masks_path = pathlib.Path("./tests/fixtures/tests_samples/COCO/coco_panoptic")
# encode them
# TODO replace by .from_pretrained microsoft/conditional-detr-resnet-50-panoptic
feature_extractor = ConditionalDetrFeatureExtractor(format="coco_panoptic")
encoding = feature_extractor(images=image, annotations=target, masks_path=masks_path, return_tensors="pt")

14
tests/models/conditional_detr/test_modeling_conditional_detr.py
View File

@@ -492,6 +492,7 @@ class ConditionalDetrModelIntegrationTests(unittest.TestCase):
with torch.no_grad():
outputs = model(pixel_values, pixel_mask)
# verify logits + box predictions
expected_shape_logits = torch.Size((1, model.config.num_queries, model.config.num_labels))
self.assertEqual(outputs.logits.shape, expected_shape_logits)
expected_slice_logits = torch.tensor(
@@ -505,3 +506,16 @@ class ConditionalDetrModelIntegrationTests(unittest.TestCase):
[[0.7733, 0.6576, 0.4496], [0.5171, 0.1184, 0.9094], [0.8846, 0.5647, 0.2486]]
).to(torch_device)
self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_slice_boxes, atol=1e-4))
# verify postprocessing
results = feature_extractor.post_process_object_detection(
outputs, threshold=0.3, target_sizes=[image.size[::-1]]
)[0]
expected_scores = torch.tensor([0.8330, 0.8313, 0.8039, 0.6829, 0.5355])
expected_labels = [75, 17, 17, 75, 63]
expected_slice_boxes = torch.tensor([38.3089, 72.1022, 177.6293, 118.4512])
self.assertEqual(len(results["scores"]), 5)
self.assertTrue(torch.allclose(results["scores"], expected_scores, atol=1e-4))
self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
self.assertTrue(torch.allclose(results["boxes"][0, :], expected_slice_boxes))

13
tests/models/deformable_detr/test_modeling_deformable_detr.py
View File

@@ -565,6 +565,19 @@ class DeformableDetrModelIntegrationTests(unittest.TestCase):
self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes)
self.assertTrue(torch.allclose(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=1e-4))
# verify postprocessing
results = feature_extractor.post_process_object_detection(
outputs, threshold=0.3, target_sizes=[image.size[::-1]]
)[0]
expected_scores = torch.tensor([0.7999, 0.7894, 0.6331, 0.4720, 0.4382])
expected_labels = [17, 17, 75, 75, 63]
expected_slice_boxes = torch.tensor([16.5028, 52.8390, 318.2544, 470.7841])
self.assertEqual(len(results["scores"]), 5)
self.assertTrue(torch.allclose(results["scores"], expected_scores, atol=1e-4))
self.assertSequenceEqual(results["labels"].tolist(), expected_labels)
self.assertTrue(torch.allclose(results["boxes"][0, :], expected_slice_boxes))
def test_inference_object_detection_head_with_box_refine_two_stage(self):
model = DeformableDetrForObjectDetection.from_pretrained(
"SenseTime/deformable-detr-with-box-refine-two-stage"