Build Models from Yacs Config¶
From a yacs config object,
models (and their sub-models) can be built by
functions such as
from detectron2.modeling import build_model model = build_model(cfg) # returns a torch.nn.Module
build_model only builds the model structure and fills it with random parameters.
See below for how to load an existing checkpoint to the model and how to use the
Load/Save a Checkpoint¶
from detectron2.checkpoint import DetectionCheckpointer DetectionCheckpointer(model).load(file_path_or_url) # load a file, usually from cfg.MODEL.WEIGHTS checkpointer = DetectionCheckpointer(model, save_dir="output") checkpointer.save("model_999") # save to output/model_999.pth
Detectron2’s checkpointer recognizes models in pytorch’s
.pth format, as well as the
in our model zoo.
See API doc
for more details about its usage.
The model files can be arbitrarily manipulated using
.pth files or
Use a Model¶
A model can be called by
outputs = model(inputs), where
inputs is a
Each dict corresponds to one image and the required keys
depend on the type of model, and whether the model is in training or evaluation mode.
For example, in order to do inference,
all existing models expect the “image” key, and optionally “height” and “width”.
The detailed format of inputs and outputs of existing models are explained below.
Training: When in training mode, all models are required to be used under an
The training statistics will be put into the storage:
from detectron2.utils.events import EventStorage with EventStorage() as storage: losses = model(inputs)
Inference: If you only want to do simple inference using an existing model, DefaultPredictor is a wrapper around model that provides such basic functionality. It includes default behavior including model loading, preprocessing, and operates on single image rather than batches. See its documentation for usage.
You can also run inference directly like this:
model.eval() with torch.no_grad(): outputs = model(inputs)
Model Input Format¶
Users can implement custom models that support any arbitrary input format.
Here we describe the standard input format that all builtin models support in detectron2.
They all take a
list[dict] as the inputs. Each dict
corresponds to information about one image.
The dict may contain the following keys:
Tensorin (C, H, W) format. The meaning of channels are defined by
cfg.INPUT.FORMAT. Image normalization, if any, will be performed inside the model using
“height”, “width”: the desired output height and width in inference, which is not necessarily the same as the height or width of the
imagefield. For example, the
imagefield contains the resized image, if resize is used as a preprocessing step. But you may want the outputs to be in original resolution. If provided, the model will produce output in this resolution, rather than in the resolution of the
imageas input into the model. This is more efficient and accurate.
“instances”: an Instances object for training, with the following fields:
“gt_boxes”: a Boxes object storing N boxes, one for each instance.
Tensorof long type, a vector of N labels, in range [0, num_categories).
“gt_keypoints”: a Keypoints object storing N keypoint sets, one for each instance.
Tensor[int]in (H, W) format. The semantic segmentation ground truth for training. Values represent category labels starting from 0.
“proposals”: an Instances object used only in Fast R-CNN style models, with the following fields:
“proposal_boxes”: a Boxes object storing P proposal boxes.
Tensor, a vector of P scores, one for each proposal.
For inference of builtin models, only “image” key is required, and “width/height” are optional.
We currently don’t define standard input format for panoptic segmentation training, because models now use custom formats produced by custom data loaders.
Model Output Format¶
When in training mode, the builtin models output a
dict[str->ScalarTensor] with all the losses.
When in inference mode, the builtin models output a
list[dict], one dict for each image.
Based on the tasks the model is doing, each dict may contain the following fields:
“instances”: Instances object with the following fields:
“pred_boxes”: Boxes object storing N boxes, one for each detected instance.
Tensor, a vector of N confidence scores.
Tensor, a vector of N labels in range [0, num_categories).
Tensorof shape (N, H, W), masks for each detected instance.
Tensorof shape (N, num_keypoint, 3). Each row in the last dimension is (x, y, score). Confidence scores are larger than 0.
Tensorof (num_categories, H, W), the semantic segmentation prediction.
“proposals”: Instances object with the following fields:
“proposal_boxes”: Boxes object storing N boxes.
“objectness_logits”: a torch vector of N confidence scores.
“panoptic_seg”: A tuple of
(pred: Tensor, segments_info: Optional[list[dict]]). The
predtensor has shape (H, W), containing the segment id of each pixel.
segments_infoexists, each dict describes one segment id in
predand has the following fields:
“id”: the segment id
“isthing”: whether the segment is a thing or stuff
“category_id”: the category id of this segment.
If a pixel’s id does not exist in
segments_info, it is considered to be void label defined in Panoptic Segmentation.
segments_infois None, all pixel values in
predmust be ≥ -1. Pixels with value -1 are assigned void labels. Otherwise, the category id of each pixel is obtained by
category_id = pixel // metadata.label_divisor.
Partially execute a model:¶
Sometimes you may want to obtain an intermediate tensor inside a model, such as the input of certain layer, the output before post-processing. Since there are typically hundreds of intermediate tensors, there isn’t an API that provides you the intermediate result you need. You have the following options:
Write a (sub)model. Following the tutorial, you can rewrite a model component (e.g. a head of a model), such that it does the same thing as the existing component, but returns the output you need.
Partially execute a model. You can create the model as usual, but use custom code to execute it instead of its
forward(). For example, the following code obtains mask features before mask head.
images = ImageList.from_tensors(...) # preprocessed input tensor model = build_model(cfg) model.eval() features = model.backbone(images.tensor) proposals, _ = model.proposal_generator(images, features) instances, _ = model.roi_heads(images, features, proposals) mask_features = [features[f] for f in model.roi_heads.in_features] mask_features = model.roi_heads.mask_pooler(mask_features, [x.pred_boxes for x in instances])
Use forward hooks. Forward hooks can help you obtain inputs or outputs of a certain module. If they are not exactly what you want, they can at least be used together with partial execution to obtain other tensors.
All options require you to read documentation and sometimes code of the existing models to understand the internal logic, in order to write code to obtain the internal tensors.