# detectron2.utils package¶

## detectron2.utils.colormap module¶

An awesome colormap for really neat visualizations. Copied from Detectron, and removed gray colors.

detectron2.utils.colormap.colormap(rgb=False, maximum=255)[source]
Parameters
• rgb (bool) – whether to return RGB colors or BGR colors.

• maximum (int) – either 255 or 1

Returns

ndarray – a float32 array of Nx3 colors, in range [0, 255] or [0, 1]

detectron2.utils.colormap.random_color(rgb=False, maximum=255)[source]
Parameters
• rgb (bool) – whether to return RGB colors or BGR colors.

• maximum (int) – either 255 or 1

Returns

ndarray – a vector of 3 numbers

## detectron2.utils.comm module¶

This file contains primitives for multi-gpu communication. This is useful when doing distributed training.

detectron2.utils.comm.get_world_size()int[source]
detectron2.utils.comm.get_rank()int[source]
detectron2.utils.comm.get_local_rank()int[source]
Returns

The rank of the current process within the local (per-machine) process group.

detectron2.utils.comm.get_local_size()int[source]
Returns

The size of the per-machine process group, i.e. the number of processes per machine.

detectron2.utils.comm.is_main_process()bool[source]
detectron2.utils.comm.synchronize()[source]

Helper function to synchronize (barrier) among all processes when using distributed training

detectron2.utils.comm.all_gather(data, group=None)[source]

Run all_gather on arbitrary picklable data (not necessarily tensors).

Parameters
• data – any picklable object

• group – a torch process group. By default, will use a group which contains all ranks on gloo backend.

Returns

list[data] – list of data gathered from each rank

detectron2.utils.comm.gather(data, dst=0, group=None)[source]

Run gather on arbitrary picklable data (not necessarily tensors).

Parameters
• data – any picklable object

• dst (int) – destination rank

• group – a torch process group. By default, will use a group which contains all ranks on gloo backend.

Returns

list[data]

on dst, a list of data gathered from each rank. Otherwise,

an empty list.

detectron2.utils.comm.shared_random_seed()[source]
Returns

int

a random number that is the same across all workers.

If workers need a shared RNG, they can use this shared seed to create one.

All workers must call this function, otherwise it will deadlock.

detectron2.utils.comm.reduce_dict(input_dict, average=True)[source]

Reduce the values in the dictionary from all processes so that process with rank 0 has the reduced results.

Parameters
• input_dict (dict) – inputs to be reduced. All the values must be scalar CUDA Tensor.

• average (bool) – whether to do average or sum

Returns

a dict with the same keys as input_dict, after reduction.

## detectron2.utils.events module¶

detectron2.utils.events.get_event_storage()[source]
Returns

The EventStorage object that’s currently being used. Throws an error if no EventStorage is currently enabled.

class detectron2.utils.events.EventWriter[source]

Bases: object

Base class for writers that obtain events from EventStorage and process them.

write()[source]
close()[source]
class detectron2.utils.events.JSONWriter(json_file, window_size=20)[source]

Write scalars to a json file.

It saves scalars as one json per line (instead of a big json) for easy parsing.

Examples parsing such a json file:

$cat metrics.json | jq -s '.[0:2]' [ { "data_time": 0.008433341979980469, "iteration": 20, "loss": 1.9228371381759644, "loss_box_reg": 0.050025828182697296, "loss_classifier": 0.5316952466964722, "loss_mask": 0.7236229181289673, "loss_rpn_box": 0.0856662318110466, "loss_rpn_cls": 0.48198649287223816, "lr": 0.007173333333333333, "time": 0.25401854515075684 }, { "data_time": 0.007216215133666992, "iteration": 40, "loss": 1.282649278640747, "loss_box_reg": 0.06222952902317047, "loss_classifier": 0.30682939291000366, "loss_mask": 0.6970193982124329, "loss_rpn_box": 0.038663312792778015, "loss_rpn_cls": 0.1471673548221588, "lr": 0.007706666666666667, "time": 0.2490077018737793 } ]$ cat metrics.json | jq '.loss_mask'
0.7126231789588928
0.689423680305481
0.6776131987571716
...

__init__(json_file, window_size=20)[source]
Parameters
• json_file (str) – path to the json file. New data will be appended if the file exists.

• window_size (int) – the window size of median smoothing for the scalars whose smoothing_hint are True.

write()[source]
close()[source]
class detectron2.utils.events.TensorboardXWriter(log_dir: str, window_size: int = 20, **kwargs)[source]

Write all scalars to a tensorboard file.

__init__(log_dir: str, window_size: int = 20, **kwargs)[source]
Parameters
• log_dir (str) – the directory to save the output events

• window_size (int) – the scalars will be median-smoothed by this window size

• kwargs – other arguments passed to torch.utils.tensorboard.SummaryWriter(…)

write()[source]
close()[source]
class detectron2.utils.events.CommonMetricPrinter(max_iter)[source]

Print common metrics to the terminal, including iteration time, ETA, memory, all losses, and the learning rate.

To print something different, please implement a similar printer by yourself.

__init__(max_iter)[source]
Parameters

max_iter (int) – the maximum number of iterations to train. Used to compute ETA.

write()[source]
class detectron2.utils.events.EventStorage(start_iter=0)[source]

Bases: object

The user-facing class that provides metric storage functionalities.

In the future we may add support for storing / logging other types of data if needed.

__init__(start_iter=0)[source]
Parameters

put_image(img_name, img_tensor)[source]

Add an img_tensor associated with img_name, to be shown on tensorboard.

Parameters
• img_name (str) – The name of the image to put into tensorboard.

• img_tensor (torch.Tensor or numpy.array) – An uint8 or float Tensor of shape [channel, height, width] where channel is 3. The image format should be RGB. The elements in img_tensor can either have values in [0, 1] (float32) or [0, 255] (uint8). The img_tensor will be visualized in tensorboard.

put_scalar(name, value, smoothing_hint=True)[source]

Add a scalar value to the HistoryBuffer associated with name.

Parameters

smoothing_hint (bool) –

a ‘hint’ on whether this scalar is noisy and should be smoothed when logged. The hint will be accessible through EventStorage.smoothing_hints(). A writer may ignore the hint and apply custom smoothing rule.

It defaults to True because most scalars we save need to be smoothed to provide any useful signal.

put_scalars(*, smoothing_hint=True, **kwargs)[source]

Put multiple scalars from keyword arguments.

Examples

storage.put_scalars(loss=my_loss, accuracy=my_accuracy, smoothing_hint=True)

put_histogram(hist_name, hist_tensor, bins=1000)[source]

Create a histogram from a tensor.

Parameters
• hist_name (str) – The name of the histogram to put into tensorboard.

• hist_tensor (torch.Tensor) – A Tensor of arbitrary shape to be converted into a histogram.

• bins (int) – Number of histogram bins.

history(name)[source]
Returns

HistoryBuffer – the scalar history for name

histories()[source]
Returns

dict[name -> HistoryBuffer] – the HistoryBuffer for all scalars

latest()[source]
Returns

dict[name -> number] – the scalars that’s added in the current iteration.

latest_with_smoothing_hint(window_size=20)[source]

Similar to latest(), but the returned values are either the un-smoothed original latest value, or a median of the given window_size, depend on whether the smoothing_hint is True.

This provides a default behavior that other writers can use.

smoothing_hints()[source]
Returns

dict[name -> bool]

the user-provided hint on whether the scalar

is noisy and needs smoothing.

step()[source]

User should call this function at the beginning of each iteration, to notify the storage of the start of a new iteration. The storage will then be able to associate the new data with the correct iteration number.

property iter
property iteration
name_scope(name)[source]
Yields

A context within which all the events added to this storage will be prefixed by the name scope.

clear_images()[source]

Delete all the stored images for visualization. This should be called after images are written to tensorboard.

clear_histograms()[source]

Delete all the stored histograms for visualization. This should be called after histograms are written to tensorboard.

## detectron2.utils.logger module¶

detectron2.utils.logger.setup_logger(output=None, distributed_rank=0, *, color=True, name='detectron2', abbrev_name=None)[source]

Initialize the detectron2 logger and set its verbosity level to “DEBUG”.

Parameters
• output (str) – a file name or a directory to save log. If None, will not save log file. If ends with “.txt” or “.log”, assumed to be a file name. Otherwise, logs will be saved to output/log.txt.

• name (str) – the root module name of this logger

• abbrev_name (str) – an abbreviation of the module, to avoid long names in logs. Set to “” to not log the root module in logs. By default, will abbreviate “detectron2” to “d2” and leave other modules unchanged.

Returns

logging.Logger – a logger

detectron2.utils.logger.log_first_n(lvl, msg, n=1, *, name=None, key='caller')[source]

Log only for the first n times.

Parameters
• lvl (int) – the logging level

• msg (str) –

• n (int) –

• name (str) – name of the logger to use. Will use the caller’s module by default.

• key (str or tuple[str]) – the string(s) can be one of “caller” or “message”, which defines how to identify duplicated logs. For example, if called with n=1, key=”caller”, this function will only log the first call from the same caller, regardless of the message content. If called with n=1, key=”message”, this function will log the same content only once, even if they are called from different places. If called with n=1, key=(“caller”, “message”), this function will not log only if the same caller has logged the same message before.

detectron2.utils.logger.log_every_n(lvl, msg, n=1, *, name=None)[source]

Log once per n times.

Parameters
• lvl (int) – the logging level

• msg (str) –

• n (int) –

• name (str) – name of the logger to use. Will use the caller’s module by default.

detectron2.utils.logger.log_every_n_seconds(lvl, msg, n=1, *, name=None)[source]

Log no more than once per n seconds.

Parameters
• lvl (int) – the logging level

• msg (str) –

• n (int) –

• name (str) – name of the logger to use. Will use the caller’s module by default.

detectron2.utils.logger.create_small_table(small_dict)[source]

Create a small table using the keys of small_dict as headers. This is only suitable for small dictionaries.

Parameters

small_dict (dict) – a result dictionary of only a few items.

Returns

str – the table as a string.

## detectron2.utils.registry module¶

class detectron2.utils.registry.Registry(name: str)[source]

Bases: object

The registry that provides name -> object mapping, to support third-party users’ custom modules.

To create a registry (e.g. a backbone registry):

BACKBONE_REGISTRY = Registry('BACKBONE')


To register an object:

@BACKBONE_REGISTRY.register()
class MyBackbone():
...


Or:

BACKBONE_REGISTRY.register(MyBackbone)

__init__(name: str) → None[source]
Parameters

name (str) – the name of this registry

register(obj: object = None) → Optional[object][source]

Register the given object under the the name obj.__name__. Can be used as either a decorator or not. See docstring of this class for usage.

get(name: str)object[source]

## detectron2.utils.memory module¶

detectron2.utils.memory.retry_if_cuda_oom(func)[source]

Makes a function retry itself after encountering pytorch’s CUDA OOM error. It will first retry after calling torch.cuda.empty_cache().

If that still fails, it will then retry by trying to convert inputs to CPUs. In this case, it expects the function to dispatch to CPU implementation. The return values may become CPU tensors as well and it’s user’s responsibility to convert it back to CUDA tensor if needed.

Parameters

func – a stateless callable that takes tensor-like objects as arguments

Returns

a callable which retries func if OOM is encountered.

Examples:

output = retry_if_cuda_oom(some_torch_function)(input1, input2)
# output may be on CPU even if inputs are on GPU


Note

1. When converting inputs to CPU, it will only look at each argument and check if it has .device and .to for conversion. Nested structures of tensors are not supported.

2. Since the function might be called more than once, it has to be stateless.

## detectron2.utils.analysis module¶

detectron2.utils.analysis.parameter_count(model: torch.nn.modules.module.Module) → DefaultDict[str, int][source]

Count parameters of a model and its submodules.

Parameters

model – a torch module

Returns

dict (str-> int)

the key is either a parameter name or a module name.

The value is the number of elements in the parameter, or in all parameters of the module. The key “” corresponds to the total number of parameters of the model.

detectron2.utils.analysis.parameter_count_table(model: torch.nn.modules.module.Module, max_depth: int = 3)str[source]

Format the parameter count of the model (and its submodules or parameters) in a nice table.

Parameters
• model – a torch module

• max_depth (int) – maximum depth to recursively print submodules or parameters

Returns

str – the table to be printed

detectron2.utils.analysis.flop_count_operators(model: torch.nn.modules.module.Module, inputs: list, **kwargs) → DefaultDict[str, float][source]

Implement operator-level flops counting using jit. This is a wrapper of fvcore.nn.flop_count, that supports standard detection models in detectron2.

Note

The function runs the input through the model to compute flops. The flops of a detection model is often input-dependent, for example, the flops of box & mask head depends on the number of proposals & the number of detected objects. Therefore, the flops counting using a single input may not accurately reflect the computation cost of a model.

Parameters
• model – a detectron2 model that takes list[dict] as input.

• inputs (list[dict]) – inputs to model, in detectron2’s standard format.

detectron2.utils.analysis.activation_count_operators(model: torch.nn.modules.module.Module, inputs: list, **kwargs) → DefaultDict[str, float][source]

Implement operator-level activations counting using jit. This is a wrapper of fvcore.nn.activation_count, that supports standard detection models in detectron2.

Note

The function runs the input through the model to compute activations. The activations of a detection model is often input-dependent, for example, the activations of box & mask head depends on the number of proposals & the number of detected objects.

Parameters
• model – a detectron2 model that takes list[dict] as input.

• inputs (list[dict]) – inputs to model, in detectron2’s standard format.

## detectron2.utils.visualizer module¶

class detectron2.utils.visualizer.ColorMode[source]

Bases: enum.Enum

Enum of different color modes to use for instance visualizations.

IMAGE = 0

Picks a random color for every instance and overlay segmentations with low opacity.

SEGMENTATION = 1

Let instances of the same category have similar colors (from metadata.thing_colors), and overlay them with high opacity. This provides more attention on the quality of segmentation.

IMAGE_BW = 2

Same as IMAGE, but convert all areas without masks to gray-scale. Only available for drawing per-instance mask predictions.

class detectron2.utils.visualizer.VisImage(img, scale=1.0)[source]

Bases: object

__init__(img, scale=1.0)[source]
Parameters
• img (ndarray) – an RGB image of shape (H, W, 3).

• scale (float) – scale the input image

save(filepath)[source]
Parameters

filepath (str) – a string that contains the absolute path, including the file name, where the visualized image will be saved.

get_image()[source]
Returns

ndarray – the visualized image of shape (H, W, 3) (RGB) in uint8 type. The shape is scaled w.r.t the input image using the given scale argument.

class detectron2.utils.visualizer.Visualizer(img_rgb, metadata=None, scale=1.0, instance_mode=<ColorMode.IMAGE: 0>)[source]

Bases: object

__init__(img_rgb, metadata=None, scale=1.0, instance_mode=<ColorMode.IMAGE: 0>)[source]
Parameters
• img_rgb – a numpy array of shape (H, W, C), where H and W correspond to the height and width of the image respectively. C is the number of color channels. The image is required to be in RGB format since that is a requirement of the Matplotlib library. The image is also expected to be in the range [0, 255].

draw_instance_predictions(predictions)[source]

Draw instance-level prediction results on an image.

Parameters

predictions (Instances) – the output of an instance detection/segmentation model. Following fields will be used to draw: “pred_boxes”, “pred_classes”, “scores”, “pred_masks” (or “pred_masks_rle”).

Returns

output (VisImage) – image object with visualizations.

draw_sem_seg(sem_seg, area_threshold=None, alpha=0.8)[source]

Draw semantic segmentation predictions/labels.

Parameters
• sem_seg (Tensor or ndarray) – the segmentation of shape (H, W). Each value is the integer label of the pixel.

• area_threshold (int) – segments with less than area_threshold are not drawn.

• alpha (float) – the larger it is, the more opaque the segmentations are.

Returns

output (VisImage) – image object with visualizations.

draw_panoptic_seg_predictions(panoptic_seg, segments_info, area_threshold=None, alpha=0.7)[source]

Draw panoptic prediction results on an image.

Parameters
• panoptic_seg (Tensor) – of shape (height, width) where the values are ids for each segment.

• segments_info (list[dict]) – Describe each segment in panoptic_seg. Each dict contains keys “id”, “category_id”, “isthing”.

• area_threshold (int) – stuff segments with less than area_threshold are not drawn.

Returns

output (VisImage) – image object with visualizations.

draw_dataset_dict(dic)[source]

Draw annotations/segmentaions in Detectron2 Dataset format.

Parameters

dic (dict) – annotation/segmentation data of one image, in Detectron2 Dataset format.

Returns

output (VisImage) – image object with visualizations.

overlay_instances(*, boxes=None, labels=None, masks=None, keypoints=None, assigned_colors=None, alpha=0.5)[source]
Parameters
• boxes (Boxes, RotatedBoxes or ndarray) – either a Boxes, or an Nx4 numpy array of XYXY_ABS format for the N objects in a single image, or a RotatedBoxes, or an Nx5 numpy array of (x_center, y_center, width, height, angle_degrees) format for the N objects in a single image,

• labels (list[str]) – the text to be displayed for each instance.

Supported types are:

• list[list[ndarray]]: contains the segmentation masks for all objects in one image. The first level of the list corresponds to individual instances. The second level to all the polygon that compose the instance, and the third level to the polygon coordinates. The third level should have the format of [x0, y0, x1, y1, …, xn, yn] (n >= 3).

• list[ndarray]: each ndarray is a binary mask of shape (H, W).

• list[dict]: each dict is a COCO-style RLE.

• keypoints (Keypoint or array like) – an array-like object of shape (N, K, 3), where the N is the number of instances and K is the number of keypoints. The last dimension corresponds to (x, y, visibility or score).

• assigned_colors (list[matplotlib.colors]) – a list of colors, where each color corresponds to each mask or box in the image. Refer to ‘matplotlib.colors’ for full list of formats that the colors are accepted in.

Returns

output (VisImage) – image object with visualizations.

overlay_rotated_instances(boxes=None, labels=None, assigned_colors=None)[source]
Parameters
• boxes (ndarray) – an Nx5 numpy array of (x_center, y_center, width, height, angle_degrees) format for the N objects in a single image.

• labels (list[str]) – the text to be displayed for each instance.

• assigned_colors (list[matplotlib.colors]) – a list of colors, where each color corresponds to each mask or box in the image. Refer to ‘matplotlib.colors’ for full list of formats that the colors are accepted in.

Returns

output (VisImage) – image object with visualizations.

draw_and_connect_keypoints(keypoints)[source]

Draws keypoints of an instance and follows the rules for keypoint connections to draw lines between appropriate keypoints. This follows color heuristics for line color.

Parameters

keypoints (Tensor) – a tensor of shape (K, 3), where K is the number of keypoints and the last dimension corresponds to (x, y, probability).

Returns

output (VisImage) – image object with visualizations.

draw_text(text, position, *, font_size=None, color='g', horizontal_alignment='center', rotation=0)[source]
Parameters
• text (str) – class label

• position (tuple) – a tuple of the x and y coordinates to place text on image.

• font_size (int, optional) – font of the text. If not provided, a font size proportional to the image width is calculated and used.

• color – color of the text. Refer to matplotlib.colors for full list of formats that are accepted.

• horizontal_alignment (str) – see matplotlib.text.Text

• rotation – rotation angle in degrees CCW

Returns

output (VisImage) – image object with text drawn.

draw_box(box_coord, alpha=0.5, edge_color='g', line_style='-')[source]
Parameters
• box_coord (tuple) – a tuple containing x0, y0, x1, y1 coordinates, where x0 and y0 are the coordinates of the image’s top left corner. x1 and y1 are the coordinates of the image’s bottom right corner.

• alpha (float) – blending efficient. Smaller values lead to more transparent masks.

• edge_color – color of the outline of the box. Refer to matplotlib.colors for full list of formats that are accepted.

• line_style (string) – the string to use to create the outline of the boxes.

Returns

output (VisImage) – image object with box drawn.

draw_rotated_box_with_label(rotated_box, alpha=0.5, edge_color='g', line_style='-', label=None)[source]
Parameters
• rotated_box (tuple) – a tuple containing (cnt_x, cnt_y, w, h, angle), where cnt_x and cnt_y are the center coordinates of the box. w and h are the width and height of the box. angle represents how many degrees the box is rotated CCW with regard to the 0-degree box.

• alpha (float) – blending efficient. Smaller values lead to more transparent masks.

• edge_color – color of the outline of the box. Refer to matplotlib.colors for full list of formats that are accepted.

• line_style (string) – the string to use to create the outline of the boxes.

• label (string) – label for rotated box. It will not be rendered when set to None.

Returns

output (VisImage) – image object with box drawn.

draw_circle(circle_coord, color, radius=3)[source]
Parameters
• circle_coord (list(int) or tuple(int)) – contains the x and y coordinates of the center of the circle.

• color – color of the polygon. Refer to matplotlib.colors for a full list of formats that are accepted.

Returns

output (VisImage) – image object with box drawn.

draw_line(x_data, y_data, color, linestyle='-', linewidth=None)[source]
Parameters
• x_data (list[int]) – a list containing x values of all the points being drawn. Length of list should match the length of y_data.

• y_data (list[int]) – a list containing y values of all the points being drawn. Length of list should match the length of x_data.

• color – color of the line. Refer to matplotlib.colors for a full list of formats that are accepted.

• linestyle – style of the line. Refer to matplotlib.lines.Line2D for a full list of formats that are accepted.

• linewidth (float or None) – width of the line. When it’s None, a default value will be computed and used.

Returns

output (VisImage) – image object with line drawn.

draw_binary_mask(binary_mask, color=None, *, edge_color=None, text=None, alpha=0.5, area_threshold=0)[source]
Parameters
• binary_mask (ndarray) – numpy array of shape (H, W), where H is the image height and W is the image width. Each value in the array is either a 0 or 1 value of uint8 type.

• color – color of the mask. Refer to matplotlib.colors for a full list of formats that are accepted. If None, will pick a random color.

• edge_color – color of the polygon edges. Refer to matplotlib.colors for a full list of formats that are accepted.

• text (str) – if None, will be drawn in the object’s center of mass.

• alpha (float) – blending efficient. Smaller values lead to more transparent masks.

• area_threshold (float) – a connected component small than this will not be shown.

Returns

output (VisImage) – image object with mask drawn.

draw_polygon(segment, color, edge_color=None, alpha=0.5)[source]
Parameters
• segment – numpy array of shape Nx2, containing all the points in the polygon.

• color – color of the polygon. Refer to matplotlib.colors for a full list of formats that are accepted.

• edge_color – color of the polygon edges. Refer to matplotlib.colors for a full list of formats that are accepted. If not provided, a darker shade of the polygon color will be used instead.

• alpha (float) – blending efficient. Smaller values lead to more transparent masks.

Returns

output (VisImage) – image object with polygon drawn.

get_output()[source]
Returns

output (VisImage) – the image output containing the visualizations added to the image.

## detectron2.utils.video_visualizer module¶

class detectron2.utils.video_visualizer.VideoVisualizer(metadata, instance_mode=<ColorMode.IMAGE: 0>)[source]

Bases: object

__init__(metadata, instance_mode=<ColorMode.IMAGE: 0>)[source]
Parameters

draw_instance_predictions(frame, predictions)[source]

Draw instance-level prediction results on an image.

Parameters
• frame (ndarray) – an RGB image of shape (H, W, C), in the range [0, 255].

• predictions (Instances) – the output of an instance detection/segmentation model. Following fields will be used to draw: “pred_boxes”, “pred_classes”, “scores”, “pred_masks” (or “pred_masks_rle”).

Returns

output (VisImage) – image object with visualizations.

draw_sem_seg(frame, sem_seg, area_threshold=None)[source]
Parameters
• sem_seg (ndarray or Tensor) – semantic segmentation of shape (H, W), each value is the integer label.

• area_threshold (Optional[int]) – only draw segmentations larger than the threshold

draw_panoptic_seg_predictions(frame, panoptic_seg, segments_info, area_threshold=None, alpha=0.5)[source]