Command Line
On this page
Command Line#
This is a reference for all the --help message from all of OpenPifPaf’s command line interfaces (CLIs).
predict#
%%bash
python3 -m openpifpaf.predict --help
usage: python3 -m openpifpaf.predict [options] images
Predict poses for given images.
positional arguments:
images input images (default: None)
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--glob GLOB glob expression for input images (for many images)
(default: None)
-o [IMAGE_OUTPUT], --image-output [IMAGE_OUTPUT]
Whether to output an image, with the option to specify
the output path or directory (default: None)
--json-output [JSON_OUTPUT]
Whether to output a json file, with the option to
specify the output path or directory (default: None)
--disable-cuda disable CUDA (default: False)
decoder configuration:
--decoder DECODER [DECODER ...]
Decoders to be considered: ['posesimilarity',
'cifcafdense', 'trackingpose', 'cifcaf', 'cifdet'].
(default: None)
--seed-threshold SEED_THRESHOLD
minimum threshold for seeds (default: 0.2)
--instance-threshold INSTANCE_THRESHOLD
filter instances by score (default is 0.0 with
--force-complete-pose and 0.15 otherwise) (default:
None)
--decoder-workers DECODER_WORKERS
number of workers for pose decoding (default: None)
--profile-decoder [PROFILE_DECODER]
specify out .prof file or nothing for default file
name (default: None)
CifCaf decoders:
--cif-th CIF_TH cif threshold (default: 0.3)
--caf-th CAF_TH caf threshold (default: 0.3)
Decoder for tracking:
--tr-single-pose-threshold TR_SINGLE_POSE_THRESHOLD
Single-pose threshold for tracking. (default: 0.3)
--tr-multi-pose-threshold TR_MULTI_POSE_THRESHOLD
multi-pose threshold for tracking. (default: 0.2)
--tr-multi-pose-n TR_MULTI_POSE_N
multi-pose n for tracking. (default: 3)
--tr-minimum-threshold TR_MINIMUM_THRESHOLD
minimum-pose threshold for tracking. (default: 0.1)
PoseSimilarity:
--posesimilarity-distance {crafted,euclidean,euclidean4,oks}
--posesimilarity-oks-inflate POSESIMILARITY_OKS_INFLATE
CifCafDense decoder:
--dense-connections [DENSE_CONNECTIONS]
trackingpose decoder:
--trackingpose-track-recovery
--trackingpose-single-seed
CifCaf decoder:
--force-complete-pose
--force-complete-caf-th FORCE_COMPLETE_CAF_TH
CAF threshold for force complete. Set to -1 to
deactivate. (default: 0.001)
--nms-before-force-complete
run an additional NMS before completing poses
(default: False)
--keypoint-threshold KEYPOINT_THRESHOLD
filter keypoints by score (default: 0.15)
--keypoint-threshold-rel KEYPOINT_THRESHOLD_REL
filter keypoint connections by relative score
(default: 0.5)
--greedy greedy decoding (default: False)
--connection-method {max,blend}
connection method to use, max is faster (default:
blend)
--cifcaf-block-joints
block joints (default: False)
--no-reverse-match
--ablation-cifseeds-nms
--ablation-cifseeds-no-rescore
--ablation-caf-no-rescore
--ablation-independent-kp
logger:
-q, --quiet only show warning messages or above (default: False)
--debug print debug messages (default: False)
--log-stats enable stats logging (default: False)
MobileNetV3:
--mobilenetv3-no-pretrain
use randomly initialized models (default: True)
shufflenetv2k:
--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS
out channels of the optional 2nd input convolution
(default: None)
--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION
dilation factor of stage 4 (default: 1)
--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL
kernel width (default: 5)
--shufflenetv2k-conv5-as-stage
--shufflenetv2k-instance-norm
--shufflenetv2k-group-norm
--shufflenetv2k-leaky-relu
MobileNetV2:
--mobilenetv2-no-pretrain
use randomly initialized models (default: True)
ShuffleNetv2:
--shufflenetv2-no-pretrain
use randomly initialized models (default: True)
SwinTransformer:
--swin-drop-path-rate SWIN_DROP_PATH_RATE
drop path (stochastic depth) rate (default: 0.2)
--swin-input-upsample
scales input image by a factor of 2 for higher res
feature maps (default: False)
--swin-use-fpn adds a FPN after the Swin network to obtain higher res
feature maps (default: False)
--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS
output channels of the FPN (None to use the default
number of channels of the Swin network) (default:
None)
--swin-fpn-level SWIN_FPN_LEVEL
FPN pyramid level, must be between 1 (highest
resolution) and 4 (lowest resolution) (default: 3)
--swin-no-pretrain use randomly initialized models (default: True)
XCiT:
--xcit-out-channels XCIT_OUT_CHANNELS
number of output channels for optional projection
layer (None for no projection layer) (default: None)
--xcit-out-maxpool adds max-pooling to backbone output feature map
(default: False)
--xcit-no-pretrain use randomly initialized models (default: True)
ResNet:
--resnet-no-pretrain use randomly initialized models (default: True)
--resnet-pool0-stride RESNET_POOL0_STRIDE
stride of zero removes the pooling op (default: 0)
--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE
stride of the input convolution (default: 2)
--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--resnet-block5-dilation RESNET_BLOCK5_DILATION
use dilated convs in block5 (default: 1)
--resnet-remove-last-block
create a network without the last block (default:
False)
CompositeField4:
--cf4-dropout CF4_DROPOUT
[experimental] zeroing probability of feature in head
input (default: 0.0)
--cf4-no-inplace-ops alternative graph without inplace ops (default: True)
network configuration:
--checkpoint CHECKPOINT
Path to a local checkpoint. Or provide one of the
following to download a pretrained model:
['shufflenetv2k30-animalpose',
'shufflenetv2k16-apollo-24',
'shufflenetv2k16-apollo-66',
'shufflenetv2k30-apollo-66', 'mobilenetv2',
'mobilenetv3small', 'mobilenetv3large', 'resnet50',
'shufflenetv2k16', 'shufflenetv2k16-withdense',
'shufflenetv2k30', 'mobilenetv3small-cocodet',
'resnet18-cocodet', 'resnet50-crowdpose',
'shufflenetv2k16-nuscenes', 'tshufflenetv2k30',
'shufflenetv2k16-wholebody',
'shufflenetv2k30-wholebody', 'swin_s', 'swin_b',
'swin_t_input_upsample'] (default: None)
--basenet BASENET base network, one of ['mobilenetv2',
'mobilenetv3large', 'mobilenetv3small', 'resnet18',
'resnet50', 'resnet101', 'resnet152', 'resnext50',
'resnext101', 'shufflenetv2x1', 'shufflenetv2x2',
'shufflenetv2k16', 'shufflenetv2k20',
'shufflenetv2kx5', 'shufflenetv2k30',
'shufflenetv2k44', 'squeezenet', 'tshufflenetv2k16',
'tshufflenetv2k30', 'tresnet50', 'cifar10net',
'swin_t', 'swin_s', 'swin_b', 'swin_b_window_12',
'swin_l', 'swin_l_window_12', 'xcit_nano_12_p16',
'xcit_tiny_12_p16', 'xcit_tiny_24_p16',
'xcit_small_12_p16', 'xcit_small_24_p16',
'xcit_medium_24_p16', 'xcit_large_24_p16',
'xcit_nano_12_p8', 'xcit_tiny_12_p8',
'xcit_tiny_24_p8', 'xcit_small_12_p8',
'xcit_small_24_p8', 'xcit_medium_24_p8',
'xcit_large_24_p8', 'effnetv2_s', 'effnetv2_m',
'effnetv2_l', 'effnetv2_xl', 'effnetv2_s16_s',
'effnetv2_s16_m', 'effnetv2_s16_l', 'effnetv2_s16_xl',
'botnet'] (default: None)
--cross-talk CROSS_TALK
[experimental] (default: 0.0)
--no-download-progress
suppress model download progress bar (default: True)
--head-consolidation {keep,create,filter_and_extend}
consolidation strategy for a checkpoint's head
networks and the heads specified by the datamodule
(default: filter_and_extend)
Predictor:
--batch-size BATCH_SIZE
processing batch size (default: 1)
--loader-workers LOADER_WORKERS
number of workers for data loading (default: None)
--long-edge LONG_EDGE
rescale the long side of the image (aspect ratio
maintained) (default: None)
--precise-rescaling use more exact image rescaling (requires scipy)
(default: True)
show:
--save-all [SAVE_ALL]
every plot is saved (optional to specify directory)
(default: None)
--show show every plot, i.e., call matplotlib show()
(default: False)
--image-width IMAGE_WIDTH
image width for matplotlib (in inches) (default: None)
--image-height IMAGE_HEIGHT
image height for matplotlib (in inches) (default:
None)
--image-dpi-factor IMAGE_DPI_FACTOR
increase dpi of output image by this factor (default:
2.0)
--image-min-dpi IMAGE_MIN_DPI
minimum dpi of image output (default: 50.0)
--show-file-extension SHOW_FILE_EXTENSION
default file extension (default: jpeg)
--textbox-alpha TEXTBOX_ALPHA
transparency of annotation text box (default: 0.5)
--text-color TEXT_COLOR
annotation text color (default: white)
--font-size FONT_SIZE
annotation font size (default: 8)
--monocolor-connections
use a single color per instance (default: False)
--line-width LINE_WIDTH
skeleton line width (default: None)
--skeleton-solid-threshold SKELETON_SOLID_THRESHOLD
set to 0.0 to draw all connections as solid lines
(default: 0.5)
--show-box show annotation bounding boxes (default: False)
--white-overlay [WHITE_OVERLAY]
increase contrast to annotations by making image
whiter (default: False)
--show-joint-scales show boxes representing joint sizes (default: False)
--show-joint-confidences
print per-joint confidences on skeleton annotations
(default: False)
--show-decoding-order
--show-frontier-order
--show-only-decoded-connections
to debug which connections were used for decoding
(default: False)
--video-fps VIDEO_FPS
output video frame rate (frames per second) (default:
10)
--video-dpi VIDEO_DPI
output video resolution (dots per inch) (default: 100)
visualizer:
--debug-indices DEBUG_INDICES [DEBUG_INDICES ...]
Indices of fields to create debug plots for of the
form headname:fieldindex, e.g. cif:5. Optionally,
specify the visualization type, e.g. cif:5:hr for the
high resolution plot only. Use comma separation to
specify multiple head names, field indices or
visualization types, e.g. cif:5,6:confidence,hr to
visualize CIF fields 5 and 6 but only show confidence
and hr. (default: [])
video#
%%bash
python3 -m openpifpaf.video --help
usage: python3 -m openpifpaf.video [options]
Video demo application.
Use --scale=0.2 to reduce the input image size to 20%.
Use --json-output for headless processing.
Example commands:
python3 -m pifpaf.video --source=0 # default webcam
python3 -m pifpaf.video --source=1 # another webcam
# streaming source
python3 -m pifpaf.video --source=http://127.0.0.1:8080/video
# file system source (any valid OpenCV source)
python3 -m pifpaf.video --source=docs/coco/000000081988.jpg
Trouble shooting:
* MacOSX: try to prefix the command with "MPLBACKEND=MACOSX".
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--source SOURCE OpenCV source url. Integer for webcams. Or ipwebcam
urls (rtsp/rtmp). Use "screen" for screen grabs.
(default: 0)
--video-output [VIDEO_OUTPUT]
video output file or "virtualcam" (default: None)
--json-output [JSON_OUTPUT]
json output file (default: None)
--separate-debug-ax
--disable-cuda disable CUDA (default: False)
ResNet:
--resnet-no-pretrain use randomly initialized models (default: True)
--resnet-pool0-stride RESNET_POOL0_STRIDE
stride of zero removes the pooling op (default: 0)
--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE
stride of the input convolution (default: 2)
--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--resnet-block5-dilation RESNET_BLOCK5_DILATION
use dilated convs in block5 (default: 1)
--resnet-remove-last-block
create a network without the last block (default:
False)
ShuffleNetv2:
--shufflenetv2-no-pretrain
use randomly initialized models (default: True)
SwinTransformer:
--swin-drop-path-rate SWIN_DROP_PATH_RATE
drop path (stochastic depth) rate (default: 0.2)
--swin-input-upsample
scales input image by a factor of 2 for higher res
feature maps (default: False)
--swin-use-fpn adds a FPN after the Swin network to obtain higher res
feature maps (default: False)
--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS
output channels of the FPN (None to use the default
number of channels of the Swin network) (default:
None)
--swin-fpn-level SWIN_FPN_LEVEL
FPN pyramid level, must be between 1 (highest
resolution) and 4 (lowest resolution) (default: 3)
--swin-no-pretrain use randomly initialized models (default: True)
XCiT:
--xcit-out-channels XCIT_OUT_CHANNELS
number of output channels for optional projection
layer (None for no projection layer) (default: None)
--xcit-out-maxpool adds max-pooling to backbone output feature map
(default: False)
--xcit-no-pretrain use randomly initialized models (default: True)
MobileNetV3:
--mobilenetv3-no-pretrain
use randomly initialized models (default: True)
MobileNetV2:
--mobilenetv2-no-pretrain
use randomly initialized models (default: True)
shufflenetv2k:
--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS
out channels of the optional 2nd input convolution
(default: None)
--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION
dilation factor of stage 4 (default: 1)
--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL
kernel width (default: 5)
--shufflenetv2k-conv5-as-stage
--shufflenetv2k-instance-norm
--shufflenetv2k-group-norm
--shufflenetv2k-leaky-relu
CompositeField4:
--cf4-dropout CF4_DROPOUT
[experimental] zeroing probability of feature in head
input (default: 0.0)
--cf4-no-inplace-ops alternative graph without inplace ops (default: True)
network configuration:
--checkpoint CHECKPOINT
Path to a local checkpoint. Or provide one of the
following to download a pretrained model:
['shufflenetv2k30-animalpose',
'shufflenetv2k16-apollo-24',
'shufflenetv2k16-apollo-66',
'shufflenetv2k30-apollo-66', 'mobilenetv2',
'mobilenetv3small', 'mobilenetv3large', 'resnet50',
'shufflenetv2k16', 'shufflenetv2k16-withdense',
'shufflenetv2k30', 'mobilenetv3small-cocodet',
'resnet18-cocodet', 'resnet50-crowdpose',
'shufflenetv2k16-nuscenes', 'tshufflenetv2k30',
'shufflenetv2k16-wholebody',
'shufflenetv2k30-wholebody', 'swin_s', 'swin_b',
'swin_t_input_upsample'] (default: None)
--basenet BASENET base network, one of ['mobilenetv2',
'mobilenetv3large', 'mobilenetv3small', 'resnet18',
'resnet50', 'resnet101', 'resnet152', 'resnext50',
'resnext101', 'shufflenetv2x1', 'shufflenetv2x2',
'shufflenetv2k16', 'shufflenetv2k20',
'shufflenetv2kx5', 'shufflenetv2k30',
'shufflenetv2k44', 'squeezenet', 'tshufflenetv2k16',
'tshufflenetv2k30', 'tresnet50', 'cifar10net',
'swin_t', 'swin_s', 'swin_b', 'swin_b_window_12',
'swin_l', 'swin_l_window_12', 'xcit_nano_12_p16',
'xcit_tiny_12_p16', 'xcit_tiny_24_p16',
'xcit_small_12_p16', 'xcit_small_24_p16',
'xcit_medium_24_p16', 'xcit_large_24_p16',
'xcit_nano_12_p8', 'xcit_tiny_12_p8',
'xcit_tiny_24_p8', 'xcit_small_12_p8',
'xcit_small_24_p8', 'xcit_medium_24_p8',
'xcit_large_24_p8', 'effnetv2_s', 'effnetv2_m',
'effnetv2_l', 'effnetv2_xl', 'effnetv2_s16_s',
'effnetv2_s16_m', 'effnetv2_s16_l', 'effnetv2_s16_xl',
'botnet'] (default: None)
--cross-talk CROSS_TALK
[experimental] (default: 0.0)
--no-download-progress
suppress model download progress bar (default: True)
--head-consolidation {keep,create,filter_and_extend}
consolidation strategy for a checkpoint's head
networks and the heads specified by the datamodule
(default: filter_and_extend)
decoder configuration:
--decoder DECODER [DECODER ...]
Decoders to be considered: ['cifcafdense',
'trackingpose', 'cifcaf', 'posesimilarity', 'cifdet'].
(default: None)
--seed-threshold SEED_THRESHOLD
minimum threshold for seeds (default: 0.2)
--instance-threshold INSTANCE_THRESHOLD
filter instances by score (default is 0.0 with
--force-complete-pose and 0.15 otherwise) (default:
None)
--decoder-workers DECODER_WORKERS
number of workers for pose decoding (default: None)
--profile-decoder [PROFILE_DECODER]
specify out .prof file or nothing for default file
name (default: None)
CifCaf decoders:
--cif-th CIF_TH cif threshold (default: 0.3)
--caf-th CAF_TH caf threshold (default: 0.3)
Decoder for tracking:
--tr-single-pose-threshold TR_SINGLE_POSE_THRESHOLD
Single-pose threshold for tracking. (default: 0.3)
--tr-multi-pose-threshold TR_MULTI_POSE_THRESHOLD
multi-pose threshold for tracking. (default: 0.2)
--tr-multi-pose-n TR_MULTI_POSE_N
multi-pose n for tracking. (default: 3)
--tr-minimum-threshold TR_MINIMUM_THRESHOLD
minimum-pose threshold for tracking. (default: 0.1)
CifCafDense decoder:
--dense-connections [DENSE_CONNECTIONS]
trackingpose decoder:
--trackingpose-track-recovery
--trackingpose-single-seed
CifCaf decoder:
--force-complete-pose
--force-complete-caf-th FORCE_COMPLETE_CAF_TH
CAF threshold for force complete. Set to -1 to
deactivate. (default: 0.001)
--nms-before-force-complete
run an additional NMS before completing poses
(default: False)
--keypoint-threshold KEYPOINT_THRESHOLD
filter keypoints by score (default: 0.15)
--keypoint-threshold-rel KEYPOINT_THRESHOLD_REL
filter keypoint connections by relative score
(default: 0.5)
--greedy greedy decoding (default: False)
--connection-method {max,blend}
connection method to use, max is faster (default:
blend)
--cifcaf-block-joints
block joints (default: False)
--no-reverse-match
--ablation-cifseeds-nms
--ablation-cifseeds-no-rescore
--ablation-caf-no-rescore
--ablation-independent-kp
PoseSimilarity:
--posesimilarity-distance {crafted,euclidean,euclidean4,oks}
--posesimilarity-oks-inflate POSESIMILARITY_OKS_INFLATE
logger:
-q, --quiet only show warning messages or above (default: False)
--debug print debug messages (default: False)
--log-stats enable stats logging (default: False)
Predictor:
--batch-size BATCH_SIZE
processing batch size (default: 1)
--loader-workers LOADER_WORKERS
number of workers for data loading (default: None)
--long-edge LONG_EDGE
rescale the long side of the image (aspect ratio
maintained) (default: None)
--precise-rescaling use more exact image rescaling (requires scipy)
(default: True)
show:
--save-all [SAVE_ALL]
every plot is saved (optional to specify directory)
(default: None)
--show show every plot, i.e., call matplotlib show()
(default: False)
--image-width IMAGE_WIDTH
image width for matplotlib (in inches) (default: None)
--image-height IMAGE_HEIGHT
image height for matplotlib (in inches) (default:
None)
--image-dpi-factor IMAGE_DPI_FACTOR
increase dpi of output image by this factor (default:
2.0)
--image-min-dpi IMAGE_MIN_DPI
minimum dpi of image output (default: 50.0)
--show-file-extension SHOW_FILE_EXTENSION
default file extension (default: jpeg)
--textbox-alpha TEXTBOX_ALPHA
transparency of annotation text box (default: 0.5)
--text-color TEXT_COLOR
annotation text color (default: white)
--font-size FONT_SIZE
annotation font size (default: 8)
--monocolor-connections
use a single color per instance (default: False)
--line-width LINE_WIDTH
skeleton line width (default: None)
--skeleton-solid-threshold SKELETON_SOLID_THRESHOLD
set to 0.0 to draw all connections as solid lines
(default: 0.5)
--show-box show annotation bounding boxes (default: False)
--white-overlay [WHITE_OVERLAY]
increase contrast to annotations by making image
whiter (default: False)
--show-joint-scales show boxes representing joint sizes (default: False)
--show-joint-confidences
print per-joint confidences on skeleton annotations
(default: False)
--show-decoding-order
--show-frontier-order
--show-only-decoded-connections
to debug which connections were used for decoding
(default: False)
--video-fps VIDEO_FPS
output video frame rate (frames per second) (default:
10)
--video-dpi VIDEO_DPI
output video resolution (dots per inch) (default: 100)
Stream:
--horizontal-flip mirror input image (default: False)
--scale SCALE input image scale factor (default: 1.0)
--start-frame START_FRAME
start frame (default: None)
--start-msec START_MSEC
start millisecond (default: None)
--crop CROP CROP CROP CROP
left top right bottom (default: None)
--rotate {left,right,180}
rotate (default: None)
--max-frames MAX_FRAMES
max frames (default: None)
visualizer:
--debug-indices DEBUG_INDICES [DEBUG_INDICES ...]
Indices of fields to create debug plots for of the
form headname:fieldindex, e.g. cif:5. Optionally,
specify the visualization type, e.g. cif:5:hr for the
high resolution plot only. Use comma separation to
specify multiple head names, field indices or
visualization types, e.g. cif:5,6:confidence,hr to
visualize CIF fields 5 and 6 but only show confidence
and hr. (default: [])
train#
%%bash
python3 -m openpifpaf.train --help
usage: python3 -m openpifpaf.train [options]
Train a pifpaf network.
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
-o OUTPUT, --output OUTPUT
output file (default: None)
--disable-cuda disable CUDA (default: False)
--ddp [experimental] DistributedDataParallel (default:
False)
--local_rank LOCAL_RANK
[experimental] for torch.distributed.launch (default:
None)
--no-sync-batchnorm [experimental] in ddp, to not use syncbatchnorm
(default: True)
logger:
-q, --quiet only show warning messages or above (default: False)
--debug print debug messages (default: False)
--log-stats enable stats logging (default: False)
SwinTransformer:
--swin-drop-path-rate SWIN_DROP_PATH_RATE
drop path (stochastic depth) rate (default: 0.2)
--swin-input-upsample
scales input image by a factor of 2 for higher res
feature maps (default: False)
--swin-use-fpn adds a FPN after the Swin network to obtain higher res
feature maps (default: False)
--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS
output channels of the FPN (None to use the default
number of channels of the Swin network) (default:
None)
--swin-fpn-level SWIN_FPN_LEVEL
FPN pyramid level, must be between 1 (highest
resolution) and 4 (lowest resolution) (default: 3)
--swin-no-pretrain use randomly initialized models (default: True)
XCiT:
--xcit-out-channels XCIT_OUT_CHANNELS
number of output channels for optional projection
layer (None for no projection layer) (default: None)
--xcit-out-maxpool adds max-pooling to backbone output feature map
(default: False)
--xcit-no-pretrain use randomly initialized models (default: True)
MobileNetV3:
--mobilenetv3-no-pretrain
use randomly initialized models (default: True)
shufflenetv2k:
--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS
out channels of the optional 2nd input convolution
(default: None)
--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION
dilation factor of stage 4 (default: 1)
--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL
kernel width (default: 5)
--shufflenetv2k-conv5-as-stage
--shufflenetv2k-instance-norm
--shufflenetv2k-group-norm
--shufflenetv2k-leaky-relu
MobileNetV2:
--mobilenetv2-no-pretrain
use randomly initialized models (default: True)
ShuffleNetv2:
--shufflenetv2-no-pretrain
use randomly initialized models (default: True)
ResNet:
--resnet-no-pretrain use randomly initialized models (default: True)
--resnet-pool0-stride RESNET_POOL0_STRIDE
stride of zero removes the pooling op (default: 0)
--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE
stride of the input convolution (default: 2)
--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--resnet-block5-dilation RESNET_BLOCK5_DILATION
use dilated convs in block5 (default: 1)
--resnet-remove-last-block
create a network without the last block (default:
False)
CompositeField4:
--cf4-dropout CF4_DROPOUT
[experimental] zeroing probability of feature in head
input (default: 0.0)
--cf4-no-inplace-ops alternative graph without inplace ops (default: True)
network configuration:
--checkpoint CHECKPOINT
Path to a local checkpoint. Or provide one of the
following to download a pretrained model:
['shufflenetv2k30-animalpose',
'shufflenetv2k16-apollo-24',
'shufflenetv2k16-apollo-66',
'shufflenetv2k30-apollo-66', 'mobilenetv2',
'mobilenetv3small', 'mobilenetv3large', 'resnet50',
'shufflenetv2k16', 'shufflenetv2k16-withdense',
'shufflenetv2k30', 'mobilenetv3small-cocodet',
'resnet18-cocodet', 'resnet50-crowdpose',
'shufflenetv2k16-nuscenes', 'tshufflenetv2k30',
'shufflenetv2k16-wholebody',
'shufflenetv2k30-wholebody', 'swin_s', 'swin_b',
'swin_t_input_upsample'] (default: None)
--basenet BASENET base network, one of ['mobilenetv2',
'mobilenetv3large', 'mobilenetv3small', 'resnet18',
'resnet50', 'resnet101', 'resnet152', 'resnext50',
'resnext101', 'shufflenetv2x1', 'shufflenetv2x2',
'shufflenetv2k16', 'shufflenetv2k20',
'shufflenetv2kx5', 'shufflenetv2k30',
'shufflenetv2k44', 'squeezenet', 'tshufflenetv2k16',
'tshufflenetv2k30', 'tresnet50', 'cifar10net',
'swin_t', 'swin_s', 'swin_b', 'swin_b_window_12',
'swin_l', 'swin_l_window_12', 'xcit_nano_12_p16',
'xcit_tiny_12_p16', 'xcit_tiny_24_p16',
'xcit_small_12_p16', 'xcit_small_24_p16',
'xcit_medium_24_p16', 'xcit_large_24_p16',
'xcit_nano_12_p8', 'xcit_tiny_12_p8',
'xcit_tiny_24_p8', 'xcit_small_12_p8',
'xcit_small_24_p8', 'xcit_medium_24_p8',
'xcit_large_24_p8', 'effnetv2_s', 'effnetv2_m',
'effnetv2_l', 'effnetv2_xl', 'effnetv2_s16_s',
'effnetv2_s16_m', 'effnetv2_s16_l', 'effnetv2_s16_xl',
'botnet'] (default: None)
--cross-talk CROSS_TALK
[experimental] (default: 0.0)
--no-download-progress
suppress model download progress bar (default: True)
--head-consolidation {keep,create,filter_and_extend}
consolidation strategy for a checkpoint's head
networks and the heads specified by the datamodule
(default: filter_and_extend)
losses:
--lambdas LAMBDAS [LAMBDAS ...]
prefactor for head losses by head (default: None)
--component-lambdas COMPONENT_LAMBDAS [COMPONENT_LAMBDAS ...]
prefactor for head losses by component (default: None)
--auto-tune-mtl [experimental] use Kendall's prescription for
adjusting the multitask weight (default: False)
--auto-tune-mtl-variance
[experimental] use Variance prescription for adjusting
the multitask weight (default: False)
--task-sparsity-weight TASK_SPARSITY_WEIGHT
[experimental] (default: 0.0)
Scale Loss:
--b-scale B_SCALE Laplace width b for scale loss (default: 1.0)
--scale-log
--scale-soft-clamp SCALE_SOFT_CLAMP
soft clamp for scale (default: 5.0)
Regression loss:
--regression-soft-clamp REGRESSION_SOFT_CLAMP
soft clamp for scale (default: 5.0)
Bce Loss:
--focal-alpha FOCAL_ALPHA
scale parameter of focal loss (default: 0.5)
--focal-gamma FOCAL_GAMMA
use focal loss with the given gamma (default: 1.0)
--bce-soft-clamp BCE_SOFT_CLAMP
soft clamp for BCE (default: 5.0)
--bce-background-clamp BCE_BACKGROUND_CLAMP
background clamp for BCE (default: -15.0)
trainer:
--epochs EPOCHS number of epochs to train (default: None)
--train-batches TRAIN_BATCHES
number of train batches (default: None)
--val-batches VAL_BATCHES
number of val batches (default: None)
--clip-grad-norm CLIP_GRAD_NORM
clip grad norm: specify largest change for single
param (default: 0.0)
--clip-grad-value CLIP_GRAD_VALUE
clip grad value: specify largest change for single
param (default: 0.0)
--log-interval LOG_INTERVAL
log loss every n steps (default: 11)
--val-interval VAL_INTERVAL
validation run every n epochs (default: 1)
--stride-apply STRIDE_APPLY
apply and reset gradients every n batches (default: 1)
--fix-batch-norm [FIX_BATCH_NORM]
fix batch norm running statistics (optionally specify
epoch) (default: False)
--ema EMA ema decay constant (default: 0.01)
--profile PROFILE enables profiling. specify path for chrome tracing
file (default: None)
encoders:
--cif-side-length CIF_SIDE_LENGTH
side length of the CIF field (default: 4)
--caf-min-size CAF_MIN_SIZE
min side length of the CAF field (default: 3)
--caf-fixed-size fixed caf size (default: False)
--caf-aspect-ratio CAF_ASPECT_RATIO
CAF width relative to its length (default: 0.0)
--encoder-no-suppress-selfhidden
[experimental] (default: True)
--encoder-suppress-invisible
[experimental] (default: False)
--encoder-suppress-collision
[experimental] (default: False)
optimizer:
--momentum MOMENTUM SGD momentum, beta1 in Adam (default: 0.9)
--beta2 BETA2 beta2 for Adam/AMSGrad (default: 0.999)
--adam-eps ADAM_EPS eps value for Adam/AMSGrad (default: 1e-06)
--no-nesterov do not use Nesterov momentum for SGD update (default:
True)
--weight-decay WEIGHT_DECAY
SGD/Adam/AMSGrad weight decay (default: 0.0)
--adam use Adam optimizer (default: False)
--amsgrad use Adam optimizer with AMSGrad option (default:
False)
learning rate scheduler:
--lr LR learning rate (default: 0.001)
--lr-decay LR_DECAY [LR_DECAY ...]
epochs at which to decay the learning rate (default:
[])
--lr-decay-factor LR_DECAY_FACTOR
learning rate decay factor (default: 0.1)
--lr-decay-epochs LR_DECAY_EPOCHS
learning rate decay duration in epochs (default: 1.0)
--lr-warm-up-start-epoch LR_WARM_UP_START_EPOCH
starting epoch for warm-up (default: 0)
--lr-warm-up-epochs LR_WARM_UP_EPOCHS
number of epochs at the beginning with lower learning
rate (default: 1)
--lr-warm-up-factor LR_WARM_UP_FACTOR
learning pre-factor during warm-up (default: 0.001)
--lr-warm-restarts LR_WARM_RESTARTS [LR_WARM_RESTARTS ...]
list of epochs to do a warm restart (default: [])
--lr-warm-restart-duration LR_WARM_RESTART_DURATION
duration of a warm restart (default: 0.5)
generic data module parameters:
--dataset DATASET
--loader-workers LOADER_WORKERS
number of workers for data loading (default: None)
--batch-size BATCH_SIZE
batch size (default: 1)
--dataset-weights DATASET_WEIGHTS [DATASET_WEIGHTS ...]
n-1 weights for the datasets (default: None)
data module Animal:
--animal-train-annotations ANIMAL_TRAIN_ANNOTATIONS
--animal-val-annotations ANIMAL_VAL_ANNOTATIONS
--animal-train-image-dir ANIMAL_TRAIN_IMAGE_DIR
--animal-val-image-dir ANIMAL_VAL_IMAGE_DIR
--animal-square-edge ANIMAL_SQUARE_EDGE
square edge of input images (default: 513)
--animal-extended-scale
augment with an extended scale range (default: False)
--animal-orientation-invariant ANIMAL_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--animal-blur ANIMAL_BLUR
augment with blur (default: 0.0)
--animal-no-augmentation
do not apply data augmentation (default: True)
--animal-rescale-images ANIMAL_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--animal-upsample ANIMAL_UPSAMPLE
head upsample stride (default: 1)
--animal-min-kp-anns ANIMAL_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--animal-bmin ANIMAL_BMIN
b minimum in pixels (default: 1)
--animal-eval-test2017
--animal-eval-testdev2017
--animal-no-eval-annotation-filter
--animal-eval-long-edge ANIMAL_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 0)
--animal-eval-extended-scale
--animal-eval-orientation-invariant ANIMAL_EVAL_ORIENTATION_INVARIANT
data module Apollo:
--apollo-train-annotations APOLLO_TRAIN_ANNOTATIONS
--apollo-val-annotations APOLLO_VAL_ANNOTATIONS
--apollo-train-image-dir APOLLO_TRAIN_IMAGE_DIR
--apollo-val-image-dir APOLLO_VAL_IMAGE_DIR
--apollo-square-edge APOLLO_SQUARE_EDGE
square edge of input images (default: 513)
--apollo-extended-scale
augment with an extended scale range (default: False)
--apollo-orientation-invariant APOLLO_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--apollo-blur APOLLO_BLUR
augment with blur (default: 0.0)
--apollo-no-augmentation
do not apply data augmentation (default: True)
--apollo-rescale-images APOLLO_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--apollo-upsample APOLLO_UPSAMPLE
head upsample stride (default: 1)
--apollo-min-kp-anns APOLLO_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--apollo-bmin APOLLO_BMIN
b minimum in pixels (default: 1)
--apollo-apply-local-centrality-weights
Weigh the CIF and CAF head during training. (default:
False)
--apollo-no-eval-annotation-filter
--apollo-eval-long-edge APOLLO_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 0)
--apollo-eval-extended-scale
--apollo-eval-orientation-invariant APOLLO_EVAL_ORIENTATION_INVARIANT
--apollo-use-24-kps The ApolloCar3D dataset can be trained with 24 or 66
kps. If you want to train a model with 24 kps activate
this flag. Change the annotations path to the json
files with 24 kps. (default: False)
data module Cifar10:
--cifar10-root-dir CIFAR10_ROOT_DIR
--cifar10-download
data module CocoDet:
--cocodet-train-annotations COCODET_TRAIN_ANNOTATIONS
--cocodet-val-annotations COCODET_VAL_ANNOTATIONS
--cocodet-train-image-dir COCODET_TRAIN_IMAGE_DIR
--cocodet-val-image-dir COCODET_VAL_IMAGE_DIR
--cocodet-square-edge COCODET_SQUARE_EDGE
square edge of input images (default: 513)
--cocodet-extended-scale
augment with an extended scale range (default: False)
--cocodet-orientation-invariant COCODET_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--cocodet-blur COCODET_BLUR
augment with blur (default: 0.0)
--cocodet-no-augmentation
do not apply data augmentation (default: True)
--cocodet-rescale-images COCODET_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--cocodet-upsample COCODET_UPSAMPLE
head upsample stride (default: 1)
data module CocoKp:
--cocokp-train-annotations COCOKP_TRAIN_ANNOTATIONS
train annotations (default: data-
mscoco/annotations/person_keypoints_train2017.json)
--cocokp-val-annotations COCOKP_VAL_ANNOTATIONS
val annotations (default: data-
mscoco/annotations/person_keypoints_val2017.json)
--cocokp-train-image-dir COCOKP_TRAIN_IMAGE_DIR
train image dir (default: data-
mscoco/images/train2017/)
--cocokp-val-image-dir COCOKP_VAL_IMAGE_DIR
val image dir (default: data-mscoco/images/val2017/)
--cocokp-square-edge COCOKP_SQUARE_EDGE
square edge of input images (default: 385)
--cocokp-with-dense train with dense connections (default: False)
--cocokp-extended-scale
augment with an extended scale range (default: False)
--cocokp-orientation-invariant COCOKP_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--cocokp-blur COCOKP_BLUR
augment with blur (default: 0.0)
--cocokp-no-augmentation
do not apply data augmentation (default: True)
--cocokp-rescale-images COCOKP_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--cocokp-upsample COCOKP_UPSAMPLE
head upsample stride (default: 1)
--cocokp-min-kp-anns COCOKP_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--cocokp-bmin COCOKP_BMIN
bmin (default: 0.1)
--cocokp-eval-test2017
--cocokp-eval-testdev2017
--coco-no-eval-annotation-filter
--coco-eval-long-edge COCO_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 641)
--coco-eval-extended-scale
--coco-eval-orientation-invariant COCO_EVAL_ORIENTATION_INVARIANT
data module CrowdPose:
--crowdpose-train-annotations CROWDPOSE_TRAIN_ANNOTATIONS
--crowdpose-val-annotations CROWDPOSE_VAL_ANNOTATIONS
--crowdpose-image-dir CROWDPOSE_IMAGE_DIR
--crowdpose-square-edge CROWDPOSE_SQUARE_EDGE
square edge of input images (default: 385)
--crowdpose-extended-scale
augment with an extended scale range (default: False)
--crowdpose-orientation-invariant CROWDPOSE_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--crowdpose-no-augmentation
do not apply data augmentation (default: True)
--crowdpose-rescale-images CROWDPOSE_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--crowdpose-upsample CROWDPOSE_UPSAMPLE
head upsample stride (default: 1)
--crowdpose-min-kp-anns CROWDPOSE_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--crowdpose-eval-test
--crowdpose-eval-long-edge CROWDPOSE_EVAL_LONG_EDGE
--crowdpose-eval-extended-scale
--crowdpose-eval-orientation-invariant CROWDPOSE_EVAL_ORIENTATION_INVARIANT
--crowdpose-index {easy,medium,hard}
data module NuScenes:
--nuscenes-train-annotations NUSCENES_TRAIN_ANNOTATIONS
--nuscenes-val-annotations NUSCENES_VAL_ANNOTATIONS
--nuscenes-train-image-dir NUSCENES_TRAIN_IMAGE_DIR
--nuscenes-val-image-dir NUSCENES_VAL_IMAGE_DIR
--nuscenes-square-edge NUSCENES_SQUARE_EDGE
square edge of input images (default: 513)
--nuscenes-extended-scale
augment with an extended scale range (default: False)
--nuscenes-orientation-invariant NUSCENES_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--nuscenes-blur NUSCENES_BLUR
augment with blur (default: 0.0)
--nuscenes-no-augmentation
do not apply data augmentation (default: True)
--nuscenes-rescale-images NUSCENES_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--nuscenes-upsample NUSCENES_UPSAMPLE
head upsample stride (default: 1)
data module Posetrack2018:
--posetrack2018-train-annotations POSETRACK2018_TRAIN_ANNOTATIONS
train annotations (default: data-
posetrack2018/annotations/train/*.json)
--posetrack2018-val-annotations POSETRACK2018_VAL_ANNOTATIONS
val annotations (default: data-
posetrack2018/annotations/val/*.json)
--posetrack2018-eval-annotations POSETRACK2018_EVAL_ANNOTATIONS
eval annotations (default: data-
posetrack2018/annotations/val/*.json)
--posetrack2018-data-root POSETRACK2018_DATA_ROOT
data root (default: data-posetrack2018)
data module Posetrack:
--posetrack-square-edge POSETRACK_SQUARE_EDGE
square edge of input images (default: 385)
--posetrack-with-dense
train with dense connections (default: False)
--posetrack-no-augmentation
do not apply data augmentation (default: True)
--posetrack-rescale-images POSETRACK_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--posetrack-upsample POSETRACK_UPSAMPLE
head upsample stride (default: 1)
--posetrack-min-kp-anns POSETRACK_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--posetrack-bmin POSETRACK_BMIN
--posetrack-sample-pairing POSETRACK_SAMPLE_PAIRING
blend two samples together (default: 0.0)
--posetrack-image-augmentations POSETRACK_IMAGE_AUGMENTATIONS
autocontrast, equalize, invert, solarize (default:
0.0)
--posetrack-max-shift POSETRACK_MAX_SHIFT
max shift (default: 30.0)
--posetrack-eval-long-edge POSETRACK_EVAL_LONG_EDGE
--posetrack-eval-extended-scale
--posetrack-eval-orientation-invariant POSETRACK_EVAL_ORIENTATION_INVARIANT
--posetrack-ablation-without-tcaf
data module Posetrack2017:
--posetrack2017-eval-annotations POSETRACK2017_EVAL_ANNOTATIONS
eval annotations (default: data-
posetrack2017/annotations/val/*.json)
--posetrack2017-data-root POSETRACK2017_DATA_ROOT
data root (default: data-posetrack2017)
data module CocoKpSt:
--cocokpst-max-shift COCOKPST_MAX_SHIFT
max shift (default: 30.0)
data module wholebody:
--wholebody-train-annotations WHOLEBODY_TRAIN_ANNOTATIONS
train annotations (default: data-mscoco/annotations/pe
rson_keypoints_train2017_wholebody_pifpaf_style.json)
--wholebody-val-annotations WHOLEBODY_VAL_ANNOTATIONS
val annotations (default: data-
mscoco/annotations/coco_wholebody_val_v1.0.json)
--wholebody-train-image-dir WHOLEBODY_TRAIN_IMAGE_DIR
train image dir (default: data-
mscoco/images/train2017/)
--wholebody-val-image-dir WHOLEBODY_VAL_IMAGE_DIR
val image dir (default: data-mscoco/images/val2017)
--wholebody-square-edge WHOLEBODY_SQUARE_EDGE
square edge of input images (default: 385)
--wholebody-extended-scale
augment with an extended scale range (default: False)
--wholebody-orientation-invariant WHOLEBODY_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--wholebody-blur WHOLEBODY_BLUR
augment with blur (default: 0.0)
--wholebody-no-augmentation
do not apply data augmentation (default: True)
--wholebody-rescale-images WHOLEBODY_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--wholebody-upsample WHOLEBODY_UPSAMPLE
head upsample stride (default: 1)
--wholebody-min-kp-anns WHOLEBODY_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--wholebody-bmin WHOLEBODY_BMIN
bmin (default: 1.0)
--wholebody-apply-local-centrality-weights
Weigh the CIF and CAF head during training. (default:
False)
--wholebody-eval-test2017
--wholebody-eval-testdev2017
--wholebody-no-eval-annotation-filter
--wholebody-eval-long-edge WHOLEBODY_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 641)
--wholebody-eval-extended-scale
--wholebody-eval-orientation-invariant WHOLEBODY_EVAL_ORIENTATION_INVARIANT
show:
--save-all [SAVE_ALL]
every plot is saved (optional to specify directory)
(default: None)
--show show every plot, i.e., call matplotlib show()
(default: False)
--image-width IMAGE_WIDTH
image width for matplotlib (in inches) (default: None)
--image-height IMAGE_HEIGHT
image height for matplotlib (in inches) (default:
None)
--image-dpi-factor IMAGE_DPI_FACTOR
increase dpi of output image by this factor (default:
2.0)
--image-min-dpi IMAGE_MIN_DPI
minimum dpi of image output (default: 50.0)
--show-file-extension SHOW_FILE_EXTENSION
default file extension (default: jpeg)
--textbox-alpha TEXTBOX_ALPHA
transparency of annotation text box (default: 0.5)
--text-color TEXT_COLOR
annotation text color (default: white)
--font-size FONT_SIZE
annotation font size (default: 8)
--monocolor-connections
use a single color per instance (default: False)
--line-width LINE_WIDTH
skeleton line width (default: None)
--skeleton-solid-threshold SKELETON_SOLID_THRESHOLD
set to 0.0 to draw all connections as solid lines
(default: 0.5)
--show-box show annotation bounding boxes (default: False)
--white-overlay [WHITE_OVERLAY]
increase contrast to annotations by making image
whiter (default: False)
--show-joint-scales show boxes representing joint sizes (default: False)
--show-joint-confidences
print per-joint confidences on skeleton annotations
(default: False)
--show-decoding-order
--show-frontier-order
--show-only-decoded-connections
to debug which connections were used for decoding
(default: False)
--video-fps VIDEO_FPS
output video frame rate (frames per second) (default:
10)
--video-dpi VIDEO_DPI
output video resolution (dots per inch) (default: 100)
visualizer:
--debug-indices DEBUG_INDICES [DEBUG_INDICES ...]
Indices of fields to create debug plots for of the
form headname:fieldindex, e.g. cif:5. Optionally,
specify the visualization type, e.g. cif:5:hr for the
high resolution plot only. Use comma separation to
specify multiple head names, field indices or
visualization types, e.g. cif:5,6:confidence,hr to
visualize CIF fields 5 and 6 but only show confidence
and hr. (default: [])
eval#
%%bash
python3 -m openpifpaf.eval --help
usage: python3 -m openpifpaf.eval [options]
Evaluation on COCO data.
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--disable-cuda disable CUDA (default: False)
--output OUTPUT output filename without file extension (default: None)
--watch [WATCH] Watch a directory for new checkpoint files. Optionally
specify the number of seconds between checks.
(default: False)
generic data module parameters:
--dataset DATASET
--loader-workers LOADER_WORKERS
number of workers for data loading (default: None)
--batch-size BATCH_SIZE
batch size (default: 1)
--dataset-weights DATASET_WEIGHTS [DATASET_WEIGHTS ...]
n-1 weights for the datasets (default: None)
data module Animal:
--animal-train-annotations ANIMAL_TRAIN_ANNOTATIONS
--animal-val-annotations ANIMAL_VAL_ANNOTATIONS
--animal-train-image-dir ANIMAL_TRAIN_IMAGE_DIR
--animal-val-image-dir ANIMAL_VAL_IMAGE_DIR
--animal-square-edge ANIMAL_SQUARE_EDGE
square edge of input images (default: 513)
--animal-extended-scale
augment with an extended scale range (default: False)
--animal-orientation-invariant ANIMAL_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--animal-blur ANIMAL_BLUR
augment with blur (default: 0.0)
--animal-no-augmentation
do not apply data augmentation (default: True)
--animal-rescale-images ANIMAL_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--animal-upsample ANIMAL_UPSAMPLE
head upsample stride (default: 1)
--animal-min-kp-anns ANIMAL_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--animal-bmin ANIMAL_BMIN
b minimum in pixels (default: 1)
--animal-eval-test2017
--animal-eval-testdev2017
--animal-no-eval-annotation-filter
--animal-eval-long-edge ANIMAL_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 0)
--animal-eval-extended-scale
--animal-eval-orientation-invariant ANIMAL_EVAL_ORIENTATION_INVARIANT
data module Apollo:
--apollo-train-annotations APOLLO_TRAIN_ANNOTATIONS
--apollo-val-annotations APOLLO_VAL_ANNOTATIONS
--apollo-train-image-dir APOLLO_TRAIN_IMAGE_DIR
--apollo-val-image-dir APOLLO_VAL_IMAGE_DIR
--apollo-square-edge APOLLO_SQUARE_EDGE
square edge of input images (default: 513)
--apollo-extended-scale
augment with an extended scale range (default: False)
--apollo-orientation-invariant APOLLO_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--apollo-blur APOLLO_BLUR
augment with blur (default: 0.0)
--apollo-no-augmentation
do not apply data augmentation (default: True)
--apollo-rescale-images APOLLO_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--apollo-upsample APOLLO_UPSAMPLE
head upsample stride (default: 1)
--apollo-min-kp-anns APOLLO_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--apollo-bmin APOLLO_BMIN
b minimum in pixels (default: 1)
--apollo-apply-local-centrality-weights
Weigh the CIF and CAF head during training. (default:
False)
--apollo-no-eval-annotation-filter
--apollo-eval-long-edge APOLLO_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 0)
--apollo-eval-extended-scale
--apollo-eval-orientation-invariant APOLLO_EVAL_ORIENTATION_INVARIANT
--apollo-use-24-kps The ApolloCar3D dataset can be trained with 24 or 66
kps. If you want to train a model with 24 kps activate
this flag. Change the annotations path to the json
files with 24 kps. (default: False)
data module Cifar10:
--cifar10-root-dir CIFAR10_ROOT_DIR
--cifar10-download
data module CocoDet:
--cocodet-train-annotations COCODET_TRAIN_ANNOTATIONS
--cocodet-val-annotations COCODET_VAL_ANNOTATIONS
--cocodet-train-image-dir COCODET_TRAIN_IMAGE_DIR
--cocodet-val-image-dir COCODET_VAL_IMAGE_DIR
--cocodet-square-edge COCODET_SQUARE_EDGE
square edge of input images (default: 513)
--cocodet-extended-scale
augment with an extended scale range (default: False)
--cocodet-orientation-invariant COCODET_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--cocodet-blur COCODET_BLUR
augment with blur (default: 0.0)
--cocodet-no-augmentation
do not apply data augmentation (default: True)
--cocodet-rescale-images COCODET_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--cocodet-upsample COCODET_UPSAMPLE
head upsample stride (default: 1)
data module CocoKp:
--cocokp-train-annotations COCOKP_TRAIN_ANNOTATIONS
train annotations (default: data-
mscoco/annotations/person_keypoints_train2017.json)
--cocokp-val-annotations COCOKP_VAL_ANNOTATIONS
val annotations (default: data-
mscoco/annotations/person_keypoints_val2017.json)
--cocokp-train-image-dir COCOKP_TRAIN_IMAGE_DIR
train image dir (default: data-
mscoco/images/train2017/)
--cocokp-val-image-dir COCOKP_VAL_IMAGE_DIR
val image dir (default: data-mscoco/images/val2017/)
--cocokp-square-edge COCOKP_SQUARE_EDGE
square edge of input images (default: 385)
--cocokp-with-dense train with dense connections (default: False)
--cocokp-extended-scale
augment with an extended scale range (default: False)
--cocokp-orientation-invariant COCOKP_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--cocokp-blur COCOKP_BLUR
augment with blur (default: 0.0)
--cocokp-no-augmentation
do not apply data augmentation (default: True)
--cocokp-rescale-images COCOKP_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--cocokp-upsample COCOKP_UPSAMPLE
head upsample stride (default: 1)
--cocokp-min-kp-anns COCOKP_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--cocokp-bmin COCOKP_BMIN
bmin (default: 0.1)
--cocokp-eval-test2017
--cocokp-eval-testdev2017
--coco-no-eval-annotation-filter
--coco-eval-long-edge COCO_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 641)
--coco-eval-extended-scale
--coco-eval-orientation-invariant COCO_EVAL_ORIENTATION_INVARIANT
data module CrowdPose:
--crowdpose-train-annotations CROWDPOSE_TRAIN_ANNOTATIONS
--crowdpose-val-annotations CROWDPOSE_VAL_ANNOTATIONS
--crowdpose-image-dir CROWDPOSE_IMAGE_DIR
--crowdpose-square-edge CROWDPOSE_SQUARE_EDGE
square edge of input images (default: 385)
--crowdpose-extended-scale
augment with an extended scale range (default: False)
--crowdpose-orientation-invariant CROWDPOSE_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--crowdpose-no-augmentation
do not apply data augmentation (default: True)
--crowdpose-rescale-images CROWDPOSE_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--crowdpose-upsample CROWDPOSE_UPSAMPLE
head upsample stride (default: 1)
--crowdpose-min-kp-anns CROWDPOSE_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--crowdpose-eval-test
--crowdpose-eval-long-edge CROWDPOSE_EVAL_LONG_EDGE
--crowdpose-eval-extended-scale
--crowdpose-eval-orientation-invariant CROWDPOSE_EVAL_ORIENTATION_INVARIANT
--crowdpose-index {easy,medium,hard}
data module NuScenes:
--nuscenes-train-annotations NUSCENES_TRAIN_ANNOTATIONS
--nuscenes-val-annotations NUSCENES_VAL_ANNOTATIONS
--nuscenes-train-image-dir NUSCENES_TRAIN_IMAGE_DIR
--nuscenes-val-image-dir NUSCENES_VAL_IMAGE_DIR
--nuscenes-square-edge NUSCENES_SQUARE_EDGE
square edge of input images (default: 513)
--nuscenes-extended-scale
augment with an extended scale range (default: False)
--nuscenes-orientation-invariant NUSCENES_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--nuscenes-blur NUSCENES_BLUR
augment with blur (default: 0.0)
--nuscenes-no-augmentation
do not apply data augmentation (default: True)
--nuscenes-rescale-images NUSCENES_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--nuscenes-upsample NUSCENES_UPSAMPLE
head upsample stride (default: 1)
data module Posetrack2018:
--posetrack2018-train-annotations POSETRACK2018_TRAIN_ANNOTATIONS
train annotations (default: data-
posetrack2018/annotations/train/*.json)
--posetrack2018-val-annotations POSETRACK2018_VAL_ANNOTATIONS
val annotations (default: data-
posetrack2018/annotations/val/*.json)
--posetrack2018-eval-annotations POSETRACK2018_EVAL_ANNOTATIONS
eval annotations (default: data-
posetrack2018/annotations/val/*.json)
--posetrack2018-data-root POSETRACK2018_DATA_ROOT
data root (default: data-posetrack2018)
data module Posetrack:
--posetrack-square-edge POSETRACK_SQUARE_EDGE
square edge of input images (default: 385)
--posetrack-with-dense
train with dense connections (default: False)
--posetrack-no-augmentation
do not apply data augmentation (default: True)
--posetrack-rescale-images POSETRACK_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--posetrack-upsample POSETRACK_UPSAMPLE
head upsample stride (default: 1)
--posetrack-min-kp-anns POSETRACK_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--posetrack-bmin POSETRACK_BMIN
--posetrack-sample-pairing POSETRACK_SAMPLE_PAIRING
blend two samples together (default: 0.0)
--posetrack-image-augmentations POSETRACK_IMAGE_AUGMENTATIONS
autocontrast, equalize, invert, solarize (default:
0.0)
--posetrack-max-shift POSETRACK_MAX_SHIFT
max shift (default: 30.0)
--posetrack-eval-long-edge POSETRACK_EVAL_LONG_EDGE
--posetrack-eval-extended-scale
--posetrack-eval-orientation-invariant POSETRACK_EVAL_ORIENTATION_INVARIANT
--posetrack-ablation-without-tcaf
data module Posetrack2017:
--posetrack2017-eval-annotations POSETRACK2017_EVAL_ANNOTATIONS
eval annotations (default: data-
posetrack2017/annotations/val/*.json)
--posetrack2017-data-root POSETRACK2017_DATA_ROOT
data root (default: data-posetrack2017)
data module CocoKpSt:
--cocokpst-max-shift COCOKPST_MAX_SHIFT
max shift (default: 30.0)
data module wholebody:
--wholebody-train-annotations WHOLEBODY_TRAIN_ANNOTATIONS
train annotations (default: data-mscoco/annotations/pe
rson_keypoints_train2017_wholebody_pifpaf_style.json)
--wholebody-val-annotations WHOLEBODY_VAL_ANNOTATIONS
val annotations (default: data-
mscoco/annotations/coco_wholebody_val_v1.0.json)
--wholebody-train-image-dir WHOLEBODY_TRAIN_IMAGE_DIR
train image dir (default: data-
mscoco/images/train2017/)
--wholebody-val-image-dir WHOLEBODY_VAL_IMAGE_DIR
val image dir (default: data-mscoco/images/val2017)
--wholebody-square-edge WHOLEBODY_SQUARE_EDGE
square edge of input images (default: 385)
--wholebody-extended-scale
augment with an extended scale range (default: False)
--wholebody-orientation-invariant WHOLEBODY_ORIENTATION_INVARIANT
augment with random orientations (default: 0.0)
--wholebody-blur WHOLEBODY_BLUR
augment with blur (default: 0.0)
--wholebody-no-augmentation
do not apply data augmentation (default: True)
--wholebody-rescale-images WHOLEBODY_RESCALE_IMAGES
overall rescale factor for images (default: 1.0)
--wholebody-upsample WHOLEBODY_UPSAMPLE
head upsample stride (default: 1)
--wholebody-min-kp-anns WHOLEBODY_MIN_KP_ANNS
filter images with fewer keypoint annotations
(default: 1)
--wholebody-bmin WHOLEBODY_BMIN
bmin (default: 1.0)
--wholebody-apply-local-centrality-weights
Weigh the CIF and CAF head during training. (default:
False)
--wholebody-eval-test2017
--wholebody-eval-testdev2017
--wholebody-no-eval-annotation-filter
--wholebody-eval-long-edge WHOLEBODY_EVAL_LONG_EDGE
set to zero to deactivate rescaling (default: 641)
--wholebody-eval-extended-scale
--wholebody-eval-orientation-invariant WHOLEBODY_EVAL_ORIENTATION_INVARIANT
decoder configuration:
--decoder DECODER [DECODER ...]
Decoders to be considered: ['cifcaf', 'cifdet',
'trackingpose', 'posesimilarity', 'cifcafdense'].
(default: None)
--seed-threshold SEED_THRESHOLD
minimum threshold for seeds (default: 0.2)
--instance-threshold INSTANCE_THRESHOLD
filter instances by score (default is 0.0 with
--force-complete-pose and 0.15 otherwise) (default:
None)
--decoder-workers DECODER_WORKERS
number of workers for pose decoding (default: None)
--profile-decoder [PROFILE_DECODER]
specify out .prof file or nothing for default file
name (default: None)
CifCaf decoders:
--cif-th CIF_TH cif threshold (default: 0.3)
--caf-th CAF_TH caf threshold (default: 0.3)
Decoder for tracking:
--tr-single-pose-threshold TR_SINGLE_POSE_THRESHOLD
Single-pose threshold for tracking. (default: 0.3)
--tr-multi-pose-threshold TR_MULTI_POSE_THRESHOLD
multi-pose threshold for tracking. (default: 0.2)
--tr-multi-pose-n TR_MULTI_POSE_N
multi-pose n for tracking. (default: 3)
--tr-minimum-threshold TR_MINIMUM_THRESHOLD
minimum-pose threshold for tracking. (default: 0.1)
CifCaf decoder:
--force-complete-pose
--force-complete-caf-th FORCE_COMPLETE_CAF_TH
CAF threshold for force complete. Set to -1 to
deactivate. (default: 0.001)
--nms-before-force-complete
run an additional NMS before completing poses
(default: False)
--keypoint-threshold KEYPOINT_THRESHOLD
filter keypoints by score (default: 0.15)
--keypoint-threshold-rel KEYPOINT_THRESHOLD_REL
filter keypoint connections by relative score
(default: 0.5)
--greedy greedy decoding (default: False)
--connection-method {max,blend}
connection method to use, max is faster (default:
blend)
--cifcaf-block-joints
block joints (default: False)
--no-reverse-match
--ablation-cifseeds-nms
--ablation-cifseeds-no-rescore
--ablation-caf-no-rescore
--ablation-independent-kp
trackingpose decoder:
--trackingpose-track-recovery
--trackingpose-single-seed
PoseSimilarity:
--posesimilarity-distance {crafted,euclidean,euclidean4,oks}
--posesimilarity-oks-inflate POSESIMILARITY_OKS_INFLATE
CifCafDense decoder:
--dense-connections [DENSE_CONNECTIONS]
logger:
-q, --quiet only show warning messages or above (default: False)
--debug print debug messages (default: False)
--log-stats enable stats logging (default: False)
MobileNetV3:
--mobilenetv3-no-pretrain
use randomly initialized models (default: True)
ShuffleNetv2:
--shufflenetv2-no-pretrain
use randomly initialized models (default: True)
shufflenetv2k:
--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS
out channels of the optional 2nd input convolution
(default: None)
--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION
dilation factor of stage 4 (default: 1)
--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL
kernel width (default: 5)
--shufflenetv2k-conv5-as-stage
--shufflenetv2k-instance-norm
--shufflenetv2k-group-norm
--shufflenetv2k-leaky-relu
MobileNetV2:
--mobilenetv2-no-pretrain
use randomly initialized models (default: True)
ResNet:
--resnet-no-pretrain use randomly initialized models (default: True)
--resnet-pool0-stride RESNET_POOL0_STRIDE
stride of zero removes the pooling op (default: 0)
--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE
stride of the input convolution (default: 2)
--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--resnet-block5-dilation RESNET_BLOCK5_DILATION
use dilated convs in block5 (default: 1)
--resnet-remove-last-block
create a network without the last block (default:
False)
XCiT:
--xcit-out-channels XCIT_OUT_CHANNELS
number of output channels for optional projection
layer (None for no projection layer) (default: None)
--xcit-out-maxpool adds max-pooling to backbone output feature map
(default: False)
--xcit-no-pretrain use randomly initialized models (default: True)
SwinTransformer:
--swin-drop-path-rate SWIN_DROP_PATH_RATE
drop path (stochastic depth) rate (default: 0.2)
--swin-input-upsample
scales input image by a factor of 2 for higher res
feature maps (default: False)
--swin-use-fpn adds a FPN after the Swin network to obtain higher res
feature maps (default: False)
--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS
output channels of the FPN (None to use the default
number of channels of the Swin network) (default:
None)
--swin-fpn-level SWIN_FPN_LEVEL
FPN pyramid level, must be between 1 (highest
resolution) and 4 (lowest resolution) (default: 3)
--swin-no-pretrain use randomly initialized models (default: True)
CompositeField4:
--cf4-dropout CF4_DROPOUT
[experimental] zeroing probability of feature in head
input (default: 0.0)
--cf4-no-inplace-ops alternative graph without inplace ops (default: True)
network configuration:
--checkpoint CHECKPOINT
Path to a local checkpoint. Or provide one of the
following to download a pretrained model:
['shufflenetv2k30-animalpose',
'shufflenetv2k16-apollo-24',
'shufflenetv2k16-apollo-66',
'shufflenetv2k30-apollo-66', 'mobilenetv2',
'mobilenetv3small', 'mobilenetv3large', 'resnet50',
'shufflenetv2k16', 'shufflenetv2k16-withdense',
'shufflenetv2k30', 'mobilenetv3small-cocodet',
'resnet18-cocodet', 'resnet50-crowdpose',
'shufflenetv2k16-nuscenes', 'tshufflenetv2k30',
'shufflenetv2k16-wholebody',
'shufflenetv2k30-wholebody', 'swin_s', 'swin_b',
'swin_t_input_upsample'] (default: None)
--basenet BASENET base network, one of ['mobilenetv2',
'mobilenetv3large', 'mobilenetv3small', 'resnet18',
'resnet50', 'resnet101', 'resnet152', 'resnext50',
'resnext101', 'shufflenetv2x1', 'shufflenetv2x2',
'shufflenetv2k16', 'shufflenetv2k20',
'shufflenetv2kx5', 'shufflenetv2k30',
'shufflenetv2k44', 'squeezenet', 'tshufflenetv2k16',
'tshufflenetv2k30', 'tresnet50', 'cifar10net',
'swin_t', 'swin_s', 'swin_b', 'swin_b_window_12',
'swin_l', 'swin_l_window_12', 'xcit_nano_12_p16',
'xcit_tiny_12_p16', 'xcit_tiny_24_p16',
'xcit_small_12_p16', 'xcit_small_24_p16',
'xcit_medium_24_p16', 'xcit_large_24_p16',
'xcit_nano_12_p8', 'xcit_tiny_12_p8',
'xcit_tiny_24_p8', 'xcit_small_12_p8',
'xcit_small_24_p8', 'xcit_medium_24_p8',
'xcit_large_24_p8', 'effnetv2_s', 'effnetv2_m',
'effnetv2_l', 'effnetv2_xl', 'effnetv2_s16_s',
'effnetv2_s16_m', 'effnetv2_s16_l', 'effnetv2_s16_xl',
'botnet'] (default: None)
--cross-talk CROSS_TALK
[experimental] (default: 0.0)
--no-download-progress
suppress model download progress bar (default: True)
--head-consolidation {keep,create,filter_and_extend}
consolidation strategy for a checkpoint's head
networks and the heads specified by the datamodule
(default: filter_and_extend)
Predictor:
--long-edge LONG_EDGE
rescale the long side of the image (aspect ratio
maintained) (default: None)
--precise-rescaling use more exact image rescaling (requires scipy)
(default: True)
show:
--save-all [SAVE_ALL]
every plot is saved (optional to specify directory)
(default: None)
--show show every plot, i.e., call matplotlib show()
(default: False)
--image-width IMAGE_WIDTH
image width for matplotlib (in inches) (default: None)
--image-height IMAGE_HEIGHT
image height for matplotlib (in inches) (default:
None)
--image-dpi-factor IMAGE_DPI_FACTOR
increase dpi of output image by this factor (default:
2.0)
--image-min-dpi IMAGE_MIN_DPI
minimum dpi of image output (default: 50.0)
--show-file-extension SHOW_FILE_EXTENSION
default file extension (default: jpeg)
--textbox-alpha TEXTBOX_ALPHA
transparency of annotation text box (default: 0.5)
--text-color TEXT_COLOR
annotation text color (default: white)
--font-size FONT_SIZE
annotation font size (default: 8)
--monocolor-connections
use a single color per instance (default: False)
--line-width LINE_WIDTH
skeleton line width (default: None)
--skeleton-solid-threshold SKELETON_SOLID_THRESHOLD
set to 0.0 to draw all connections as solid lines
(default: 0.5)
--show-box show annotation bounding boxes (default: False)
--white-overlay [WHITE_OVERLAY]
increase contrast to annotations by making image
whiter (default: False)
--show-joint-scales show boxes representing joint sizes (default: False)
--show-joint-confidences
print per-joint confidences on skeleton annotations
(default: False)
--show-decoding-order
--show-frontier-order
--show-only-decoded-connections
to debug which connections were used for decoding
(default: False)
--video-fps VIDEO_FPS
output video frame rate (frames per second) (default:
10)
--video-dpi VIDEO_DPI
output video resolution (dots per inch) (default: 100)
visualizer:
--debug-indices DEBUG_INDICES [DEBUG_INDICES ...]
Indices of fields to create debug plots for of the
form headname:fieldindex, e.g. cif:5. Optionally,
specify the visualization type, e.g. cif:5:hr for the
high resolution plot only. Use comma separation to
specify multiple head names, field indices or
visualization types, e.g. cif:5,6:confidence,hr to
visualize CIF fields 5 and 6 but only show confidence
and hr. (default: [])
Evaluator:
--eval-loader-warmup EVAL_LOADER_WARMUP
--eval-n-images EVAL_N_IMAGES
--eval-show-final-image
show the final image (default: False)
--eval-show-final-ground-truth
show the final image with ground truth annotations
(default: False)
--eval-no-skip-epoch0
do not skip eval for epoch 0 (default: True)
--eval-no-skip-existing
skip if output eval file exists already (default:
True)
--eval-write-predictions
write a json and a zip file of the predictions
(default: False)
export_onnx#
%%bash
python3 -m openpifpaf.export_onnx --help
usage: python3 -m openpifpaf.export_onnx [-h] [--version]
[--swin-drop-path-rate SWIN_DROP_PATH_RATE]
[--swin-input-upsample]
[--swin-use-fpn]
[--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS]
[--swin-fpn-level SWIN_FPN_LEVEL]
[--swin-no-pretrain]
[--xcit-out-channels XCIT_OUT_CHANNELS]
[--xcit-out-maxpool]
[--xcit-no-pretrain]
[--shufflenetv2-no-pretrain]
[--mobilenetv3-no-pretrain]
[--mobilenetv2-no-pretrain]
[--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE]
[--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS]
[--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION]
[--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL]
[--shufflenetv2k-conv5-as-stage]
[--shufflenetv2k-instance-norm | --shufflenetv2k-group-norm]
[--shufflenetv2k-leaky-relu]
[--resnet-no-pretrain]
[--resnet-pool0-stride RESNET_POOL0_STRIDE]
[--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE]
[--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE]
[--resnet-block5-dilation RESNET_BLOCK5_DILATION]
[--resnet-remove-last-block]
[--cf4-dropout CF4_DROPOUT]
[--cf4-no-inplace-ops]
[--checkpoint CHECKPOINT]
[--basenet BASENET]
[--cross-talk CROSS_TALK]
[--no-download-progress]
[--head-consolidation {keep,create,filter_and_extend}]
[--outfile OUTFILE] [--simplify]
[--check] [--input-width INPUT_WIDTH]
[--input-height INPUT_HEIGHT]
Export a checkpoint as an ONNX model.
Applies onnx utilities to improve the exported model and
also tries to simplify the model with onnx-simplifier.
https://github.com/onnx/onnx/blob/master/docs/PythonAPIOverview.md
https://github.com/daquexian/onnx-simplifier
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--outfile OUTFILE
--simplify
--check
--input-width INPUT_WIDTH
--input-height INPUT_HEIGHT
SwinTransformer:
--swin-drop-path-rate SWIN_DROP_PATH_RATE
drop path (stochastic depth) rate (default: 0.2)
--swin-input-upsample
scales input image by a factor of 2 for higher res
feature maps (default: False)
--swin-use-fpn adds a FPN after the Swin network to obtain higher res
feature maps (default: False)
--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS
output channels of the FPN (None to use the default
number of channels of the Swin network) (default:
None)
--swin-fpn-level SWIN_FPN_LEVEL
FPN pyramid level, must be between 1 (highest
resolution) and 4 (lowest resolution) (default: 3)
--swin-no-pretrain use randomly initialized models (default: True)
XCiT:
--xcit-out-channels XCIT_OUT_CHANNELS
number of output channels for optional projection
layer (None for no projection layer) (default: None)
--xcit-out-maxpool adds max-pooling to backbone output feature map
(default: False)
--xcit-no-pretrain use randomly initialized models (default: True)
ShuffleNetv2:
--shufflenetv2-no-pretrain
use randomly initialized models (default: True)
MobileNetV3:
--mobilenetv3-no-pretrain
use randomly initialized models (default: True)
MobileNetV2:
--mobilenetv2-no-pretrain
use randomly initialized models (default: True)
shufflenetv2k:
--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS
out channels of the optional 2nd input convolution
(default: None)
--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION
dilation factor of stage 4 (default: 1)
--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL
kernel width (default: 5)
--shufflenetv2k-conv5-as-stage
--shufflenetv2k-instance-norm
--shufflenetv2k-group-norm
--shufflenetv2k-leaky-relu
ResNet:
--resnet-no-pretrain use randomly initialized models (default: True)
--resnet-pool0-stride RESNET_POOL0_STRIDE
stride of zero removes the pooling op (default: 0)
--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE
stride of the input convolution (default: 2)
--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--resnet-block5-dilation RESNET_BLOCK5_DILATION
use dilated convs in block5 (default: 1)
--resnet-remove-last-block
create a network without the last block (default:
False)
CompositeField4:
--cf4-dropout CF4_DROPOUT
[experimental] zeroing probability of feature in head
input (default: 0.0)
--cf4-no-inplace-ops alternative graph without inplace ops (default: True)
network configuration:
--checkpoint CHECKPOINT
Path to a local checkpoint. Or provide one of the
following to download a pretrained model:
['shufflenetv2k30-animalpose',
'shufflenetv2k16-apollo-24',
'shufflenetv2k16-apollo-66',
'shufflenetv2k30-apollo-66', 'mobilenetv2',
'mobilenetv3small', 'mobilenetv3large', 'resnet50',
'shufflenetv2k16', 'shufflenetv2k16-withdense',
'shufflenetv2k30', 'mobilenetv3small-cocodet',
'resnet18-cocodet', 'resnet50-crowdpose',
'shufflenetv2k16-nuscenes', 'tshufflenetv2k30',
'shufflenetv2k16-wholebody',
'shufflenetv2k30-wholebody', 'swin_s', 'swin_b',
'swin_t_input_upsample'] (default: None)
--basenet BASENET base network, one of ['mobilenetv2',
'mobilenetv3large', 'mobilenetv3small', 'resnet18',
'resnet50', 'resnet101', 'resnet152', 'resnext50',
'resnext101', 'shufflenetv2x1', 'shufflenetv2x2',
'shufflenetv2k16', 'shufflenetv2k20',
'shufflenetv2kx5', 'shufflenetv2k30',
'shufflenetv2k44', 'squeezenet', 'tshufflenetv2k16',
'tshufflenetv2k30', 'tresnet50', 'cifar10net',
'swin_t', 'swin_s', 'swin_b', 'swin_b_window_12',
'swin_l', 'swin_l_window_12', 'xcit_nano_12_p16',
'xcit_tiny_12_p16', 'xcit_tiny_24_p16',
'xcit_small_12_p16', 'xcit_small_24_p16',
'xcit_medium_24_p16', 'xcit_large_24_p16',
'xcit_nano_12_p8', 'xcit_tiny_12_p8',
'xcit_tiny_24_p8', 'xcit_small_12_p8',
'xcit_small_24_p8', 'xcit_medium_24_p8',
'xcit_large_24_p8', 'effnetv2_s', 'effnetv2_m',
'effnetv2_l', 'effnetv2_xl', 'effnetv2_s16_s',
'effnetv2_s16_m', 'effnetv2_s16_l', 'effnetv2_s16_xl',
'botnet'] (default: None)
--cross-talk CROSS_TALK
[experimental] (default: 0.0)
--no-download-progress
suppress model download progress bar (default: True)
--head-consolidation {keep,create,filter_and_extend}
consolidation strategy for a checkpoint's head
networks and the heads specified by the datamodule
(default: filter_and_extend)
export_coreml#
%%bash
python3 -m openpifpaf.export_coreml --help
usage: python3 -m openpifpaf.export_coreml [-h] [--version]
[--swin-drop-path-rate SWIN_DROP_PATH_RATE]
[--swin-input-upsample]
[--swin-use-fpn]
[--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS]
[--swin-fpn-level SWIN_FPN_LEVEL]
[--swin-no-pretrain]
[--xcit-out-channels XCIT_OUT_CHANNELS]
[--xcit-out-maxpool]
[--xcit-no-pretrain]
[--mobilenetv3-no-pretrain]
[--mobilenetv2-no-pretrain]
[--shufflenetv2-no-pretrain]
[--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE]
[--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS]
[--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION]
[--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL]
[--shufflenetv2k-conv5-as-stage]
[--shufflenetv2k-instance-norm | --shufflenetv2k-group-norm]
[--shufflenetv2k-leaky-relu]
[--resnet-no-pretrain]
[--resnet-pool0-stride RESNET_POOL0_STRIDE]
[--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE]
[--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE]
[--resnet-block5-dilation RESNET_BLOCK5_DILATION]
[--resnet-remove-last-block]
[--cf4-dropout CF4_DROPOUT]
[--cf4-no-inplace-ops]
[--checkpoint CHECKPOINT]
[--basenet BASENET]
[--cross-talk CROSS_TALK]
[--no-download-progress]
[--head-consolidation {keep,create,filter_and_extend}]
[--outfile OUTFILE]
[--input-width INPUT_WIDTH]
[--input-height INPUT_HEIGHT]
[--minimum-deployment-target {iOS13,iOS14}]
Export a checkpoint as a CoreML model.
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--outfile OUTFILE
--input-width INPUT_WIDTH
--input-height INPUT_HEIGHT
--minimum-deployment-target {iOS13,iOS14}
SwinTransformer:
--swin-drop-path-rate SWIN_DROP_PATH_RATE
drop path (stochastic depth) rate (default: 0.2)
--swin-input-upsample
scales input image by a factor of 2 for higher res
feature maps (default: False)
--swin-use-fpn adds a FPN after the Swin network to obtain higher res
feature maps (default: False)
--swin-fpn-out-channels SWIN_FPN_OUT_CHANNELS
output channels of the FPN (None to use the default
number of channels of the Swin network) (default:
None)
--swin-fpn-level SWIN_FPN_LEVEL
FPN pyramid level, must be between 1 (highest
resolution) and 4 (lowest resolution) (default: 3)
--swin-no-pretrain use randomly initialized models (default: True)
XCiT:
--xcit-out-channels XCIT_OUT_CHANNELS
number of output channels for optional projection
layer (None for no projection layer) (default: None)
--xcit-out-maxpool adds max-pooling to backbone output feature map
(default: False)
--xcit-no-pretrain use randomly initialized models (default: True)
MobileNetV3:
--mobilenetv3-no-pretrain
use randomly initialized models (default: True)
MobileNetV2:
--mobilenetv2-no-pretrain
use randomly initialized models (default: True)
ShuffleNetv2:
--shufflenetv2-no-pretrain
use randomly initialized models (default: True)
shufflenetv2k:
--shufflenetv2k-input-conv2-stride SHUFFLENETV2K_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--shufflenetv2k-input-conv2-outchannels SHUFFLENETV2K_INPUT_CONV2_OUTCHANNELS
out channels of the optional 2nd input convolution
(default: None)
--shufflenetv2k-stage4-dilation SHUFFLENETV2K_STAGE4_DILATION
dilation factor of stage 4 (default: 1)
--shufflenetv2k-kernel SHUFFLENETV2K_KERNEL
kernel width (default: 5)
--shufflenetv2k-conv5-as-stage
--shufflenetv2k-instance-norm
--shufflenetv2k-group-norm
--shufflenetv2k-leaky-relu
ResNet:
--resnet-no-pretrain use randomly initialized models (default: True)
--resnet-pool0-stride RESNET_POOL0_STRIDE
stride of zero removes the pooling op (default: 0)
--resnet-input-conv-stride RESNET_INPUT_CONV_STRIDE
stride of the input convolution (default: 2)
--resnet-input-conv2-stride RESNET_INPUT_CONV2_STRIDE
stride of the optional 2nd input convolution (default:
0)
--resnet-block5-dilation RESNET_BLOCK5_DILATION
use dilated convs in block5 (default: 1)
--resnet-remove-last-block
create a network without the last block (default:
False)
CompositeField4:
--cf4-dropout CF4_DROPOUT
[experimental] zeroing probability of feature in head
input (default: 0.0)
--cf4-no-inplace-ops alternative graph without inplace ops (default: True)
network configuration:
--checkpoint CHECKPOINT
Path to a local checkpoint. Or provide one of the
following to download a pretrained model:
['shufflenetv2k30-animalpose',
'shufflenetv2k16-apollo-24',
'shufflenetv2k16-apollo-66',
'shufflenetv2k30-apollo-66', 'mobilenetv2',
'mobilenetv3small', 'mobilenetv3large', 'resnet50',
'shufflenetv2k16', 'shufflenetv2k16-withdense',
'shufflenetv2k30', 'mobilenetv3small-cocodet',
'resnet18-cocodet', 'resnet50-crowdpose',
'shufflenetv2k16-nuscenes', 'tshufflenetv2k30',
'shufflenetv2k16-wholebody',
'shufflenetv2k30-wholebody', 'swin_s', 'swin_b',
'swin_t_input_upsample'] (default: None)
--basenet BASENET base network, one of ['mobilenetv2',
'mobilenetv3large', 'mobilenetv3small', 'resnet18',
'resnet50', 'resnet101', 'resnet152', 'resnext50',
'resnext101', 'shufflenetv2x1', 'shufflenetv2x2',
'shufflenetv2k16', 'shufflenetv2k20',
'shufflenetv2kx5', 'shufflenetv2k30',
'shufflenetv2k44', 'squeezenet', 'tshufflenetv2k16',
'tshufflenetv2k30', 'tresnet50', 'cifar10net',
'swin_t', 'swin_s', 'swin_b', 'swin_b_window_12',
'swin_l', 'swin_l_window_12', 'xcit_nano_12_p16',
'xcit_tiny_12_p16', 'xcit_tiny_24_p16',
'xcit_small_12_p16', 'xcit_small_24_p16',
'xcit_medium_24_p16', 'xcit_large_24_p16',
'xcit_nano_12_p8', 'xcit_tiny_12_p8',
'xcit_tiny_24_p8', 'xcit_small_12_p8',
'xcit_small_24_p8', 'xcit_medium_24_p8',
'xcit_large_24_p8', 'effnetv2_s', 'effnetv2_m',
'effnetv2_l', 'effnetv2_xl', 'effnetv2_s16_s',
'effnetv2_s16_m', 'effnetv2_s16_l', 'effnetv2_s16_xl',
'botnet'] (default: None)
--cross-talk CROSS_TALK
[experimental] (default: 0.0)
--no-download-progress
suppress model download progress bar (default: True)
--head-consolidation {keep,create,filter_and_extend}
consolidation strategy for a checkpoint's head
networks and the heads specified by the datamodule
(default: filter_and_extend)
benchmark#
%%bash
python3 -m openpifpaf.benchmark --help
usage: python3 -m openpifpaf.benchmark [-h] [--version] [--output OUTPUT]
[--checkpoints CHECKPOINTS [CHECKPOINTS ...]]
[--iccv2019-ablation]
[--v012-ablation-1] [--v012-ablation-2]
[--v012-ablation-3] [--v012-ablation-4]
[--debug]
Benchmark.
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--output OUTPUT output file name (default: None)
--checkpoints CHECKPOINTS [CHECKPOINTS ...]
checkpoints to evaluate (default: ['resnet50',
'shufflenetv2k16', 'shufflenetv2k30'])
--iccv2019-ablation
--v012-ablation-1
--v012-ablation-2
--v012-ablation-3
--v012-ablation-4
logging:
--debug print debug messages (default: False)
logs#
%%bash
python3 -m openpifpaf.logs --help
usage: python3 -m openpifpaf.logs [options] log_files
Configuring and visualizing log files.
positional arguments:
log_file path to log file(s)
optional arguments:
-h, --help show this help message and exit
--version show program's version number and exit
--label LABEL [LABEL ...]
label(s) in the same order as files (default: None)
--eval-suffix EVAL_SUFFIX
suffix of evaluation files to look for (default:
.eval-*.stats.json)
--first-epoch FIRST_EPOCH
epoch (can be float) of first data point to plot
(default: 1e-06)
--no-share-y dont share y access (default: True)
-o OUTPUT, --output OUTPUT
output prefix (default is log_file + .) (default:
None)
--show-mtl-sigmas
logger:
-q, --quiet only show warning messages or above (default: False)
--debug print debug messages (default: False)
--log-stats enable stats logging (default: False)
show:
--save-all [SAVE_ALL]
every plot is saved (optional to specify directory)
(default: None)
--show show every plot, i.e., call matplotlib show()
(default: False)
--image-width IMAGE_WIDTH
image width for matplotlib (in inches) (default: None)
--image-height IMAGE_HEIGHT
image height for matplotlib (in inches) (default:
None)
--image-dpi-factor IMAGE_DPI_FACTOR
increase dpi of output image by this factor (default:
2.0)
--image-min-dpi IMAGE_MIN_DPI
minimum dpi of image output (default: 50.0)
--show-file-extension SHOW_FILE_EXTENSION
default file extension (default: jpeg)
--textbox-alpha TEXTBOX_ALPHA
transparency of annotation text box (default: 0.5)
--text-color TEXT_COLOR
annotation text color (default: white)
--font-size FONT_SIZE
annotation font size (default: 8)
--monocolor-connections
use a single color per instance (default: False)
--line-width LINE_WIDTH
skeleton line width (default: None)
--skeleton-solid-threshold SKELETON_SOLID_THRESHOLD
set to 0.0 to draw all connections as solid lines
(default: 0.5)
--show-box show annotation bounding boxes (default: False)
--white-overlay [WHITE_OVERLAY]
increase contrast to annotations by making image
whiter (default: False)
--show-joint-scales show boxes representing joint sizes (default: False)
--show-joint-confidences
print per-joint confidences on skeleton annotations
(default: False)
--show-decoding-order
--show-frontier-order
--show-only-decoded-connections
to debug which connections were used for decoding
(default: False)
--video-fps VIDEO_FPS
output video frame rate (frames per second) (default:
10)
--video-dpi VIDEO_DPI
output video resolution (dots per inch) (default: 100)