def test_net(visualise, cache_scoremaps):
logging.basicConfig(level=logging.INFO)
cfg = load_config()
dataset = create_dataset(cfg)
dataset.set_shuffle(False)
dataset.set_test_mode(True)
sess, inputs, outputs = setup_pose_prediction(cfg)
if cache_scoremaps:
out_dir = cfg.scoremap_dir
if not os.path.exists(out_dir):
os.makedirs(out_dir)
num_images = dataset.num_images
predictions = np.zeros((num_images,), dtype=np.object)
for k in range(num_images):
print("processing image {}/{}".format(k, num_images - 1))
batch = dataset.next_batch()
outputs_np = sess.run(outputs, feed_dict={inputs: batch[Batch.inputs]})
scmap, locref = extract_cnn_output(outputs_np, cfg)
pose = argmax_pose_predict(scmap, locref, cfg.stride)
pose_refscale = np.copy(pose)
pose_refscale[:, 0:2] /= cfg.global_scale
predictions[k] = pose_refscale
if visualise:
img = np.squeeze(batch[Batch.inputs]).astype("uint8")
visualize.show_heatmaps(cfg, img, scmap, pose)
visualize.waitforbuttonpress()
if cache_scoremaps:
base = os.path.basename(batch[Batch.data_item].im_path)
raw_name = os.path.splitext(base)[0]
out_fn = os.path.join(out_dir, raw_name + ".mat")
scipy.io.savemat(out_fn, mdict={"scoremaps": scmap.astype("float32")})
out_fn = os.path.join(out_dir, raw_name + "_locreg" + ".mat")
if cfg.location_refinement:
scipy.io.savemat(
out_fn, mdict={"locreg_pred": locref.astype("float32")}
)
scipy.io.savemat("predictions.mat", mdict={"joints": predictions})
sess.close()
def load_deeplabcut():
"""
Loads TensorFlow with predefined in config DeepLabCut model
:return: tuple of DeepLabCut config, TensorFlow session, inputs and outputs
"""
model = os.path.join(MODEL_PATH, models_folder, MODEL_NAME)
cfg = load_config(os.path.join(model, 'test/pose_cfg.yaml'))
snapshots = sorted([sn.split('.')[0] for sn in os.listdir(model + '/train/') if "index" in sn])
cfg['init_weights'] = model + '/train/' + snapshots[-1]
sess, inputs, outputs = predict.setup_pose_prediction(cfg)
return cfg, sess, inputs, outputs
def evaluate_network(config,
Shuffles=[1],
trainingsetindex=0,
plotting=None,
show_errors=True,
comparisonbodyparts="all",
gputouse=None,
rescale=False):
"""
Evaluates the network based on the saved models at different stages of the training network.\n
The evaluation results are stored in the .h5 and .csv file under the subdirectory 'evaluation_results'.
Change the snapshotindex parameter in the config file to 'all' in order to evaluate all the saved models.
Parameters
----------
config : string
Full path of the config.yaml file as a string.
Shuffles: list, optional
List of integers specifying the shuffle indices of the training dataset. The default is [1]
trainingsetindex: int, optional
Integer specifying which TrainingsetFraction to use. By default the first (note that TrainingFraction is a list in config.yaml). This
variable can also be set to "all".
plotting: bool, optional
Plots the predictions on the train and test images. The default is ``False``; if provided it must be either ``True`` or ``False``
show_errors: bool, optional
Display train and test errors. The default is `True``
comparisonbodyparts: list of bodyparts, Default is "all".
The average error will be computed for those body parts only (Has to be a subset of the body parts).
gputouse: int, optional. Natural number indicating the number of your GPU (see number in nvidia-smi). If you do not have a GPU put None.
See: https://nvidia.custhelp.com/app/answers/detail/a_id/3751/~/useful-nvidia-smi-queries
rescale: bool, default False
Evaluate the model at the 'global_scale' variable (as set in the test/pose_config.yaml file for a particular project). I.e. every
image will be resized according to that scale and prediction will be compared to the resized ground truth. The error will be reported
in pixels at rescaled to the *original* size. I.e. For a [200,200] pixel image evaluated at global_scale=.5, the predictions are calculated
on [100,100] pixel images, compared to 1/2*ground truth and this error is then multiplied by 2!. The evaluation images are also shown for the
original size!
Examples
--------
If you do not want to plot
>>> deeplabcutcore.evaluate_network('/analysis/project/reaching-task/config.yaml', Shuffles=[1])
--------
If you want to plot
>>> deeplabcutcore.evaluate_network('/analysis/project/reaching-task/config.yaml',Shuffles=[1],True)
"""
import os
#import skimage.color
#from skimage.io import imread
from deeplabcutcore.utils.auxfun_videos import imread, imresize
from deeplabcutcore.pose_estimation_tensorflow.nnet import predict
from deeplabcutcore.pose_estimation_tensorflow.config import load_config
from deeplabcutcore.pose_estimation_tensorflow.dataset.pose_dataset import data_to_input
from deeplabcutcore.utils import auxiliaryfunctions
import tensorflow as tf
if 'TF_CUDNN_USE_AUTOTUNE' in os.environ:
del os.environ[
'TF_CUDNN_USE_AUTOTUNE'] #was potentially set during training
tf.compat.v1.reset_default_graph()
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' #
# tf.logging.set_verbosity(tf.logging.WARN)
start_path = os.getcwd()
# Read file path for pose_config file. >> pass it on
cfg = auxiliaryfunctions.read_config(config)
if gputouse is not None: #gpu selectinon
os.environ['CUDA_VISIBLE_DEVICES'] = str(gputouse)
if trainingsetindex == 'all':
TrainingFractions = cfg["TrainingFraction"]
else:
if trainingsetindex < len(
cfg["TrainingFraction"]) and trainingsetindex >= 0:
TrainingFractions = [
cfg["TrainingFraction"][int(trainingsetindex)]
]
else:
raise Exception('Please check the trainingsetindex! ',
trainingsetindex,
' should be an integer from 0 .. ',
int(len(cfg["TrainingFraction"]) - 1))
# Loading human annotatated data
trainingsetfolder = auxiliaryfunctions.GetTrainingSetFolder(cfg)
Data = pd.read_hdf(
os.path.join(cfg["project_path"], str(trainingsetfolder),
'CollectedData_' + cfg["scorer"] + '.h5'),
'df_with_missing')
# Get list of body parts to evaluate network for
comparisonbodyparts = auxiliaryfunctions.IntersectionofBodyPartsandOnesGivenbyUser(
cfg, comparisonbodyparts)
# Make folder for evaluation
auxiliaryfunctions.attempttomakefolder(
str(cfg["project_path"] + "/evaluation-results/"))
for shuffle in Shuffles:
for trainFraction in TrainingFractions:
##################################################
# Load and setup CNN part detector
##################################################
datafn, metadatafn = auxiliaryfunctions.GetDataandMetaDataFilenames(
trainingsetfolder, trainFraction, shuffle, cfg)
modelfolder = os.path.join(
cfg["project_path"],
str(
auxiliaryfunctions.GetModelFolder(trainFraction, shuffle,
cfg)))
path_test_config = Path(modelfolder) / 'test' / 'pose_cfg.yaml'
# Load meta data
data, trainIndices, testIndices, trainFraction = auxiliaryfunctions.LoadMetadata(
os.path.join(cfg["project_path"], metadatafn))
try:
dlc_cfg = load_config(str(path_test_config))
except FileNotFoundError:
raise FileNotFoundError(
"It seems the model for shuffle %s and trainFraction %s does not exist."
% (shuffle, trainFraction))
#change batch size, if it was edited during analysis!
dlc_cfg['batch_size'] = 1 #in case this was edited for analysis.
#Create folder structure to store results.
evaluationfolder = os.path.join(
cfg["project_path"],
str(
auxiliaryfunctions.GetEvaluationFolder(
trainFraction, shuffle, cfg)))
auxiliaryfunctions.attempttomakefolder(evaluationfolder,
recursive=True)
#path_train_config = modelfolder / 'train' / 'pose_cfg.yaml'
# Check which snapshots are available and sort them by # iterations
Snapshots = np.array([
fn.split('.')[0]
for fn in os.listdir(os.path.join(str(modelfolder), 'train'))
if "index" in fn
])
try: #check if any where found?
Snapshots[0]
except IndexError:
raise FileNotFoundError(
"Snapshots not found! It seems the dataset for shuffle %s and trainFraction %s is not trained.\nPlease train it before evaluating.\nUse the function 'train_network' to do so."
% (shuffle, trainFraction))
increasing_indices = np.argsort(
[int(m.split('-')[1]) for m in Snapshots])
Snapshots = Snapshots[increasing_indices]
if cfg["snapshotindex"] == -1:
snapindices = [-1]
elif cfg["snapshotindex"] == "all":
snapindices = range(len(Snapshots))
elif cfg["snapshotindex"] < len(Snapshots):
snapindices = [cfg["snapshotindex"]]
else:
raise ValueError(
"Invalid choice, only -1 (last), any integer up to last, or all (as string)!"
)
final_result = []
########################### RESCALING (to global scale)
if rescale == True:
scale = dlc_cfg['global_scale']
Data = pd.read_hdf(
os.path.join(cfg["project_path"], str(trainingsetfolder),
'CollectedData_' + cfg["scorer"] + '.h5'),
'df_with_missing') * scale
else:
scale = 1
##################################################
# Compute predictions over images
##################################################
for snapindex in snapindices:
dlc_cfg['init_weights'] = os.path.join(
str(modelfolder), 'train', Snapshots[snapindex]
) #setting weights to corresponding snapshot.
trainingsiterations = (
dlc_cfg['init_weights'].split(os.sep)[-1]
).split(
'-'
)[-1] #read how many training siterations that corresponds to.
# Name for deeplabcut net (based on its parameters)
DLCscorer, DLCscorerlegacy = auxiliaryfunctions.GetScorerName(
cfg, shuffle, trainFraction, trainingsiterations)
print("Running ", DLCscorer, " with # of trainingiterations:",
trainingsiterations)
notanalyzed, resultsfilename, DLCscorer = auxiliaryfunctions.CheckifNotEvaluated(
str(evaluationfolder), DLCscorer, DLCscorerlegacy,
Snapshots[snapindex])
if notanalyzed:
# Specifying state of model (snapshot / training state)
sess, inputs, outputs = predict.setup_pose_prediction(
dlc_cfg)
Numimages = len(Data.index)
PredicteData = np.zeros(
(Numimages, 3 * len(dlc_cfg['all_joints_names'])))
print("Analyzing data...")
for imageindex, imagename in tqdm(enumerate(Data.index)):
image = imread(os.path.join(cfg['project_path'],
imagename),
mode='RGB')
if scale != 1:
image = imresize(image, scale)
#image = skimage.color.gray2rgb(image)
image_batch = data_to_input(image)
# Compute prediction with the CNN
outputs_np = sess.run(outputs,
feed_dict={inputs: image_batch})
scmap, locref = predict.extract_cnn_output(
outputs_np, dlc_cfg)
# Extract maximum scoring location from the heatmap, assume 1 person
pose = predict.argmax_pose_predict(
scmap, locref, dlc_cfg.stride)
PredicteData[imageindex, :] = pose.flatten(
) # NOTE: thereby cfg_test['all_joints_names'] should be same order as bodyparts!
sess.close() #closes the current tf session
index = pd.MultiIndex.from_product(
[[DLCscorer], dlc_cfg['all_joints_names'],
['x', 'y', 'likelihood']],
names=['scorer', 'bodyparts', 'coords'])
# Saving results
DataMachine = pd.DataFrame(PredicteData,
columns=index,
index=Data.index.values)
DataMachine.to_hdf(resultsfilename,
'df_with_missing',
format='table',
mode='w')
print("Done and results stored for snapshot: ",
Snapshots[snapindex])
DataCombined = pd.concat([Data.T, DataMachine.T],
axis=0,
sort=False).T
RMSE, RMSEpcutoff = pairwisedistances(
DataCombined, cfg["scorer"], DLCscorer, cfg["pcutoff"],
comparisonbodyparts)
testerror = np.nanmean(
RMSE.iloc[testIndices].values.flatten())
trainerror = np.nanmean(
RMSE.iloc[trainIndices].values.flatten())
testerrorpcutoff = np.nanmean(
RMSEpcutoff.iloc[testIndices].values.flatten())
trainerrorpcutoff = np.nanmean(
RMSEpcutoff.iloc[trainIndices].values.flatten())
results = [
trainingsiterations,
int(100 * trainFraction), shuffle,
np.round(trainerror, 2),
np.round(testerror, 2), cfg["pcutoff"],
np.round(trainerrorpcutoff, 2),
np.round(testerrorpcutoff, 2)
]
final_result.append(results)
if show_errors == True:
print("Results for",
trainingsiterations, " training iterations:",
int(100 * trainFraction), shuffle,
"train error:",
np.round(trainerror, 2), "pixels. Test error:",
np.round(testerror, 2), " pixels.")
print("With pcutoff of",
cfg["pcutoff"], " train error:",
np.round(trainerrorpcutoff,
2), "pixels. Test error:",
np.round(testerrorpcutoff, 2), "pixels")
if scale != 1:
print(
"The predictions have been calculated for rescaled images (and rescaled ground truth). Scale:",
scale)
print(
"Thereby, the errors are given by the average distances between the labels by DLC and the scorer."
)
if plotting == True:
print("Plotting...")
foldername = os.path.join(
str(evaluationfolder), 'LabeledImages_' +
DLCscorer + '_' + Snapshots[snapindex])
auxiliaryfunctions.attempttomakefolder(foldername)
Plotting(
cfg, comparisonbodyparts, DLCscorer, trainIndices,
DataCombined * 1. / scale, foldername
) #Rescaling coordinates to have figure in original size!
tf.compat.v1.reset_default_graph()
#print(final_result)
else:
DataMachine = pd.read_hdf(resultsfilename,
'df_with_missing')
if plotting == True:
DataCombined = pd.concat([Data.T, DataMachine.T],
axis=0,
sort=False).T
print(
"Plotting...(attention scale might be inconsistent in comparison to when data was analyzed; i.e. if you used rescale)"
)
foldername = os.path.join(
str(evaluationfolder), 'LabeledImages_' +
DLCscorer + '_' + Snapshots[snapindex])
auxiliaryfunctions.attempttomakefolder(foldername)
Plotting(cfg, comparisonbodyparts, DLCscorer,
trainIndices, DataCombined * 1. / scale,
foldername)
if len(final_result) > 0: #Only append if results were calculated
make_results_file(final_result, evaluationfolder, DLCscorer)
print(
"The network is evaluated and the results are stored in the subdirectory 'evaluation_results'."
)
print(
"If it generalizes well, choose the best model for prediction and update the config file with the appropriate index for the 'snapshotindex'.\nUse the function 'analyze_video' to make predictions on new videos."
)
print(
"Otherwise consider retraining the network (see DeepLabCut workflow Fig 2)"
)
#returning to intial folder
os.chdir(str(start_path))
def analyze_time_lapse_frames(config,
directory,
frametype='.png',
shuffle=1,
trainingsetindex=0,
gputouse=None,
save_as_csv=False,
rgb=True):
"""
Analyzed all images (of type = frametype) in a folder and stores the output in one file.
You can crop the frames (before analysis), by changing 'cropping'=True and setting 'x1','x2','y1','y2' in the config file.
Output: The labels are stored as MultiIndex Pandas Array, which contains the name of the network, body part name, (x, y) label position \n
in pixels, and the likelihood for each frame per body part. These arrays are stored in an efficient Hierarchical Data Format (HDF) \n
in the same directory, where the video is stored. However, if the flag save_as_csv is set to True, the data can also be exported in \n
comma-separated values format (.csv), which in turn can be imported in many programs, such as MATLAB, R, Prism, etc.
Parameters
----------
config : string
Full path of the config.yaml file as a string.
directory: string
Full path to directory containing the frames that shall be analyzed
frametype: string, optional
Checks for the file extension of the frames. Only images with this extension are analyzed. The default is ``.png``
shuffle: int, optional
An integer specifying the shuffle index of the training dataset used for training the network. The default is 1.
trainingsetindex: int, optional
Integer specifying which TrainingsetFraction to use. By default the first (note that TrainingFraction is a list in config.yaml).
gputouse: int, optional. Natural number indicating the number of your GPU (see number in nvidia-smi). If you do not have a GPU put None.
See: https://nvidia.custhelp.com/app/answers/detail/a_id/3751/~/useful-nvidia-smi-queries
save_as_csv: bool, optional
Saves the predictions in a .csv file. The default is ``False``; if provided it must be either ``True`` or ``False``
rbg: bool, optional.
Whether to load image as rgb; Note e.g. some tiffs do not alow that option in imread, then just set this to false.
Examples
--------
If you want to analyze all frames in /analysis/project/timelapseexperiment1
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml','/analysis/project/timelapseexperiment1')
--------
If you want to analyze all frames in /analysis/project/timelapseexperiment1
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml','/analysis/project/timelapseexperiment1')
--------
Note: for test purposes one can extract all frames from a video with ffmeg, e.g. ffmpeg -i testvideo.avi thumb%04d.png
"""
if 'TF_CUDNN_USE_AUTOTUNE' in os.environ:
del os.environ[
'TF_CUDNN_USE_AUTOTUNE'] #was potentially set during training
if gputouse is not None: #gpu selection
os.environ['CUDA_VISIBLE_DEVICES'] = str(gputouse)
tf.compat.v1.reset_default_graph()
start_path = os.getcwd(
) #record cwd to return to this directory in the end
cfg = auxiliaryfunctions.read_config(config)
trainFraction = cfg['TrainingFraction'][trainingsetindex]
modelfolder = os.path.join(
cfg["project_path"],
str(auxiliaryfunctions.GetModelFolder(trainFraction, shuffle, cfg)))
path_test_config = Path(modelfolder) / 'test' / 'pose_cfg.yaml'
try:
dlc_cfg = load_config(str(path_test_config))
except FileNotFoundError:
raise FileNotFoundError(
"It seems the model for shuffle %s and trainFraction %s does not exist."
% (shuffle, trainFraction))
# Check which snapshots are available and sort them by # iterations
try:
Snapshots = np.array([
fn.split('.')[0]
for fn in os.listdir(os.path.join(modelfolder, 'train'))
if "index" in fn
])
except FileNotFoundError:
raise FileNotFoundError(
"Snapshots not found! It seems the dataset for shuffle %s has not been trained/does not exist.\n Please train it before using it to analyze videos.\n Use the function 'train_network' to train the network for shuffle %s."
% (shuffle, shuffle))
if cfg['snapshotindex'] == 'all':
print(
"Snapshotindex is set to 'all' in the config.yaml file. Running video analysis with all snapshots is very costly! Use the function 'evaluate_network' to choose the best the snapshot. For now, changing snapshot index to -1!"
)
snapshotindex = -1
else:
snapshotindex = cfg['snapshotindex']
increasing_indices = np.argsort([int(m.split('-')[1]) for m in Snapshots])
Snapshots = Snapshots[increasing_indices]
print("Using %s" % Snapshots[snapshotindex], "for model", modelfolder)
##################################################
# Load and setup CNN part detector
##################################################
# Check if data already was generated:
dlc_cfg['init_weights'] = os.path.join(modelfolder, 'train',
Snapshots[snapshotindex])
trainingsiterations = (dlc_cfg['init_weights'].split(
os.sep)[-1]).split('-')[-1]
#update batchsize (based on parameters in config.yaml)
dlc_cfg['batch_size'] = cfg['batch_size']
# Name for scorer:
DLCscorer, DLCscorerlegacy = auxiliaryfunctions.GetScorerName(
cfg, shuffle, trainFraction, trainingsiterations=trainingsiterations)
sess, inputs, outputs = predict.setup_pose_prediction(dlc_cfg)
# update number of outputs and adjust pandas indices
dlc_cfg['num_outputs'] = cfg.get('num_outputs', 1)
xyz_labs_orig = ['x', 'y', 'likelihood']
suffix = [str(s + 1) for s in range(dlc_cfg['num_outputs'])]
suffix[0] = '' # first one has empty suffix for backwards compatibility
xyz_labs = [x + s for s in suffix for x in xyz_labs_orig]
pdindex = pd.MultiIndex.from_product(
[[DLCscorer], dlc_cfg['all_joints_names'], xyz_labs],
names=['scorer', 'bodyparts', 'coords'])
if gputouse is not None: #gpu selectinon
os.environ['CUDA_VISIBLE_DEVICES'] = str(gputouse)
##################################################
# Loading the images
##################################################
#checks if input is a directory
if os.path.isdir(directory) == True:
"""
Analyzes all the frames in the directory.
"""
print("Analyzing all frames in the directory: ", directory)
os.chdir(directory)
framelist = np.sort(
[fn for fn in os.listdir(os.curdir) if (frametype in fn)])
vname = Path(directory).stem
notanalyzed, dataname, DLCscorer = auxiliaryfunctions.CheckifNotAnalyzed(
directory, vname, DLCscorer, DLCscorerlegacy, flag='framestack')
if notanalyzed:
nframes = len(framelist)
if nframes > 0:
start = time.time()
PredictedData, nframes, nx, ny = GetPosesofFrames(
cfg, dlc_cfg, sess, inputs, outputs, directory, framelist,
nframes, dlc_cfg['batch_size'], rgb)
stop = time.time()
if cfg['cropping'] == True:
coords = [cfg['x1'], cfg['x2'], cfg['y1'], cfg['y2']]
else:
coords = [0, nx, 0, ny]
dictionary = {
"start": start,
"stop": stop,
"run_duration": stop - start,
"Scorer": DLCscorer,
"config file": dlc_cfg,
"batch_size": dlc_cfg["batch_size"],
"num_outputs": dlc_cfg["num_outputs"],
"frame_dimensions": (ny, nx),
"nframes": nframes,
"cropping": cfg['cropping'],
"cropping_parameters": coords
}
metadata = {'data': dictionary}
print("Saving results in %s..." % (directory))
auxiliaryfunctions.SaveData(PredictedData[:nframes, :],
metadata, dataname, pdindex,
framelist, save_as_csv)
print(
"The folder was analyzed. Now your research can truly start!"
)
print(
"If the tracking is not satisfactory for some frome, consider expanding the training set."
)
else:
print(
"No frames were found. Consider changing the path or the frametype."
)
os.chdir(str(start_path))
def analyze_videos(config,
videos,
videotype='avi',
shuffle=1,
trainingsetindex=0,
gputouse=None,
save_as_csv=False,
destfolder=None,
batchsize=None,
crop=None,
get_nframesfrommetadata=True,
TFGPUinference=True,
dynamic=(False, .5, 10)):
"""
Makes prediction based on a trained network. The index of the trained network is specified by parameters in the config file (in particular the variable 'snapshotindex')
Output: The labels are stored as MultiIndex Pandas Array, which contains the name of the network, body part name, (x, y) label position \n
in pixels, and the likelihood for each frame per body part. These arrays are stored in an efficient Hierarchical Data Format (HDF) \n
in the same directory, where the video is stored. However, if the flag save_as_csv is set to True, the data can also be exported in \n
comma-separated values format (.csv), which in turn can be imported in many programs, such as MATLAB, R, Prism, etc.
Parameters
----------
config : string
Full path of the config.yaml file as a string.
videos : list
A list of strings containing the full paths to videos for analysis or a path to the directory, where all the videos with same extension are stored.
videotype: string, optional
Checks for the extension of the video in case the input to the video is a directory.\n Only videos with this extension are analyzed. The default is ``.avi``
shuffle: int, optional
An integer specifying the shuffle index of the training dataset used for training the network. The default is 1.
trainingsetindex: int, optional
Integer specifying which TrainingsetFraction to use. By default the first (note that TrainingFraction is a list in config.yaml).
gputouse: int, optional. Natural number indicating the number of your GPU (see number in nvidia-smi). If you do not have a GPU put None.
See: https://nvidia.custhelp.com/app/answers/detail/a_id/3751/~/useful-nvidia-smi-queries
save_as_csv: bool, optional
Saves the predictions in a .csv file. The default is ``False``; if provided it must be either ``True`` or ``False``
destfolder: string, optional
Specifies the destination folder for analysis data (default is the path of the video). Note that for subsequent analysis this
folder also needs to be passed.
batchsize: int, default from pose_cfg.yaml
Change batch size for inference; if given overwrites value in pose_cfg.yaml
crop: list, optional (default=None)
List of cropping coordinates as [x1, x2, y1, y2].
Note that the same cropping parameters will then be used for all videos.
If different video crops are desired, run 'analyze_videos' on individual videos
with the corresponding cropping coordinates.
TFGPUinference: bool, default: True
Perform inference on GPU with Tensorflow code. Introduced in "Pretraining boosts out-of-domain robustness for pose estimation" by
Alexander Mathis, Mert Yüksekgönül, Byron Rogers, Matthias Bethge, Mackenzie W. Mathis Source: https://arxiv.org/abs/1909.11229
dynamic: triple containing (state,detectiontreshold,margin)
If the state is true, then dynamic cropping will be performed. That means that if an object is detected (i.e. any body part > detectiontreshold),
then object boundaries are computed according to the smallest/largest x position and smallest/largest y position of all body parts. This window is
expanded by the margin and from then on only the posture within this crop is analyzed (until the object is lost, i.e. <detectiontreshold). The
current position is utilized for updating the crop window for the next frame (this is why the margin is important and should be set large
enough given the movement of the animal).
Examples
--------
Windows example for analyzing 1 video
>>> deeplabcutcore.analyze_videos('C:\\myproject\\reaching-task\\config.yaml',['C:\\yourusername\\rig-95\\Videos\\reachingvideo1.avi'])
--------
If you want to analyze only 1 video
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml',['/analysis/project/videos/reachingvideo1.avi'])
--------
If you want to analyze all videos of type avi in a folder:
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml',['/analysis/project/videos'],videotype='.avi')
--------
If you want to analyze multiple videos
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml',['/analysis/project/videos/reachingvideo1.avi','/analysis/project/videos/reachingvideo2.avi'])
--------
If you want to analyze multiple videos with shuffle = 2
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml',['/analysis/project/videos/reachingvideo1.avi','/analysis/project/videos/reachingvideo2.avi'],shuffle=2)
--------
If you want to analyze multiple videos with shuffle = 2 and save results as an additional csv file too
>>> deeplabcutcore.analyze_videos('/analysis/project/reaching-task/config.yaml',['/analysis/project/videos/reachingvideo1.avi','/analysis/project/videos/reachingvideo2.avi'],shuffle=2,save_as_csv=True)
--------
"""
if 'TF_CUDNN_USE_AUTOTUNE' in os.environ:
del os.environ[
'TF_CUDNN_USE_AUTOTUNE'] #was potentially set during training
if gputouse is not None: #gpu selection
os.environ['CUDA_VISIBLE_DEVICES'] = str(gputouse)
tf.compat.v1.reset_default_graph()
start_path = os.getcwd(
) #record cwd to return to this directory in the end
cfg = auxiliaryfunctions.read_config(config)
trainFraction = cfg['TrainingFraction'][trainingsetindex]
if crop is not None:
cfg['cropping'] = True
cfg['x1'], cfg['x2'], cfg['y1'], cfg['y2'] = crop
print("Overwriting cropping parameters:", crop)
print(
"These are used for all videos, but won't be save to the cfg file."
)
modelfolder = os.path.join(
cfg["project_path"],
str(auxiliaryfunctions.GetModelFolder(trainFraction, shuffle, cfg)))
path_test_config = Path(modelfolder) / 'test' / 'pose_cfg.yaml'
try:
dlc_cfg = load_config(str(path_test_config))
except FileNotFoundError:
raise FileNotFoundError(
"It seems the model for shuffle %s and trainFraction %s does not exist."
% (shuffle, trainFraction))
# Check which snapshots are available and sort them by # iterations
try:
Snapshots = np.array([
fn.split('.')[0]
for fn in os.listdir(os.path.join(modelfolder, 'train'))
if "index" in fn
])
except FileNotFoundError:
raise FileNotFoundError(
"Snapshots not found! It seems the dataset for shuffle %s has not been trained/does not exist.\n Please train it before using it to analyze videos.\n Use the function 'train_network' to train the network for shuffle %s."
% (shuffle, shuffle))
if cfg['snapshotindex'] == 'all':
print(
"Snapshotindex is set to 'all' in the config.yaml file. Running video analysis with all snapshots is very costly! Use the function 'evaluate_network' to choose the best the snapshot. For now, changing snapshot index to -1!"
)
snapshotindex = -1
else:
snapshotindex = cfg['snapshotindex']
increasing_indices = np.argsort([int(m.split('-')[1]) for m in Snapshots])
Snapshots = Snapshots[increasing_indices]
print("Using %s" % Snapshots[snapshotindex], "for model", modelfolder)
##################################################
# Load and setup CNN part detector
##################################################
# Check if data already was generated:
dlc_cfg['init_weights'] = os.path.join(modelfolder, 'train',
Snapshots[snapshotindex])
trainingsiterations = (dlc_cfg['init_weights'].split(
os.sep)[-1]).split('-')[-1]
# Update number of output and batchsize
dlc_cfg['num_outputs'] = cfg.get('num_outputs',
dlc_cfg.get('num_outputs', 1))
if batchsize == None:
#update batchsize (based on parameters in config.yaml)
dlc_cfg['batch_size'] = cfg['batch_size']
else:
dlc_cfg['batch_size'] = batchsize
cfg['batch_size'] = batchsize
if dynamic[0]: #state=true
#(state,detectiontreshold,margin)=dynamic
print("Starting analysis in dynamic cropping mode with parameters:",
dynamic)
dlc_cfg['num_outputs'] = 1
TFGPUinference = False
dlc_cfg['batch_size'] = 1
print(
"Switching batchsize to 1, num_outputs (per animal) to 1 and TFGPUinference to False (all these features are not supported in this mode)."
)
# Name for scorer:
DLCscorer, DLCscorerlegacy = auxiliaryfunctions.GetScorerName(
cfg, shuffle, trainFraction, trainingsiterations=trainingsiterations)
if dlc_cfg['num_outputs'] > 1:
if TFGPUinference:
print(
"Switching to numpy-based keypoint extraction code, as multiple point extraction is not supported by TF code currently."
)
TFGPUinference = False
print("Extracting ", dlc_cfg['num_outputs'], "instances per bodypart")
xyz_labs_orig = ['x', 'y', 'likelihood']
suffix = [str(s + 1) for s in range(dlc_cfg['num_outputs'])]
suffix[
0] = '' # first one has empty suffix for backwards compatibility
xyz_labs = [x + s for s in suffix for x in xyz_labs_orig]
else:
xyz_labs = ['x', 'y', 'likelihood']
#sess, inputs, outputs = predict.setup_pose_prediction(dlc_cfg)
if TFGPUinference:
sess, inputs, outputs = predict.setup_GPUpose_prediction(dlc_cfg)
else:
sess, inputs, outputs = predict.setup_pose_prediction(dlc_cfg)
pdindex = pd.MultiIndex.from_product(
[[DLCscorer], dlc_cfg['all_joints_names'], xyz_labs],
names=['scorer', 'bodyparts', 'coords'])
##################################################
# Datafolder
##################################################
Videos = auxiliaryfunctions.Getlistofvideos(videos, videotype)
if len(Videos) > 0:
#looping over videos
for video in Videos:
DLCscorer = AnalyzeVideo(video, DLCscorer, DLCscorerlegacy,
trainFraction, cfg, dlc_cfg, sess, inputs,
outputs, pdindex, save_as_csv, destfolder,
TFGPUinference, dynamic)
os.chdir(str(start_path))
print(
"The videos are analyzed. Now your research can truly start! \n You can create labeled videos with 'create_labeled_video'."
)
print(
"If the tracking is not satisfactory for some videos, consider expanding the training set. You can use the function 'extract_outlier_frames' to extract any outlier frames!"
)
return DLCscorer #note: this is either DLCscorer or DLCscorerlegacy depending on what was used!
else:
os.chdir(str(start_path))
print("No video/s found. Please check your path!")
return DLCscorer
def load_model(cfg, shuffle=1, trainingsetindex=0, TFGPUinference=True, modelprefix=""):
"""
Loads a tensorflow session with a DLC model from the associated configuration
Return a tensorflow session with DLC model given cfg and shuffle
Parameters:
-----------
cfg : dict
Configuration read from the project's main config.yaml file
shuffle : int, optional
which shuffle to use
trainingsetindex : int. optional
which training fraction to use, identified by its index
TFGPUinference : bool, optional
use tensorflow inference model? default = True
Returns:
--------
sess : tensorflow session
tensorflow session with DLC model from the provided configuration, shuffle, and trainingsetindex
checkpoint file path : string
the path to the checkpoint file associated with the loaded model
"""
########################
### find snapshot to use
########################
train_fraction = cfg["TrainingFraction"][trainingsetindex]
model_folder = os.path.join(
cfg["project_path"],
str(
auxiliaryfunctions.GetModelFolder(
train_fraction, shuffle, cfg, modelprefix=modelprefix
)
),
)
path_test_config = os.path.normpath(model_folder + "/test/pose_cfg.yaml")
path_train_config = os.path.normpath(model_folder + "/train/pose_cfg.yaml")
try:
dlc_cfg = load_config(str(path_test_config))
dlc_cfg_train = load_config(str(path_train_config))
except FileNotFoundError:
raise FileNotFoundError(
"It seems the model for shuffle %s and trainFraction %s does not exist."
% (shuffle, train_fraction)
)
# Check which snapshots are available and sort them by # iterations
try:
Snapshots = np.array(
[
fn.split(".")[0]
for fn in os.listdir(os.path.join(model_folder, "train"))
if "index" in fn
]
)
except FileNotFoundError:
raise FileNotFoundError(
"Snapshots not found! It seems the dataset for shuffle %s has not been trained/does not exist.\n Please train it before trying to export.\n Use the function 'train_network' to train the network for shuffle %s."
% (shuffle, shuffle)
)
if len(Snapshots) == 0:
raise FileNotFoundError(
"The train folder for iteration %s and shuffle %s exists, but no snapshots were found.\n Please train this model before trying to export.\n Use the function 'train_network' to train the network for iteration %s shuffle %s."
% (cfg["iteration"], shuffle, cfg["iteration"], shuffle)
)
if cfg["snapshotindex"] == "all":
print(
"Snapshotindex is set to 'all' in the config.yaml file. Changing snapshot index to -1!"
)
snapshotindex = -1
else:
snapshotindex = cfg["snapshotindex"]
increasing_indices = np.argsort([int(m.split("-")[1]) for m in Snapshots])
Snapshots = Snapshots[increasing_indices]
####################################
### Load and setup CNN part detector
####################################
# Check if data already was generated:
dlc_cfg["init_weights"] = os.path.join(
model_folder, "train", Snapshots[snapshotindex]
)
trainingsiterations = (dlc_cfg["init_weights"].split(os.sep)[-1]).split("-")[-1]
dlc_cfg["num_outputs"] = cfg.get("num_outputs", dlc_cfg.get("num_outputs", 1))
dlc_cfg["batch_size"] = None
# load network
if TFGPUinference:
sess, _, _ = predict.setup_GPUpose_prediction(dlc_cfg)
output = ["concat_1"]
else:
sess, _, _ = predict.setup_pose_prediction(dlc_cfg)
if dlc_cfg["location_refinement"]:
output = ["Sigmoid", "pose/locref_pred/block4/BiasAdd"]
else:
output = ["Sigmoid", "pose/part_pred/block4/BiasAdd"]
input = TF.get_default_graph().get_operations()[0].name
return sess, input, output, dlc_cfg_train
