def getData():
Dpath = '/home/denitome/PythonProjects/sfm_new/'
json_file = ut.getCaffeCpm(
) + '/jsonDatasets/H36M_annotations_testSet.json'
(data, num_elem) = ut.loadJsonFile(json_file)
# 2d predictions
test = ut.sio.loadmat(Dpath + 'Data/testSet_with_predictions.mat')
mask_camera = test['mask_camera'][0]
mask_action = test['mask_action'][0]
preds = test['pred']
# 3d predictions
p_res_dict = ut.sio.loadmat(Dpath + 'predicted_3d_poses_testset.mat')
preds_3d = p_res_dict['pred_models']
# mask person
idx = np.where(mask_action == 2)[0]
t = np.where((idx[:-1] + 1) != idx[1:])[0][0] + 1
fno_person = idx[t]
mask_person = np.ones(num_elem) * 9
mask_person[fno_person:] = 11
# all masks
masks = {}
masks['camera'] = mask_camera
masks['action'] = mask_action
masks['person'] = mask_person
return (data, preds, preds_3d, masks)
def getLossOnValidationSet(NN, models_dir, json_file, masks, prototxt, output,
samplingRate):
# get models
files = [f for f in ut.os.listdir(models_dir) if f.endswith('.caffemodel')]
files = sorted(files, key=getIter)
print 'Loading json file with annotations...'
(data, num_elem) = ut.loadJsonFile(json_file)
print 'Loading mask file...'
masks = ut.sio.loadmat(masks)
print 'Done.'
results = [dict() for x in range(len(files))]
for i in range(len(files)):
model_dir = '%s/%s' % (models_dir, files[i])
iterNumber = getIter(model_dir)
print '-------------------------------'
print ' Evaluating iteration: %d' % iterNumber
print '-------------------------------'
net = ut.loadNetFromPath(model_dir, prototxt)
val = runCaffeOnModel(NN, net, data[0:num_elem:samplingRate], masks,
iterNumber)
results[i] = val
# save json
with open(output, 'w+') as out:
json.dump(results, out)
return results
def getLossOnValidationSet(NN, models_dir, json_file, masks, prototxt, output, samplingRate):
# get models
files = [f for f in ut.os.listdir(models_dir) if f.endswith('.caffemodel')]
files = sorted(files, key=getIter)
print 'Loading json file with annotations...'
(data, num_elem) = ut.loadJsonFile(json_file)
print 'Loading mask file...'
masks = ut.sio.loadmat(masks)
print 'Done.'
results = [dict() for x in range(len(files))]
for i in range(len(files)):
model_dir = '%s/%s' % (models_dir, files[i])
iterNumber = getIter(model_dir)
print '-------------------------------'
print ' Evaluating iteration: %d' % iterNumber
print '-------------------------------'
net = ut.loadNetFromPath(model_dir, prototxt)
val = runCaffeOnModel(NN, net, data[0:num_elem:samplingRate], masks, iterNumber)
results[i] = val
# save json
with open(output, 'w+') as out:
json.dump(results, out)
return results
def produceInput(NN):
json_file = ut.getCaffeCpm() + '/jsonDatasets/H36M_annotations_testSet.json'
(data, num_elem) = ut.loadJsonFile(json_file)
curr_data = data[0]
img = ut.cv2.imread(curr_data['img_paths'])
joints = np.array(curr_data['joint_self'])
center = ut.getCenterJoint(joints)
img_width = img.shape[1]
img_height = img.shape[0]
box_points = ut.getBoundingBox(joints, center, NN['offset'], img_width, img_height)
# manipulate image and joints for caffe
(img_croppad, joints) = ut.cropImage(img, box_points, joints)
(resizedImage, joints) = ut.resizeImage(img_croppad, NN['inputSize'], joints)
resizedImage = np.divide(resizedImage, float(256))
resizedImage = np.subtract(resizedImage, 0.5)
# generate labels and center channel
input_size = NN['inputSize']
center_channel = ut.generateGaussian(NN['sigma_center'], input_size, [input_size/2, input_size/2])
fno = int(curr_data['annolist_index'])
camera = curr_data['camera']
action = curr_data['action']
person = curr_data['person']
metadata_ch = ut.generateMaskChannel(NN['inputSize'], fno, camera, action, person)
imgch = np.concatenate((resizedImage, center_channel[:,:,np.newaxis], metadata_ch), axis=2)
imgch = np.transpose(imgch, (2, 0, 1))
return imgch
def main():
parser = argparse.ArgumentParser()
parser.add_argument('prototxt', metavar='prototxt', type=str, help='Path to the prototxt file')
parser.add_argument('caffemodel', metavar='caffemodel', type=str, help='Path to the caffemodel')
parser.add_argument('-o', dest='out_directory', help='Output directory where to store the predictions')
parser.add_argument('--json_file', dest='json_file', type=str, help='Path to the json file containing the test-set data')
parser.add_argument('---with_cpu', action='store_true', dest='with_cpu', help='Caffe uses CPU for feature extraction')
parser.add_argument('--num_parts', default=1, type=int, dest='num_parts', help='Split the test set into num_parts. Default set to 1')
parser.add_argument('--run_part', default=1, type=int, dest='run_part', help='Run part of the splitted test set. Default set to 1')
parser.add_argument('--merge_parts_dir', dest='merge_parts_dir', type=str, help='Merge part contained in the directory')
args = parser.parse_args()
checkFilesExistance(args.prototxt, args.caffemodel)
# set environment
NN = ut.load_configuration(gpu=(not args.with_cpu))
ut.setCaffeMode(NN['GPU'])
# get json file path
json_file = ut.getCaffeCpm() + '/jsonDatasets/H36M_annotations_testSet.json'
if args.json_file is not None:
json_file = args.json_file
# load caffe model
net = ut.loadNetFromPath(args.caffemodel, args.prototxt)
(data, num_elem) = ut.loadJsonFile(json_file)
# Execution in multiple machines
elems_per_part = int(np.floor(num_elem/int(args.num_parts)))
offset_data = 0 # for not perfect divisions
offset = elems_per_part*(int(args.run_part)-1) # depending the part we are running
if (args.run_part == args.num_parts):
offset_data = num_elem - elems_per_part*int(args.num_parts)
idx_part = range(offset, offset + elems_per_part + offset_data)
# Set output file path
output_file = getNameOutputFile(args.caffemodel)
if args.out_directory is not None:
output_file = args.out_directory
output_file += '/predictions.mat'
if args.merge_parts_dir is not None:
(preds, errors, covs, frame_num) = mergeParts(NN, args.merge_parts_dir, num_elem, elems_per_part)
ut.sio.savemat(output_file, {'preds':preds,'errors':errors,'covs':covs,'frame_num':frame_num})
print 'mean error: %r' % errors.mean()
return
# run model on test set
(preds, errors, covs, frame_num) = executeOnTestSet(NN, net, data[idx_part], elems_per_part)
print 'Mean error on the test set: %r' % errors.mean()
# save results
if (args.num_parts > 1):
output_file = re.sub('\.caffemodel', '_predictions_part', args.caffemodel)
output_file += '%d.mat' % (int(args.run_part))
ut.sio.savemat(output_file, {'preds':preds,'errors':errors,'covs':covs,'frame_num':frame_num})
def produceInput(NN):
json_file = ut.getCaffeCpm(
) + '/jsonDatasets/H36M_annotations_testSet.json'
(data, num_elem) = ut.loadJsonFile(json_file)
curr_data = data[0]
img = ut.cv2.imread(curr_data['img_paths'])
joints = np.array(curr_data['joint_self'])
center = ut.getCenterJoint(joints)
img_width = img.shape[1]
img_height = img.shape[0]
box_points = ut.getBoundingBox(joints, center, NN['offset'], img_width,
img_height)
# manipulate image and joints for caffe
(img_croppad, joints) = ut.cropImage(img, box_points, joints)
(resizedImage, joints) = ut.resizeImage(img_croppad, NN['inputSize'],
joints)
resizedImage = np.divide(resizedImage, float(256))
resizedImage = np.subtract(resizedImage, 0.5)
# generate labels and center channel
input_size = NN['inputSize']
center_channel = ut.generateGaussian(NN['sigma_center'], input_size,
[input_size / 2, input_size / 2])
fno = int(curr_data['annolist_index'])
camera = curr_data['camera']
action = curr_data['action']
person = curr_data['person']
metadata_ch = ut.generateMaskChannel(NN['inputSize'], fno, camera, action,
person)
imgch = np.concatenate(
(resizedImage, center_channel[:, :, np.newaxis], metadata_ch), axis=2)
imgch = np.transpose(imgch, (2, 0, 1))
return imgch
def getData():
Dpath = '/home/denitome/PythonProjects/sfm_new/'
json_file = ut.getCaffeCpm() + '/jsonDatasets/H36M_annotations_testSet.json'
(data, num_elem) = ut.loadJsonFile(json_file)
# 2d predictions
test = ut.sio.loadmat(Dpath+'Data/testSet_with_predictions.mat')
mask_camera = test['mask_camera'][0]
mask_action = test['mask_action'][0]
preds = test['pred']
# 3d predictions
p_res_dict = ut.sio.loadmat(Dpath+'predicted_3d_poses_testset.mat')
preds_3d = p_res_dict['pred_models']
# mask person
idx = np.where(mask_action == 2)[0]
t = np.where((idx[:-1]+1) != idx[1:])[0][0] + 1
fno_person = idx[t]
mask_person = np.ones(num_elem)*9
mask_person[fno_person:] = 11
# all masks
masks = {}
masks['camera'] = mask_camera
masks['action'] = mask_action
masks['person'] = mask_person
return (data, preds, preds_3d, masks)
return i+fno
return -1
## SET ENVIRONMENT
caffemodel = ut.getCaffeCpm() + '/prototxt/caffemodel/manifold_diffarch3/pose_iter_22000.caffemodel'
prototxt = ut.getCaffeCpm() + '/prototxt/pose_deploy_singleimg.prototxt'
output_dir = '/home/denitome/Desktop/imgs/tmp/'
checkFilesExistance(prototxt, caffemodel)
NN = ut.load_configuration(gpu=True)
ut.setCaffeMode(NN['GPU'])
# load caffe model
json_file = ut.getCaffeCpm() + '/jsonDatasets/H36M_annotations_testSet.json'
net = ut.loadNetFromPath(caffemodel, prototxt)
(data, num_elem) = ut.loadJsonFile(json_file)
idx = getIndex(data, camera=1, person=9, action=2, fno=10)
#idx = getIndex(data, camera=1, person=11, action=14, fno=950)
executeOnFrame(NN, net, data[idx], output_dir, proj=True, astr='_tmp', show=True, stage_3d_err=True, hm_joint=6)
#tmp_data = data
#data=data[idx]
# TODO: remove
##differr = errors_orig-errors
##idx = np.argmax(differr[differr < 31.76])
#phi = np.log(1+(errors_orig -errors)/errors3d)
#idx = np.where(phi == phi[phi < 0.84].max())[0][0]
#executeOnFrame(NN, net, data[idx], output_dir, proj=True, astr=('_%d' % idx), show=True, stage_3d_err=False)
## SET ENVIRONMENT
caffemodel = ut.getCaffeCpm(
) + '/prototxt/caffemodel/manifold_diffarch3/pose_iter_22000.caffemodel'
prototxt = ut.getCaffeCpm() + '/prototxt/pose_deploy_singleimg.prototxt'
output_dir = '/home/denitome/Desktop/imgs/tmp/'
checkFilesExistance(prototxt, caffemodel)
NN = ut.load_configuration(gpu=True)
ut.setCaffeMode(NN['GPU'])
# load caffe model
json_file = ut.getCaffeCpm() + '/jsonDatasets/H36M_annotations_testSet.json'
net = ut.loadNetFromPath(caffemodel, prototxt)
(data, num_elem) = ut.loadJsonFile(json_file)
idx = getIndex(data, camera=1, person=9, action=2, fno=10)
#idx = getIndex(data, camera=1, person=11, action=14, fno=950)
executeOnFrame(NN,
net,
data[idx],
output_dir,
proj=True,
astr='_tmp',
show=True,
stage_3d_err=True,
hm_joint=6)
#tmp_data = data
#data=data[idx]