def compile_results(args,views=range(2)):
all_dist = [[],[]]
rims = [[],[]]
rlocs = [[],[]]
rinfo = [[],[]]
rpredlocs = [[],[]]
for view in views:
for split_num in range(3):
args['view'] = view
args['split_num'] = split_num
curconf = get_conf(args)
tf.reset_default_graph()
self = PoseUNet.PoseUNet(curconf,name=get_name(args))
dist, ims, preds, predlocs, locs, info = self.classify_val(0)
all_dist[view].append(dist)
rims[view].append(ims)
rlocs[view].append(locs)
rinfo[view].append(info)
rpredlocs[view].append(predlocs)
all_dist[view] = np.concatenate(all_dist[view],0)
rims[view] = np.concatenate(rims[view],0)
rlocs[view] = np.concatenate(rlocs[view],0)
rinfo[view] = np.concatenate(rinfo[view], 0)
rpredlocs[view] = np.concatenate(rpredlocs[view], 0)
return all_dist, rims, rlocs, rinfo, rpredlocs
def train(args):
split_num = int(args['split_num'])
view = int(args['view'])
split_type = args['split_type']
if 'restore' in args.keys():
restart = ast.literal_eval(args['restore'])
else:
restart = False
conf = get_conf(args)
print('split:{}, view:{} split_type:{}'.format(split_num,view,split_type))
self = PoseUNet.PoseUNet(conf,name=get_name(args))
self.train_unet(restart,0)
def __init__(self, conf, name='pose_u_att', unet_name='pose_unet'):
PoseCommon.PoseCommon.__init__(self, conf, name)
self.dep_nets = PoseUNet.PoseUNet(conf, unet_name)
self.net_name = 'pose_u_att'
self.dep_nets.keep_prob = 1.
self.db_name = '_att'
self.min_score = -1
self.att_layers = conf.att_layers
self.att_hist = conf.att_hist # amount of history for attention
self.joint = True
self.debug_layers = None
self.sel_layers = None
self.sel_layers_ndx = None
def main(argv):
parser = argparse.ArgumentParser()
parser.add_argument("conffile", help="config file", required=True)
parser.add_argument("confname",
help="conf name for importing",
required=True)
parser.add_argument("net_name",
help="Name the network to classify with",
required=True)
parser.add_argument("movie", help="movie to classify", required=True)
parser.add_argument("out name",
help="file to save results to",
required=True)
args = parser.parse_args()
imp_mod = importlib.import_module(args.conffile)
conf = imp_mod.__dict__[args.confname]
oname = re.sub('!', '__', conf.getexpname(args.movie))
self = PoseUNet.PoseUNet(conf, args.net_name)
sess = self.init_net_meta(0, True)
predList = self.classify_movie(args.movie, sess, flipud=True)
cap = cv2.VideoCapture(args.movie)
height = int(cap.get(cvc.FRAME_HEIGHT))
width = int(cap.get(cvc.FRAME_WIDTH))
rescale = conf.unet_rescale
orig_crop_loc = conf.cropLoc[(height, width)]
crop_loc = [int(x / rescale) for x in orig_crop_loc]
end_pad = [
int((height - conf.imsz[0]) / rescale) - crop_loc[0],
int((width - conf.imsz[1]) / rescale) - crop_loc[1]
]
pp = [(0, 0), (crop_loc[0], end_pad[0]), (crop_loc[1], end_pad[1]), (0, 0)]
predScores = np.pad(predList[1], pp, mode='constant', constant_values=-1.)
predLocs = predList[0]
predLocs[:, :, 0] += orig_crop_loc[1]
predLocs[:, :, 1] += orig_crop_loc[0]
hdf5storage.savemat(args.outname, {
'locs': predLocs,
'scores': predScores,
'expname': args.movie
},
appendmat=False,
truncate_existing=True)
print('Done Detecting:%s' % oname)
from stephenHeadConfig import conf as conf import PoseUMDN import tensorflow as tf import PoseTools import math import PoseUNet import cv2 import PoseCommon mov = '/home/mayank/Dropbox (HHMI)/temp/stephen/C002H001S0001_c_2.avi' out = '/home/mayank/temp/C002H001S0001_c_2_res_unet.avi' conf.mdn_min_sigma = 3. conf.mdn_max_sigma = 3. self = PoseUNet.PoseUNet(conf,'pose_unet_droput_0p7') # self = PoseUMDN.PoseUMDN(conf) self.train_unet(False,0) # self.train_umdn(False,0,joint=True) val_dist, val_ims, val_preds, val_predlocs, val_locs= self.classify_val(0,-1) # self.create_pred_movie(mov,out,flipud=True) ## tf.reset_default_graph() self.init_train(0) self.pred = self.create_network() self.create_saver() sess = tf.InteractiveSession() self.init_and_restore(sess,True,['loss','dist']) info= []
def main(argv):
# defaulttrackerpath = "/groups/branson/home/bransonk/tracking/code/poseTF/matlab/compute3Dfrom2D/for_redistribution_files_only/run_compute3Dfrom2D.sh"
defaulttrackerpath = "/groups/branson/bransonlab/mayank/PoseTF/matlab/compiled/run_compute3Dfrom2D_compiled.sh"
# defaultmcrpath = "/groups/branson/bransonlab/projects/olympiad/MCR/v91"
defaultmcrpath = "/groups/branson/bransonlab/mayank/MCR/v92"
parser = argparse.ArgumentParser()
parser.add_argument("-s", dest="sfilename",
help="text file with list of side view videos",
required=True)
parser.add_argument("-f", dest="ffilename",
help="text file with list of front view videos. The list of side view videos and front view videos should match up",
required=True)
parser.add_argument("-d", dest="dltfilename",
help="text file with list of DLTs, one per fly as 'flynum,/path/to/dltfile'",
required=True)
parser.add_argument("-o", dest="outdir",
help="temporary output directory to store intermediate computations",
required=True)
parser.add_argument("-r", dest="redo",
help="if specified will recompute everything",
action="store_true")
parser.add_argument("-rt", dest="redo_tracking",
help="if specified will only recompute tracking",
action="store_true")
parser.add_argument("-gpu", dest='gpunum',
help="GPU to use [optional]")
parser.add_argument("-makemovie", dest='makemovie',
help="if specified will make results movie", action="store_true")
parser.add_argument("-trackerpath", dest='trackerpath',
help="Absolute path to the compiled MATLAB tracker script run_compute3Dfrom2D.sh",
default=defaulttrackerpath)
parser.add_argument("-mcrpath", dest='mcrpath',
help="Absolute path to MCR",
default=defaultmcrpath)
parser.add_argument("-ncores", dest="ncores",
help="Number of cores to assign to each MATLAB tracker job",
default=1)
parser.add_argument("-name", dest="name",
help="Name of the detector",
default='pose_unet')
group = parser.add_mutually_exclusive_group()
group.add_argument("-only_detect", dest='detect', action="store_true",
help="Do only the detection part of tracking which requires GPU")
group.add_argument("-only_track", dest='track', action="store_true",
help="Do only the tracking part of the tracking which requires MATLAB")
args = parser.parse_args()
if args.redo is None:
args.redo = False
if args.redo_tracking is None:
args.redo_tracking = False
if args.detect is False and args.track is False:
args.detect = True
args.track = True
args.outdir = os.path.abspath(args.outdir)
with open(args.sfilename, "r") as text_file:
smovies = text_file.readlines()
smovies = [x.rstrip() for x in smovies]
with open(args.ffilename, "r") as text_file:
fmovies = text_file.readlines()
fmovies = [x.rstrip() for x in fmovies]
print(smovies)
print(fmovies)
print(len(smovies))
print(len(fmovies))
if args.track:
if len(smovies) != len(fmovies):
print("Side and front movies must match")
raise exit(0)
# read in dltfile
dltdict = {}
f = open(args.dltfilename, 'r')
for l in f:
lparts = l.split(',')
if len(lparts) != 2:
print("Error splitting dlt file line %s into two parts" % l)
raise exit(0)
dltdict[float(lparts[0])] = lparts[1].strip()
f.close()
# compiled matlab command
matscript = args.trackerpath + " " + args.mcrpath
if args.detect:
import numpy as np
import tensorflow as tf
from scipy import io
from cvc import cvc
import localSetup
import PoseTools
import multiResData
import cv2
import PoseUNet
import PoseCommon
for ff in smovies + fmovies:
if not os.path.isfile(ff):
print("Movie %s not found" % (ff))
raise exit(0)
if args.gpunum is not None:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpunum
for view in range(2): # 0 for front and 1 for side
if args.detect:
tf.reset_default_graph()
if view == 1: # For Side
from stephenHeadConfig import sideconf as conf
extrastr = '_side'
valmovies = smovies
else: # For FRONT
from stephenHeadConfig import conf as conf
extrastr = '_front'
valmovies = fmovies
if args.detect:
self = PoseUNet.PoseUNet(conf,args.name)
self.create_ph_fd()
self.create_fd()
self.pred = self.create_network()
sess = tf.Session()
self.create_saver()
self.restore(sess, True)
PoseCommon.initialize_remaining_vars(sess)
for ndx in range(len(valmovies)):
mname, _ = os.path.splitext(os.path.basename(valmovies[ndx]))
oname = re.sub('!', '__', conf.getexpname(valmovies[ndx]))
pname = os.path.join(args.outdir, oname + extrastr)
print(oname)
# detect
if args.detect and os.path.isfile(valmovies[ndx]) and \
(args.redo or not os.path.isfile(pname + '.mat')):
predList = self.classify_movie(valmovies[ndx], sess)
if args.makemovie:
PoseTools.create_pred_movie(conf, predList, valmovies[ndx], pname + '.avi', outtype)
cap = cv2.VideoCapture(valmovies[ndx])
height = int(cap.get(cvc.FRAME_HEIGHT))
width = int(cap.get(cvc.FRAME_WIDTH))
orig_crop_loc = conf.cropLoc[(height, width)]
crop_loc = [old_div(x, 4) for x in orig_crop_loc]
end_pad = [old_div(height, 4) - crop_loc[0] - old_div(conf.imsz[0], 4),
old_div(width, 4) - crop_loc[1] - old_div(conf.imsz[1], 4)]
pp = [(0, 0), (crop_loc[0], end_pad[0]), (crop_loc[1], end_pad[1]), (0, 0), (0, 0)]
predScores = np.pad(predList[1], pp, mode='constant', constant_values=-1.)
predLocs = predList[0]
predLocs[:, :, :, 0] += orig_crop_loc[1]
predLocs[:, :, :, 1] += orig_crop_loc[0]
io.savemat(pname + '.mat', {'locs': predLocs, 'scores': predScores[..., 0], 'expname': valmovies[ndx]})
print('Detecting:%s' % oname)
# track
if args.track and view == 1:
oname_side = re.sub('!', '__', conf.getexpname(smovies[ndx]))
oname_front = re.sub('!', '__', conf.getexpname(fmovies[ndx]))
pname_side = os.path.join(args.outdir, oname_side + '_side.mat')
pname_front = os.path.join(args.outdir, oname_front + '_front.mat')
# 3d trajectories
basename_front, _ = os.path.splitext(fmovies[ndx])
basename_side, _ = os.path.splitext(smovies[ndx])
savefile = basename_side + '_3Dres.mat'
# savefile = os.path.join(args.outdir , oname_side + '_3Dres.mat')
trkfile_front = basename_front + '.trk'
trkfile_side = basename_side + '.trk'
redo_tracking = args.redo or args.redo_tracking
if os.path.isfile(savefile) and os.path.isfile(trkfile_front) and \
os.path.isfile(trkfile_side) and not redo_tracking:
print("%s, %s, and %s exist, skipping tracking" % (savefile, trkfile_front, trkfile_side))
continue
flynum = conf.getflynum(smovies[ndx])
# print "Parsed fly number as %d"%flynum
kinematfile = os.path.abspath(dltdict[flynum])
jobid = oname_side
scriptfile = os.path.join(args.outdir, jobid + '_track.sh')
logfile = os.path.join(args.outdir, jobid + '_track.log')
# print "matscript = " + matscript
# print "pname_front = " + pname_front
# print "pname_side = " + pname_side
# print "kinematfile = " + kinematfile
# make script to be qsubbed
scriptf = open(scriptfile, 'w')
scriptf.write('if [ -d %s ]\n' % args.outdir)
scriptf.write(' then export MCR_CACHE_ROOT=%s/mcrcache%s\n' % (args.outdir, jobid))
scriptf.write('fi\n')
scriptf.write('%s "%s" "%s" "%s" "%s" "%s" "%s"\n' % (
matscript, savefile, pname_front, pname_side, kinematfile, trkfile_front, trkfile_side))
scriptf.close()
os.chmod(scriptfile,
stat.S_IRUSR | stat.S_IRGRP | stat.S_IWUSR | stat.S_IWGRP | stat.S_IXUSR | stat.S_IXGRP)
cmd = "ssh login1 'source /etc/profile; qsub -pe batch %d -N %s -j y -b y -o '%s' -cwd '\"%s\"''" % (
args.ncores, jobid, logfile, scriptfile)
print(cmd)
call(cmd, shell=True)
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/witinAssayData/cx_JRC_SS03500_CsChr_RigB_20150811T114536',
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/witinAssayData/cx_GMR_SS00243_CsChr_RigD_20150812T155340'
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/withGenotypeData/cx_GMR_SS00030_CsChr_RigD_20150826T145307'
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/withGenotypeData/cx_GMR_SS00038_CsChr_RigC_20150908T140450'
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/withGenotypeData/cx_GMR_SS00168_CsChr_RigC_20150909T110456',
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/withGenotypeData/cx_GMR_SS00238_CsChr_RigD_20150826T143306',
#'/groups/branson/home/robiea/Projects_data/Labeler_APT/Austin_labelerprojects_expandedbehaviors/GT/withGenotypeData/cx_JHS_K_85321_CsChr_RigB_20151021T095119',
dirs = [
'/groups/branson/home/robiea/Projects_data/Labeler_APT/cx_GMR_SS00038_CsChr_RigB_20150729T150617',
'/groups/branson/home/robiea/Projects_data/Labeler_APT/cx_GMR_SS00030_CsChr_RigC_20150826T144616',
'/groups/branson/home/robiea/Projects_data/Labeler_APT/cx_JHS_K_85321_CsChr_RigD_20150909T163219',
'/groups/branson/home/robiea/Projects_data/Labeler_APT/cx_GMR_SS00168_CsChr_RigD_20150909T111218',
'/groups/branson/home/robiea/Projects_data/Labeler_APT/cx_GMR_SS00238_CsChr_RigC_20151007T150343',
]
self = PoseUNet.PoseUNet(conf, name='pose_unet_full_20180521')
sess = self.init_net_meta(1)
model_file = self.get_latest_model_file()
def pred_fn(all_f):
bsize = conf.batch_size
xs, _ = PoseTools.preprocess_ims(all_f,
in_locs=np.zeros(
[bsize, self.conf.n_classes, 2]),
conf=self.conf,
distort=False,
scale=self.conf.unet_rescale)
self.fd[self.ph['x']] = xs
self.fd[self.ph['phase_train']] = False
import PoseTools
self = PoseTrain.PoseTrain(conf)
val_dist, val_ims, val_preds, val_predlocs, val_locs = self.classify_val_base()
kk = np.percentile(val_dist, [80, 90, 95], axis=0)
pp = np.transpose(val_ims[0, ...].astype('uint8'), [1, 2, 0])
PoseTools.create_result_image(pp, val_locs[0, ...], kk)
## U-Net.
from poseConfig import felipeConfig as conf
import PoseUNet
conf.mdn_min_sigma = 10.
conf.mdn_max_sigma = 15.
self = PoseUNet.PoseUNet(conf, name='pose_unet')
self.train_unet(restore=False, train_type=0)
## U-Net results
from poseConfig import felipeConfig as conf
import PoseUNet
import os
import PoseTools
os.environ['CUDA_VISIBLE_DEVICES'] = ''
conf.mdn_min_sigma = 10.
conf.mdn_max_sigma = 15.
self = PoseUNet.PoseUNet(conf, name='pose_unet')
val_dist, val_ims, val_preds, val_predlocs, val_locs = self.classify_val()
## for debug..
import os
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
import PoseUNet
from poseConfig import aliceConfig as conf
import h5py
import multiResData
import tensorflow as tf
import numpy as np
import PoseTools
import pickle
from scipy import io as sio
import math
self = PoseUNet.PoseUNet(conf, name='pose_unet_128_bs8')
_, dirs, _ = multiResData.load_val_data(conf)
for ndx in range(len(dirs)):
dirs[ndx] = dirs[ndx].replace('$dataroot',
'/home/mayank/work/FlySpaceTime')
trx_file = dirs[2].replace('movie.ufmf', 'registered_trx.mat')
self.create_pred_movie_trx(dirs[2],
'/home/mayank/temp/aliceOut_JHS_fly6_0_300.avi',
trx_file,
6,
max_frames=300,
start_at=0,
trace=True)
##