def eval():
# global pred_file_name
# pred_file_name = fileName.split('/')
# pred_file_name = pred_file_name[-1].replace(".zip","_prediction.csv")
# print fileName
files = glob(data + "/" + "*.zip")
files.sort()
print len(files), "found"
for fileName in files[669 - 470:]:
# pred_file_name = fileName.split('/')
# pred_file_name = pred_file_name[-1].replace(".zip","_prediction.csv")
print fileName
# s_time = time.time()
smp = pp.GestureSample(fileName)
# print "loading", (time.time()-s_time)/1000.,"ms"
s_time = time.time()
n = smp.data['numFrames']
dv, uv, gv = smp.depth, smp.user, smp.rgb
cur_fr = 1
new_shape = (step, 128, 128)
s = []
d, u, g = [empty((n_f, ) + vid_res + (3, ), "uint8") for _ in range(3)]
# take first n_f frames
for v in dv, uv, gv:
pp.go_to_frame(v, cur_fr)
for i, fr in enumerate(range(cur_fr, cur_fr + n_f)):
s.append(smp.getSkeleton(fr))
d[i], u[i], g[i] = [v.read()[1] for v in dv, uv, gv]
d, u, g = [pp.to_grayscale(v) for v in d, u, g]
u[u < 128], u[u >= 128] = 0, 1
depth, user, gray, skelet = d, u, g, s
user_o = user.copy()
depth_o = depth.copy()
gray_o = gray.copy()
user_depth = depth_o[user_o == 1]
skelet, c = pp.proc_skelet(array(skelet).copy())
user = pp.proc_user(user)
_, depth, c = pp.proc_depth(depth.copy(), user.copy(), user_o,
array(skelet).copy())
gray, c = pp.proc_gray(
gray.copy(), user,
array(skelet).copy()) #user.copy!!!!!!!!!!!!!!!!!!!
# video = array([gray,depth],dtype="uint8")
# video = video.swapaxes(0,1) #(body-hand,gray-depth,fr,h,w)
# video = video[0:1].astype("uint8")
# x_.set_value(video,borrow=True)
# pred = evalu_model()[0][0]
# print pred.argmax()+1,"\t", "\t",pred.max(),"\t",cur_fr, "-",cur_fr+n_f
cur_fr += n_f
while cur_fr + step < n:
# time_start = time.time()
sn = []
dn, un, gn = [
empty((step, ) + vid_res + (3, ), "uint8") for _ in range(3)
]
# for v in dv,uv,gv: pp.go_to_frame(v, cur_fr)
for i, fr in enumerate(range(cur_fr, cur_fr + step)):
sn.append(smp.getSkeleton(fr))
dn[i], un[i], gn[i] = [v.read()[1] for v in dv, uv, gv]
dn, un, gn = [pp.to_grayscale(v) for v in dn, un, gn]
un[un < 128], un[un >= 128] = 0, 1
s = s[step:] + sn
# s.extend(sn)
skelet, c = pp.proc_skelet(s, _3D=False)
len_dump = len(depth_o[:step][user_o[:step] == 1])
un_d = dn[un == 1]
user_o[:-step] = user_o[step:]
user_o[-step:] = un.copy()
un = pp.proc_user(un, 3)
user[:-step] = user[step:]
user[-step:] = un.copy()
depth_o[:-step] = depth_o[step:]
depth_o[-step:] = dn.copy()
gray_o[:-step] = gray_o[step:]
gray_o[-step:] = gn.copy()
_, depth, c = pp.proc_depth(depth_o.copy(), user.copy(), user_o,
skelet)
gray, c = pp.proc_gray(gray_o.copy(), user, skelet)
traj2D, traj3D, ori, pheight, hand, center = skelet
# # ---------------------------- DEPTH ------------------------------------------
# thresh_noise = 200
# scaler = 4
# traj2D,traj3D,ori,pheight,hand,center = skelet
# traj2D = traj2D[:,-step:]
# # print traj2D.shape
# #stats
# # user_depth = depth_o[user_o==1] #11.1%
# user_depth = hstack((user_depth[len_dump:],un_d))
# dn_ = dn.copy()
# med = average(user_depth)
# med = 255 - med
# std = user_depth.std()
# depth_b = pp.cut_body(dn_, center, pheight, hand,new_shape)
# # user_b = pp.cut_body(un.copy(), center, pheight, hand,new_shape)
# depth_h = pp.cut_hand(dn_, traj2D, hand,new_shape)
# # user_h = pp.cut_hand(un.copy(), traj2D, hand,new_shape)
# new_d = empty((2,step,128,128), dtype="uint8")
# for i,d in enumerate((depth_b,depth_h)):
# d = 255 - d
# # nuser_d = d[us==0]
# # nuser_d[nuser_d>thresh_noise] = 0
# # d[us==0] = nuser_d
# d[d>thresh_noise] = 0
# #d = inpaint(d, thresh_noise)
# thresh_d = med-3*std
# # d[d<thresh_d] = thresh_d
# d = d-thresh_d
# d = clip(d*scaler, 0, 255)
# d = d.astype("uint8")
# d = pp.medianblur(d)
# new_d[i] = d
# dn = new_d.astype("uint8")
# depth[:,:-step]=depth[:,step:]
# depth[:,-step:] = dn.copy()
# # ---------------------------- DEPTH ------------------------------------------
# # ---------------------------- GRAY ------------------------------------------
# gray_b = pp.cut_body(gn.copy(), center, pheight, hand, new_shape)
# gray_h = pp.cut_hand(gn.copy(), traj2D, hand,new_shape)
# new_gray = empty((2,step,128,128), dtype="uint8")
# new_gray[0] = gray_b
# new_gray[1] = gray_h
# gn = new_gray.astype("uint8")
# gray[:,:-step]=gray[:,step:]
# gray[:,-step:] = gn.copy()
# # ---------------------------- GRAY ------------------------------------------
video = empty((
1,
2,
) + gray.shape, dtype="uint8")
video[0, 0] = gray.copy()
video[0, 1] = depth.copy()
video = video.swapaxes(1, 2) #(body-hand,gray-depth,fr,h,w)
# if hand == "left":
# for i in range(video.shape[1]):
# for j in range(video.shape[2]):
# for k in range(video.shape[3]):
# video[0,i,j,k] = cv2.flip(video[0,i,j,k],1)
# video = video[0:1].astype("uint8")
# ---------------------------AUGMENT--------------------------------------
v_new = empty((1, 2, 2) + vid_shape, dtype="uint8")
# p = pheight
ratio = 0.25
for i in xrange(video.shape[0]): #batch
if pheight < 10: pheight = 100
scale = ratio + randi(2) / 100.
ofs = pheight * scale
mid = video.shape[-1] / 2.
sli = None
if ofs < mid:
start = int(round(mid - ofs))
end = int(round(mid + ofs))
sli = slice(start, end)
for j in xrange(video.shape[2]): #maps
for k in xrange(video.shape[3]): #frames
#body
img = video[i, 0, j, k]
img = cut_img(img, 5)
img = misc.imresize(img, (h, h))
# if j==0: img = 255-misc.imfilter(img,"contour")
v_new[i, 0, j, k] = img
#hand
img = video[i, 1, j, k]
img = img[sli, sli]
img = misc.imresize(img, (h, h))
v_new[i, 1, j, k] = img
# ---------------------------AUGMENT--------------------------------------
"""
x_.set_value(v_new.astype("float32"),borrow=True)
print "predicting"
pred = evalu_model()[0][0]
pred_p = pred.max()
pred_idx = pred.argmax()+1
fps = int(1./((time.time()-time_start)/step))
fr_start = cur_fr+step-n_f
fr_end = cur_fr+step
print pred_idx,"\t", "\t",pred_p,"\t",fr_start, "-",fr_end,"\t",fps,'fps'
predict(pred_idx,pred_p,fr_start)
# print v_new.shape
# for i in xrange(v_new.shape[0]):
# for j in xrange(v_new.shape[1]):
# for k in xrange(v_new.shape[2]):
# play_vid(v_new[i,j,k],wait=0)
"""
# print "put"
q.put((v_new, cur_fr, n, fileName))
cur_fr += step
def eval():
# global pred_file_name
# pred_file_name = fileName.split('/')
# pred_file_name = pred_file_name[-1].replace(".zip","_prediction.csv")
# print fileName
files = glob(data+"/"+"*.zip")
files.sort()
print len(files), "found"
for fileName in files[669-470:]:
# pred_file_name = fileName.split('/')
# pred_file_name = pred_file_name[-1].replace(".zip","_prediction.csv")
print fileName
# s_time = time.time()
smp = pp.GestureSample(fileName)
# print "loading", (time.time()-s_time)/1000.,"ms"
s_time = time.time()
n = smp.data['numFrames']
dv,uv,gv = smp.depth, smp.user, smp.rgb
cur_fr = 1
new_shape = (step,128,128)
s = []
d,u,g = [empty((n_f,)+vid_res+(3,), "uint8") for _ in range(3)]
# take first n_f frames
for v in dv,uv,gv: pp.go_to_frame(v, cur_fr)
for i,fr in enumerate(range(cur_fr,cur_fr+n_f)):
s.append(smp.getSkeleton(fr))
d[i],u[i],g[i] = [v.read()[1] for v in dv,uv,gv]
d,u,g = [pp.to_grayscale(v) for v in d,u,g]
u[u<128], u[u>=128] = 0, 1
depth,user,gray,skelet = d,u,g,s
user_o = user.copy()
depth_o = depth.copy()
gray_o = gray.copy()
user_depth = depth_o[user_o==1]
skelet,c =pp.proc_skelet(array(skelet).copy())
user = pp.proc_user(user)
_,depth,c = pp.proc_depth(depth.copy(), user.copy(), user_o, array(skelet).copy())
gray,c = pp.proc_gray(gray.copy(), user, array(skelet).copy()) #user.copy!!!!!!!!!!!!!!!!!!!
# video = array([gray,depth],dtype="uint8")
# video = video.swapaxes(0,1) #(body-hand,gray-depth,fr,h,w)
# video = video[0:1].astype("uint8")
# x_.set_value(video,borrow=True)
# pred = evalu_model()[0][0]
# print pred.argmax()+1,"\t", "\t",pred.max(),"\t",cur_fr, "-",cur_fr+n_f
cur_fr += n_f
while cur_fr+step<n:
# time_start = time.time()
sn=[]
dn,un,gn = [empty((step,)+vid_res+(3,), "uint8") for _ in range(3)]
# for v in dv,uv,gv: pp.go_to_frame(v, cur_fr)
for i,fr in enumerate(range(cur_fr,cur_fr+step)):
sn.append(smp.getSkeleton(fr))
dn[i],un[i],gn[i] = [v.read()[1] for v in dv,uv,gv]
dn,un,gn = [pp.to_grayscale(v) for v in dn,un,gn]
un[un<128], un[un>=128] = 0,1
s = s[step:] + sn
# s.extend(sn)
skelet,c =pp.proc_skelet(s,_3D=False)
len_dump = len(depth_o[:step][user_o[:step]==1])
un_d = dn[un==1]
user_o[:-step]=user_o[step:]
user_o[-step:] = un.copy()
un = pp.proc_user(un,3)
user[:-step]=user[step:]
user[-step:] = un.copy()
depth_o[:-step]=depth_o[step:]
depth_o[-step:] = dn.copy()
gray_o[:-step]=gray_o[step:]
gray_o[-step:] = gn.copy()
_,depth,c = pp.proc_depth(depth_o.copy(), user.copy(), user_o, skelet)
gray,c = pp.proc_gray(gray_o.copy(), user, skelet)
traj2D,traj3D,ori,pheight,hand,center = skelet
# # ---------------------------- DEPTH ------------------------------------------
# thresh_noise = 200
# scaler = 4
# traj2D,traj3D,ori,pheight,hand,center = skelet
# traj2D = traj2D[:,-step:]
# # print traj2D.shape
# #stats
# # user_depth = depth_o[user_o==1] #11.1%
# user_depth = hstack((user_depth[len_dump:],un_d))
# dn_ = dn.copy()
# med = average(user_depth)
# med = 255 - med
# std = user_depth.std()
# depth_b = pp.cut_body(dn_, center, pheight, hand,new_shape)
# # user_b = pp.cut_body(un.copy(), center, pheight, hand,new_shape)
# depth_h = pp.cut_hand(dn_, traj2D, hand,new_shape)
# # user_h = pp.cut_hand(un.copy(), traj2D, hand,new_shape)
# new_d = empty((2,step,128,128), dtype="uint8")
# for i,d in enumerate((depth_b,depth_h)):
# d = 255 - d
# # nuser_d = d[us==0]
# # nuser_d[nuser_d>thresh_noise] = 0
# # d[us==0] = nuser_d
# d[d>thresh_noise] = 0
# #d = inpaint(d, thresh_noise)
# thresh_d = med-3*std
# # d[d<thresh_d] = thresh_d
# d = d-thresh_d
# d = clip(d*scaler, 0, 255)
# d = d.astype("uint8")
# d = pp.medianblur(d)
# new_d[i] = d
# dn = new_d.astype("uint8")
# depth[:,:-step]=depth[:,step:]
# depth[:,-step:] = dn.copy()
# # ---------------------------- DEPTH ------------------------------------------
# # ---------------------------- GRAY ------------------------------------------
# gray_b = pp.cut_body(gn.copy(), center, pheight, hand, new_shape)
# gray_h = pp.cut_hand(gn.copy(), traj2D, hand,new_shape)
# new_gray = empty((2,step,128,128), dtype="uint8")
# new_gray[0] = gray_b
# new_gray[1] = gray_h
# gn = new_gray.astype("uint8")
# gray[:,:-step]=gray[:,step:]
# gray[:,-step:] = gn.copy()
# # ---------------------------- GRAY ------------------------------------------
video = empty((1,2,)+gray.shape,dtype="uint8")
video[0,0] = gray.copy()
video[0,1] = depth.copy()
video = video.swapaxes(1,2) #(body-hand,gray-depth,fr,h,w)
# if hand == "left":
# for i in range(video.shape[1]):
# for j in range(video.shape[2]):
# for k in range(video.shape[3]):
# video[0,i,j,k] = cv2.flip(video[0,i,j,k],1)
# video = video[0:1].astype("uint8")
# ---------------------------AUGMENT--------------------------------------
v_new = empty((1,2,2)+vid_shape,dtype="uint8")
# p = pheight
ratio = 0.25
for i in xrange(video.shape[0]): #batch
if pheight < 10: pheight = 100
scale = ratio+randi(2)/100.
ofs = pheight*scale
mid = video.shape[-1]/2.
sli = None
if ofs < mid:
start = int(round(mid-ofs))
end = int(round(mid+ofs))
sli = slice(start,end)
for j in xrange(video.shape[2]): #maps
for k in xrange(video.shape[3]): #frames
#body
img = video[i,0,j,k]
img = cut_img(img,5)
img = misc.imresize(img,(h,h))
# if j==0: img = 255-misc.imfilter(img,"contour")
v_new[i,0,j,k] = img
#hand
img = video[i,1,j,k]
img = img[sli,sli]
img = misc.imresize(img,(h,h))
v_new[i,1,j,k] = img
# ---------------------------AUGMENT--------------------------------------
"""
x_.set_value(v_new.astype("float32"),borrow=True)
print "predicting"
pred = evalu_model()[0][0]
pred_p = pred.max()
pred_idx = pred.argmax()+1
fps = int(1./((time.time()-time_start)/step))
fr_start = cur_fr+step-n_f
fr_end = cur_fr+step
print pred_idx,"\t", "\t",pred_p,"\t",fr_start, "-",fr_end,"\t",fps,'fps'
predict(pred_idx,pred_p,fr_start)
# print v_new.shape
# for i in xrange(v_new.shape[0]):
# for j in xrange(v_new.shape[1]):
# for k in xrange(v_new.shape[2]):
# play_vid(v_new[i,j,k],wait=0)
"""
# print "put"
q.put((v_new,cur_fr,n, fileName))
cur_fr += step
def eval():
files = glob(data + "/" + "*.zip")
files.sort()
q2.put("Id,Sequence")
print len(files), "found"
for fileName in files[:]:
num = fileName.split('/')
# pred_file_name = pred_file_name[-1].replace(".zip","_prediction.csv")
num = num[-1].replace(".zip", "")[-4:]
q2.put("\n" + str(num))
print fileName
smp = pp.GestureSample(fileName)
s_time = time.time()
n = smp.data['numFrames']
dv, uv, gv = smp.depth, smp.user, smp.rgb
cur_fr = 1
new_shape = (step, 128, 128)
s = []
d, u, g = [empty((n_f, ) + vid_res + (3, ), "uint8") for _ in range(3)]
for v in dv, uv, gv:
pp.go_to_frame(v, cur_fr)
for i, fr in enumerate(range(cur_fr, cur_fr + n_f)):
s.append(smp.getSkeleton(fr))
d[i], u[i], g[i] = [v.read()[1] for v in dv, uv, gv]
d, u, g = [pp.to_grayscale(v) for v in d, u, g]
u[u < 128], u[u >= 128] = 0, 1
depth, user, gray, skelet = d, u, g, s
user_o = user.copy()
depth_o = depth.copy()
gray_o = gray.copy()
user_depth = depth_o[user_o == 1]
skelet, c = pp.proc_skelet(array(skelet).copy())
user = pp.proc_user(user)
_, depth, c = pp.proc_depth(depth.copy(), user.copy(), user_o,
array(skelet).copy())
gray, c = pp.proc_gray(
gray.copy(), user,
array(skelet).copy()) #user.copy!!!!!!!!!!!!!!!!!!!
cur_fr += n_f
while cur_fr + step < n:
sn = []
dn, un, gn = [
empty((step, ) + vid_res + (3, ), "uint8") for _ in range(3)
]
for i, fr in enumerate(range(cur_fr, cur_fr + step)):
sn.append(smp.getSkeleton(fr))
dn[i], un[i], gn[i] = [v.read()[1] for v in dv, uv, gv]
dn, un, gn = [pp.to_grayscale(v) for v in dn, un, gn]
un[un < 128], un[un >= 128] = 0, 1
s = s[step:] + sn
skelet, c = pp.proc_skelet(s, _3D=False)
len_dump = len(depth_o[:step][user_o[:step] == 1])
un_d = dn[un == 1]
user_o[:-step] = user_o[step:]
user_o[-step:] = un.copy()
un = pp.proc_user(un, 3)
user[:-step] = user[step:]
user[-step:] = un.copy()
depth_o[:-step] = depth_o[step:]
depth_o[-step:] = dn.copy()
gray_o[:-step] = gray_o[step:]
gray_o[-step:] = gn.copy()
_, depth, c = pp.proc_depth(depth_o.copy(), user.copy(), user_o,
skelet)
gray, c = pp.proc_gray(gray_o.copy(), user, skelet)
traj2D, traj3D, ori, pheight, hand, center = skelet
video = empty((
1,
2,
) + gray.shape, dtype="uint8")
video[0, 0] = gray.copy()
video[0, 1] = depth.copy()
video = video.swapaxes(1, 2) #(body-hand,gray-depth,fr,h,w)
# ---------------------------AUGMENT--------------------------------------
v_new = empty((1, 2, 2) + vid_shape, dtype="uint8")
# p = pheight
ratio = 0.25
for i in xrange(video.shape[0]): #batch
if pheight < 10: pheight = 100
scale = ratio + randi(2) / 100.
ofs = pheight * scale
mid = video.shape[-1] / 2.
sli = None
if ofs < mid:
start = int(round(mid - ofs))
end = int(round(mid + ofs))
sli = slice(start, end)
for j in xrange(video.shape[2]): #maps
for k in xrange(video.shape[3]): #frames
#body
img = video[i, 0, j, k]
img = cut_img(img, 5)
img = misc.imresize(img, (h, h))
v_new[i, 0, j, k] = img
#hand
img = video[i, 1, j, k]
img = img[sli, sli]
img = misc.imresize(img, (h, h))
v_new[i, 1, j, k] = img
# ---------------------------AUGMENT--------------------------------------
q.put((v_new, cur_fr, n, fileName))
cur_fr += step
def eval():
files = glob(data + "/" + "*.zip")
files.sort()
q2.put("Id,Sequence")
print len(files), "found"
for fileName in files[:]:
num = fileName.split("/")
# pred_file_name = pred_file_name[-1].replace(".zip","_prediction.csv")
num = num[-1].replace(".zip", "")[-4:]
q2.put("\n" + str(num))
print fileName
smp = pp.GestureSample(fileName)
s_time = time.time()
n = smp.data["numFrames"]
dv, uv, gv = smp.depth, smp.user, smp.rgb
cur_fr = 1
new_shape = (step, 128, 128)
s = []
d, u, g = [empty((n_f,) + vid_res + (3,), "uint8") for _ in range(3)]
for v in dv, uv, gv:
pp.go_to_frame(v, cur_fr)
for i, fr in enumerate(range(cur_fr, cur_fr + n_f)):
s.append(smp.getSkeleton(fr))
d[i], u[i], g[i] = [v.read()[1] for v in dv, uv, gv]
d, u, g = [pp.to_grayscale(v) for v in d, u, g]
u[u < 128], u[u >= 128] = 0, 1
depth, user, gray, skelet = d, u, g, s
user_o = user.copy()
depth_o = depth.copy()
gray_o = gray.copy()
user_depth = depth_o[user_o == 1]
skelet, c = pp.proc_skelet(array(skelet).copy())
user = pp.proc_user(user)
_, depth, c = pp.proc_depth(depth.copy(), user.copy(), user_o, array(skelet).copy())
gray, c = pp.proc_gray(gray.copy(), user, array(skelet).copy()) # user.copy!!!!!!!!!!!!!!!!!!!
cur_fr += n_f
while cur_fr + step < n:
sn = []
dn, un, gn = [empty((step,) + vid_res + (3,), "uint8") for _ in range(3)]
for i, fr in enumerate(range(cur_fr, cur_fr + step)):
sn.append(smp.getSkeleton(fr))
dn[i], un[i], gn[i] = [v.read()[1] for v in dv, uv, gv]
dn, un, gn = [pp.to_grayscale(v) for v in dn, un, gn]
un[un < 128], un[un >= 128] = 0, 1
s = s[step:] + sn
skelet, c = pp.proc_skelet(s, _3D=False)
len_dump = len(depth_o[:step][user_o[:step] == 1])
un_d = dn[un == 1]
user_o[:-step] = user_o[step:]
user_o[-step:] = un.copy()
un = pp.proc_user(un, 3)
user[:-step] = user[step:]
user[-step:] = un.copy()
depth_o[:-step] = depth_o[step:]
depth_o[-step:] = dn.copy()
gray_o[:-step] = gray_o[step:]
gray_o[-step:] = gn.copy()
_, depth, c = pp.proc_depth(depth_o.copy(), user.copy(), user_o, skelet)
gray, c = pp.proc_gray(gray_o.copy(), user, skelet)
traj2D, traj3D, ori, pheight, hand, center = skelet
video = empty((1, 2) + gray.shape, dtype="uint8")
video[0, 0] = gray.copy()
video[0, 1] = depth.copy()
video = video.swapaxes(1, 2) # (body-hand,gray-depth,fr,h,w)
# ---------------------------AUGMENT--------------------------------------
v_new = empty((1, 2, 2) + vid_shape, dtype="uint8")
# p = pheight
ratio = 0.25
for i in xrange(video.shape[0]): # batch
if pheight < 10:
pheight = 100
scale = ratio + randi(2) / 100.0
ofs = pheight * scale
mid = video.shape[-1] / 2.0
sli = None
if ofs < mid:
start = int(round(mid - ofs))
end = int(round(mid + ofs))
sli = slice(start, end)
for j in xrange(video.shape[2]): # maps
for k in xrange(video.shape[3]): # frames
# body
img = video[i, 0, j, k]
img = cut_img(img, 5)
img = misc.imresize(img, (h, h))
v_new[i, 0, j, k] = img
# hand
img = video[i, 1, j, k]
img = img[sli, sli]
img = misc.imresize(img, (h, h))
v_new[i, 1, j, k] = img
# ---------------------------AUGMENT--------------------------------------
q.put((v_new, cur_fr, n, fileName))
cur_fr += step