self.get_gp = theano.function([], error_pos, allow_input_downcast=True)
self.total_params = [
W, W_in, W_out
] + self.gp_mlp.params + self.gen.params + [b_in, b, b_out]
print "Discrminator built"
def step(self, u_t, h_tm1, W, W_in, W_out, b_in, b_out):
h_t1 = T.nnet.relu(T.dot(u_t, W_in) + b_in, alpha=0.2)
h_t = T.nnet.sigmoid(h_t1 + T.dot(h_tm1, W))
y_t = T.nnet.sigmoid(T.dot(h_t, W_out) + b_out)
return h_t, y_t
if __name__ == '__main__':
vr = VR('../data/person01_boxing_d4_uncomp.avi')
video_inp = T.dmatrix("video_inp")
input = []
for i in range(0, 20):
input.append(
numpy.asarray(cv2.resize(vr.next_frame(), (80, 60)).flatten()))
print numpy.min(input), numpy.max(input)
new_matrix = (input - numpy.min(input)) / (numpy.max(input) -
numpy.min(input) * 1.0)
input = new_matrix.T
print input
disc = Discriminator(memory=1000,
time_step=20,
h_size=500,
n_hid=1000,
mlp_hid=1000,
lr=0.0002,
video_size=80 * 60,
video_inp=video_inp,
srng=srng)
count_vid = 0
disc_err = 0
fool_err = 0
patience = 10
benevolent = 2
punishement = 1.0001
threshold = 2.0
flag = 1
for l in range(0, 10000):
for f in onlyfiles:
vr = VR('../data/' + f)
count = vr.count_frame()
vr = VR('../data/' + f)
for j in range(0, count - time_step - 1):
input = []
for i in range(0, time_step):
input.append(
numpy.asarray(
cv2.resize(vr.return_frame(j + 1),
(80, 60)).flatten()))
new_matrix = (input - numpy.min(input)) / (numpy.max(input) -
numpy.min(input) * 1.0)
input = new_matrix.T
disc_err = disc.get_disc(input)
fool_err = disc.get_gp()
#print disc.train_out()
time_step = 20
eps=0.0
video_inp = T.matrix("video_inp")
srng = RandomStreams(seed=12345)
disc = Discriminator(memory=1000,time_step=time_step,h_size=500,n_hid=1000,mlp_hid=1000,lr=0.0002,video_size=80*60,video_inp=video_inp,srng=srng)
count_vid = 0
disc_err = 0
fool_err =0
patience = 10
benevolent = 2
punishement = 1.0001
threshold = 2.0
flag = 1
for l in range(0,10000):
for f in onlyfiles:
vr = VR('../data/'+f)
count = vr.count_frame()
vr = VR('../data/'+f)
for j in range(0,count-time_step-1):
input=[]
for i in range(0,time_step):
input.append(numpy.asarray(cv2.resize(vr.return_frame(j+1),(80,60)).flatten()))
new_matrix = (input - numpy.min(input))/(numpy.max(input)-numpy.min(input)*1.0)
input = new_matrix.T
disc_err = disc.get_disc(input)
fool_err = disc.get_gp()
#print disc.train_out()
if flag ==1 and disc_err <threshold:
flag = 0
if flag == 0 and fool_err <threshold:
flag = 1
self.get_gp = theano.function([],
error_pos,allow_input_downcast=True
)
self.total_params = [W,W_in,W_out]+self.gp_mlp.params+self.gen.params+[b_in,b,b_out]
print "Discrminator built"
def step(self,u_t, h_tm1, W, W_in, W_out,b_in,b_out):
h_t1 = T.nnet.relu(T.dot(u_t, W_in)+b_in,alpha=0.2)
h_t = T.nnet.sigmoid(h_t1+ T.dot(h_tm1, W))
y_t = T.nnet.sigmoid(T.dot(h_t, W_out)+b_out)
return h_t, y_t
if __name__=='__main__':
vr = VR('../data/person01_boxing_d4_uncomp.avi')
video_inp = T.dmatrix("video_inp")
input=[]
for i in range(0,20):
input.append(numpy.asarray(cv2.resize(vr.next_frame(),(80,60)).flatten()))
print numpy.min(input),numpy.max(input)
new_matrix = (input - numpy.min(input))/(numpy.max(input)-numpy.min(input)*1.0)
input = new_matrix.T
print input
disc = Discriminator(memory=1000,time_step=20,h_size=500,n_hid=1000,mlp_hid=500,lr=0.001,video_size=80*60,video_inp=video_inp)
print disc.get_auto(input).shape
print disc.train_auto(input)
print disc.train_disc(input)
print disc.train_gp()
for i in range(1,1000):
