print 'MLP:', n_in_MLP, "->", net.hidden, "->", net.n_class, ""
if use.depth:
if net.n_convnets == 2:
out = [x[:, :, 0, :, :, :], x[:, :,
1, :, :, :]] # 2 nets: body and hand
# build 3D ConvNet
layers = [] # all architecture layers
insp = []
for stage in xrange(net.n_stages):
for i in xrange(len(out)): # for body and hand
# normalization
if use.norm and stage == 0:
gray_norm = NormLayer(out[i][:, 0:1],
method="lcn",
use_divisor=use.norm_div).output
gray_norm = std_norm(gray_norm, axis=[-3, -2, -1])
depth_norm = var_norm(out[i][:, 1:])
out[i] = T.concatenate([gray_norm, depth_norm], axis=1)
elif use.norm:
out[i] = NormLayer(out[i], method="lcn",
use_divisor=use.norm_div).output
out[i] = std_norm(out[i], axis=[-3, -2, -1])
# convolutions
out[i] *= net.scaler[stage][i]
layers.append(
ConvLayer(
out[i],
**conv_args(stage, i, batch, net, use, tr.rng,
tr.video_shapes)))
def build():
# constants
floatX = config.floatX
enum = enumerate
file = gzip.GzipFile("params.zip", 'rb')
params = load(file)
file.close()
print params
W = [[None, None], [None, None], [None, None]]
b = [[None, None], [None, None], [None, None]]
W[0][0], b[0][0], W[0][1], b[0][1], W[1][0], b[1][0], W[1][1], b[1][1], W[
2][0], b[2][0], W[2][1], b[2][1], Wh, bh, Ws, bs = params
#-----------------------------FLIP KERNEL------------------------------------------
W = array(W)
W_new = [[None, None], [None, None], [None, None]]
for i in range(W.shape[0]):
for j in range(W.shape[1]):
w = W[i, j].get_value()
print w.shape, w.dtype
for k in range(w.shape[0]):
for l in range(w.shape[1]):
for m in range(w.shape[2]):
w[k, l, m] = cv2.flip(w[k, l, m], -1)
W_new[i][j] = shared(array(w, dtype=floatX), borrow=True)
W = W_new
#-----------------------------FLIP KERNEL------------------------------------------
rng = random.RandomState(
1337) # this will make sure results are always the same
batch_size = 1
in_shape = (1, 2, 2, 32, 64, 64
) # (batchsize, maps, frames, w, h) input video shapes
traj_shape = (batch_size, 3, 32
) # (batchsize, input shape of the trajectory
# hyper parameters
# ------------------------------------------------------------------------------
# use techniques/methods
class use:
drop = True # dropout
depth = True # use depth map as input
aug = False # data augmentation
load = False # load params.p file
traj = False # trajectory
trajconv = False # convolutions on trajectory
valid2 = False
fast_conv = False
norm_div = False
norm = True # normalization layer
mom = True # momentum
# regularization
class reg:
L1_traj = .0 # degree/amount of regularization
L2_traj = .0 # 1: only L1, 0: only L2
L1_vid = .0 # degree/amount of regularization
L2_vid = .0 # 1: only L1, 0: only L2
class trajconv:
append = False # append convolutions result to original traject
filter_size = 5
layers = 3 # number of convolution layers
res_shape = traj_shape[-1] - layers * (filter_size - 1)
class net:
shared_stages = [] # stages where weights are shared
shared_convnets = [
] # convnets that share weights ith beighbouring convnet
n_convnets = 2 # number of convolutional networks in the architecture
maps = [2, 16, 32, 64] # feature maps in each convolutional network
# maps = [2,5,25,25] # feature maps in each convolutional network
kernels = [(1, 7, 7), (1, 8, 8),
(1, 6, 6)] # convolution kernel shapes
pools = [(2, 2, 2), (2, 2, 2), (2, 2, 2)] # pool/subsampling shapes
hidden_traj = 200 # hidden units in MLP
hidden_vid = 300 # hidden units in MLP
W_scale = 0.01
b_scale = 0.1
norm_method = "lcn" # normalisation method: lcn = local contrast normalisation
pool_method = "max" # maxpool
fusion = "early" # early or late fusion
hidden = hidden_traj + hidden_vid if fusion == "late" else 500 # hidden units in MLP
n_class = 21
activation = relu
n_stages = len(net.kernels)
video_shapes = [in_shape[-3:]]
def _shared(val, borrow=True):
return shared(array(val, dtype=floatX), borrow=borrow)
def ndtensor(n):
return TensorType(floatX, (False, ) * n) # n-dimensional tensor
for i in xrange(n_stages):
k, p, v = array(net.kernels[i]), array(net.pools[i]), array(
video_shapes[i])
conv_s = tuple(v - k + 1)
video_shapes.append(tuple((v - k + 1) / p))
n_in_MLP = net.maps[-1] * net.n_convnets * prod(video_shapes[-1])
def conv_args(stage, i):
""" ConvLayer arguments, i: stage index """
args = {
'batch_size': 1,
'activation': activation,
'rng': rng,
'n_in_maps': net.maps[stage],
'n_out_maps': net.maps[stage + 1],
'kernel_shape': net.kernels[stage],
'video_shape': video_shapes[stage],
"fast_conv": use.fast_conv,
"layer_name": "Conv" + str(stage),
"W_scale": net.W_scale,
"b_scale": net.b_scale,
"stride": 1,
"W": W[stage][i],
"b": b[stage][i]
}
return args
# print conv_args(0,0)
x = ndtensor(len(in_shape))(name='x') # video input
def var_norm(_x, imgs=True, axis=[-3, -2, -1]):
if imgs:
return (_x - T.mean(_x, axis=axis, keepdims=True)) / T.maximum(
1e-4, T.std(_x, axis=axis, keepdims=True))
return (_x - T.mean(_x)) / T.maximum(1e-4, T.std(_x))
def std_norm(_x, axis=[-3, -2, -1]):
return _x / T.maximum(1e-4, T.std(_x, axis=axis, keepdims=True))
out = [x[:, 0], x[:, 1]]
for stage in xrange(n_stages):
for i in xrange(len(out)): # for each convnet of the stage
if stage == 0:
gray_norm = NormLayer(out[i][:, 0:1],
method="lcn",
use_divisor=False).output
gray_norm = std_norm(gray_norm)
depth_norm = var_norm(out[i][:, 1:])
out[i] = T.concatenate([gray_norm, depth_norm], axis=1)
else:
out[i] = NormLayer(out[i], method="lcn",
use_divisor=False).output
out[i] = std_norm(out[i])
out[i] = ConvLayer(out[i], **conv_args(stage, i)).output
out[i] = PoolLayer(out[i],
net.pools[stage],
method=net.pool_method).output
out = [out[i].flatten(2) for i in range(len(out))]
out = T.concatenate(out, axis=1)
#hidden layer
out = HiddenLayer(out,
W=Wh,
b=bh,
n_in=n_in_MLP,
n_out=net.hidden,
rng=rng,
activation=activation).output
logreg = LogRegr(out,
W=Ws,
b=bs,
rng=rng,
activation=activation,
n_in=net.hidden,
n_out=net.n_class)
pred = logreg.p_y_given_x
x_ = _shared(empty(in_shape))
print "compiling..."
eval_model = function([], [pred], givens={x: x_}, on_unused_input='ignore')
print "compiling done"
return eval_model, x_
for stage in xrange(n_stages):
for i in xrange(len(out)): # for each convnet of the stage
# normalization
# if use.norm and stage==0:
# gray_norm = NormLayer(out[i][:,0:1], method="lcn",
# use_divisor=use.norm_div).output
# gray_norm = std_norm(gray_norm,axis=[-3,-2,-1])
# depth_norm = var_norm(out[i][:,1:])
# out[i] = T.concatenate([gray_norm,depth_norm],axis=1)
# elif use.norm:
# out[i] = NormLayer(out[i], method="lcn",use_divisor=use.norm_div).output
# # out[i] = std_norm(out[i],axis=[-3,-2,-1])
# else:
if stage == 0:
gray_norm = NormLayer(out[i][:, 0:1],
method="lcn",
kernel_size=9,
use_divisor=use.norm_div).output
# gray_norm = std_norm(gray_norm,axis=[-3,-2,-1])
depth_norm = var_norm(out[i][:, 1:2])
# user_norm = var_norm(out[i][:,2:3])
out[i] = T.concatenate([gray_norm, depth_norm], axis=1)
# out[i] = T.concatenate([gray_norm,depth_norm, user_norm],axis=1)
# out[i] = mean_sub(out[i],axis=[-3,-2,-1])
# out[i] = std_norm(out[i],axis=[-4,-3,-2,-1])
# out[i] = var_norm(out[i],axis=[-3,-2,-1])
elif stage == 1:
out[i] = NormLayer(out[i],
method="lcn",
else:
if net.n_convnets == 1: out = [x[:, 0, 0:1]]
else:
out = [x[:, 0, 0:1], x[:, 1,
0:1]] # 2 nets without depth: left and right
# build 3D ConvNet
for stage in xrange(n_stages):
for i in xrange(len(out)): # for each convnet of the stage
# convolutions
layers.append(ConvLayer(out[i], **conv_args(stage, i)))
out[i] = layers[-1].output
# normalization
if use.norm:
out[i] = NormLayer(out[i],
method=net.norm_method,
use_divisor=use.norm_div).output
# pooling, subsamping
out[i] = PoolLayer(out[i], net.pools[stage],
method=net.pool_method).output
# out[i] = normalize(out[i])
# flatten all convnets outputs
out = [out[i].flatten(2) for i in range(len(out))]
vid_ = T.concatenate(out, axis=1)
debug = vid_
# traject convolution
# ------------------------------------------------------------------------------
def build():
use.load = True # we load the CNN parameteres here
x = ndtensor(len(tr.in_shape))(name='x') # video input
x_ = _shared(empty(tr.in_shape))
conv_shapes = []
for i in xrange(net.n_stages):
k, p, v = array(net.kernels[i]), array(net.pools[i]), array(
tr.video_shapes[i])
conv_s = tuple(v - k + 1)
conv_shapes.append(conv_s)
tr.video_shapes.append(tuple((v - k + 1) / p))
print "stage", i
print " conv", tr.video_shapes[i], "->", conv_s
print " pool", conv_s, "->", tr.video_shapes[i + 1], "x", net.maps[i +
1]
# number of inputs for MLP = (# maps last stage)*(# convnets)*(resulting video shape) + trajectory size
n_in_MLP = net.maps[-1] * net.n_convnets * prod(tr.video_shapes[-1])
print 'MLP:', n_in_MLP, "->", net.hidden, "->", net.n_class, ""
if use.depth:
if net.n_convnets == 2:
out = [x[:, :, 0, :, :, :], x[:, :,
1, :, :, :]] # 2 nets: body and hand
# build 3D ConvNet
layers = [] # all architecture layers
insp = []
for stage in xrange(net.n_stages):
for i in xrange(len(out)): # for body and hand
# normalization
if use.norm and stage == 0:
gray_norm = NormLayer(out[i][:, 0:1],
method="lcn",
use_divisor=use.norm_div).output
gray_norm = std_norm(gray_norm, axis=[-3, -2, -1])
depth_norm = var_norm(out[i][:, 1:])
out[i] = T.concatenate([gray_norm, depth_norm], axis=1)
elif use.norm:
out[i] = NormLayer(out[i],
method="lcn",
use_divisor=use.norm_div).output
out[i] = std_norm(out[i], axis=[-3, -2, -1])
# convolutions
out[i] *= net.scaler[stage][i]
layers.append(
ConvLayer(
out[i],
**conv_args(stage, i, batch, net, use, tr.rng,
tr.video_shapes)))
out[i] = layers[-1].output
out[i] = PoolLayer(out[i],
net.pools[stage],
method=net.pool_method).output
if tr.inspect: insp.append(T.mean(out[i]))
# flatten all convnets outputs
for i in xrange(len(out)):
out[i] = std_norm(out[i], axis=[-3, -2, -1])
out = [out[i].flatten(2) for i in range(len(out))]
vid_ = T.concatenate(out, axis=1)
# dropout
if use.drop:
drop.p_vid = shared(float32(drop.p_vid_val))
drop.p_hidden = shared(float32(drop.p_hidden_val))
drop.p_vid.set_value(float32(0.)) # dont use dropout when testing
drop.p_hidden.set_value(float32(0.)) # dont use dropout when testing
vid_ = DropoutLayer(vid_, rng=tr.rng, p=drop.p_vid).output
# MLP
# ------------------------------------------------------------------------------
# fusion
if net.fusion == "early":
out = vid_
# hidden layer
Wh, bh = load_params(use) # This is test, wudi added this!
layers.append(
HiddenLayer(out,
W=Wh,
b=bh,
n_in=n_in_MLP,
n_out=net.hidden,
rng=tr.rng,
W_scale=net.W_scale[-2],
b_scale=net.b_scale[-2],
activation=relu))
out = layers[-1].output
if tr.inspect:
insp = T.stack(insp[0], insp[1], insp[2], insp[3], insp[4], insp[5],
T.mean(out))
else:
insp = T.stack(0, 0)
if use.drop: out = DropoutLayer(out, rng=tr.rng, p=drop.p_hidden).output
#maxout
# softmax layer
Ws, bs = load_params(use) # This is test, wudi added this!
layers.append(
LogRegr(out,
W=Ws,
b=bs,
rng=tr.rng,
activation=lin,
n_in=net.hidden,
W_scale=net.W_scale[-1],
b_scale=net.b_scale[-1],
n_out=net.n_class))
"""
layers[-1] : softmax layer
layers[-2] : hidden layer (video if late fusion)
layers[-3] : hidden layer (trajectory, only if late fusion)
"""
# prediction
y_pred = layers[-1].y_pred
p_y_given_x = layers[-1].p_y_given_x
####################################################################
####################################################################
print "\n%s\n\tcompiling\n%s" % (('-' * 30, ) * 2)
####################################################################
####################################################################
# compile functions
# ------------------------------------------------------------------------------
print 'compiling test_model'
eval_model = function([], [y_pred, p_y_given_x],
givens={x: x_},
on_unused_input='ignore')
return eval_model, x_
def __init__(self,
x,
use,
lr,
batch,
net,
reg,
drop,
mom,
tr,
res_dir,
load_path=""):
self.out = []
self.layers = []
self.insp_mean = []
self.insp_std = []
for c in (use, lr, batch, net, reg, drop, mom, tr):
write(c.__name__ + ":", res_dir)
_s = c.__dict__
del _s['__module__'], _s['__doc__']
for key in _s.keys():
val = str(_s[key])
if val.startswith("<static"):
val = str(_s[key].__func__.__name__)
if val.startswith("<Cuda"): continue
if val.startswith("<Tensor"): continue
write(" " + key + ": " + val, res_dir)
####################################################################
####################################################################
print "\n%s\n\tbuilding\n%s" % (('-' * 30, ) * 2)
####################################################################
####################################################################
# ConvNet
# ------------------------------------------------------------------------------
# calculate resulting video shapes for all stages
print net.n_stages
conv_shapes = []
for i in xrange(net.n_stages):
k, p, v = array(net.kernels[i]), array(net.pools[i]), array(
tr.video_shapes[i])
conv_s = tuple(v - k + 1)
conv_shapes.append(conv_s)
tr.video_shapes.append(tuple((v - k + 1) / p))
print "stage", i
if use.depth and i == 0:
print " conv", tr.video_shapes[
i], "x 2 ->", conv_s #for body and hand
else:
print " conv", tr.video_shapes[i], "->", conv_s
print " pool", conv_s, "->", tr.video_shapes[i +
1], "x", net.maps[i +
1]
# number of inputs for MLP = (# maps last stage)*(# convnets)*(resulting video shape) + trajectory size
n_in_MLP = net.maps[-1] * net.n_convnets * prod(tr.video_shapes[-1])
print 'debug1'
if use.depth:
if net.n_convnets == 2:
out = [x[:, :, 0, :, :, :],
x[:, :, 1, :, :, :]] # 2 nets: body and hand
# build 3D ConvNet
for stage in xrange(net.n_stages):
for i in xrange(len(out)): # for body and hand
# normalization
if use.norm and stage == 0:
gray_norm = NormLayer(out[i][:, 0:1],
method="lcn",
use_divisor=use.norm_div).output
gray_norm = std_norm(gray_norm, axis=[-3, -2, -1])
depth_norm = var_norm(out[i][:, 1:])
out[i] = T.concatenate([gray_norm, depth_norm], axis=1)
elif use.norm:
out[i] = NormLayer(out[i],
method="lcn",
use_divisor=use.norm_div).output
out[i] = std_norm(out[i], axis=[-3, -2, -1])
# convolutions
out[i] *= net.scaler[stage][i]
print 'debug2'
self.layers.append(
ConvLayer(
out[i],
**conv_args(stage, i, batch, net, use, tr.rng,
tr.video_shapes, load_path)))
out[i] = self.layers[-1].output
out[i] = PoolLayer(out[i],
net.pools[stage],
method=net.pool_method).output
if tr.inspect:
self.insp_mean.append(T.mean(out[i]))
self.insp_std.append(T.std(out[i]))
print 'debug2'
# flatten all convnets outputs
for i in xrange(len(out)):
out[i] = std_norm(out[i], axis=[-3, -2, -1])
out = [out[i].flatten(2) for i in range(len(out))]
vid_ = T.concatenate(out, axis=1)
print 'debug3'
# dropout
if use.drop:
drop.p_vid = shared(float32(drop.p_vid_val))
drop.p_hidden = shared(float32(drop.p_hidden_val))
vid_ = DropoutLayer(vid_, rng=tr.rng, p=drop.p_hidden).output
#maxout
if use.maxout:
vid_ = maxout(vid_, (batch.micro, n_in_MLP))
net.activation = lin
n_in_MLP /= 2
# net.hidden *= 2
# MLP
# ------------------------------------------------------------------------------
# fusion
if net.fusion == "early":
out = vid_
# hidden layer
if use.load:
W, b = load_params(use, load_path)
self.layers.append(
HiddenLayer(out,
n_in=n_in_MLP,
n_out=net.hidden_vid,
rng=tr.rng,
W=W,
b=b,
W_scale=net.W_scale[-2],
b_scale=net.b_scale[-2],
activation=net.activation))
else:
self.layers.append(
HiddenLayer(out,
n_in=n_in_MLP,
n_out=net.hidden_vid,
rng=tr.rng,
W_scale=net.W_scale[-2],
b_scale=net.b_scale[-2],
activation=net.activation))
out = self.layers[-1].output
#if tr.inspect:
#self.insp_mean = T.stack(self.insp_mean)
#self.insp_std = T.stack(self.insp_std)
#self.insp = T.stack(self.insp[0],self.insp[1],self.insp[2],self.insp[3],self.insp[4],self.insp[5], T.mean(out))
#else: self.insp = T.stack(0,0)
# out = normalize(out)
if use.drop:
out = DropoutLayer(out, rng=tr.rng, p=drop.p_hidden).output
#maxout
if use.maxout:
out = maxout(out, (batch.micro, net.hidden))
net.hidden /= 2
print 'debug3'
# now assembly all the output
self.out = out
self.n_in_MLP = n_in_MLP
# if net.n_convnets==2: out = [x[:,0], x[:,1]] # 2 nets: left and right
# else: out = [x[:,0,0:1], x[:,0,1:2], x[:,1,0:1], x[:,1,1:2]] # 4 nets
# else: out = [x[:,0,0:1], x[:,1,1:2]] # 2 nets without depth: left and right
out = [x[:, 0]]
# build 3D ConvNet
for stage in xrange(n_stages):
for i in xrange(len(out)): # for each convnet of the stage
# convolutions
layers.append(ConvLayer(out[i], **conv_args(stage, i)))
out[i] = layers[-1].output
# normalization
if use.norm:
out[i] = NormLayer(out[i],
method=net.norm_method,
use_divisor=False).output
# pooling, subsamping
out[i] = PoolLayer(out[i], net.pools[stage],
method=net.pool_method).output
# flatten all convnets outputs
# out = [out[i].flatten(2) for i in range(len(out))]
# vid_ = T.concatenate(out, axis=1)
vid_ = T.concatenate([out[0].flatten(2)], axis=1)
# traject convolution
# ------------------------------------------------------------------------------
if use.trajconv:
t_conv = t.reshape((prod(traj_shape[:-1]), 1, 1, traj_shape[-1]))