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,
if trajconv.append:
traj_ = T.concatenate([t.flatten(2), t_conv.flatten(2)], axis=1)
else:
traj_ = t_conv.flatten(2)
n_in_MLP -= traj_size
n_in_MLP += conv_length
elif use.traj:
traj_ = var_norm(t.flatten(2), axis=1)
# elif use.traj: traj_ = t.flatten(2)
# insp = T.stack(T.min(vid_), T.mean(vid_), T.max(vid_), T.std(vid_), batch.micro*n_in_MLP - vid_.nonzero()[0].size)#, T.min(traj_), T.mean(traj_), T.max(traj_), T.std(traj_))
# dropout
if use.drop:
if use.traj: traj_ = DropoutLayer(traj_, rng=rng, p=drop.p_traj).output
vid_ = DropoutLayer(vid_, rng=rng, p=drop.p_vid).output
def lin(X):
return X
def maxout(X, X_shape):
shape = X_shape[:-1] + (X_shape[-1] / 4, ) + (4, )
out = X.reshape(shape)
return T.max(out, axis=-1)
#maxout
if use.maxout:
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
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, res_dir, load_path):
self.layers = [] # only contain the layers from fusion
self.insp_mean = [] # inspection for each layer mean activation
self.insp_std = [] # inspection for each layer std activation
self.params = [] # parameter list
self.idx_mini = T.lscalar(name="idx_mini") # minibatch index
self.idx_micro = T.lscalar(name="idx_micro") # microbatch index
# symbolic variables
self.x = ndtensor(len(tr.in_shape))(name='x') # video input
self.y = T.ivector(name='y') # labels
# symbolic variables
self.x_skeleton = ndtensor(len(tr._skeleon_in_shape))(
name='x_skeleton') # video input
if use.drop:
drop.p_vid = shared(float32(drop.p_vid_val))
drop.p_hidden = shared(float32(drop.p_hidden_val))
video_cnn = conv3d_chalearn(self.x, use, lr, batch, net, reg, drop, mom, \
tr, res_dir, load_path)
dbn = GRBM_DBN(numpy_rng=random.RandomState(123), n_ins=891, \
hidden_layers_sizes=[2000, 2000, 1000], n_outs=101, input_x=self.x_skeleton, label=self.y )
# we load the pretrained DBN skeleton parameteres here
if use.load == True:
dbn.load(os.path.join(load_path, 'dbn_2015-06-19-11-34-24.npy'))
#####################################################################
# fuse the ConvNet output with skeleton output -- need to change here
######################################################################
out = T.concatenate([video_cnn.out, dbn.sigmoid_layers[-1].output],
axis=1)
#####################################################################
# wudi add the mean and standard deviation of the activation values to exam the neural net
# Reference: Understanding the difficulty of training deep feedforward neural networks, Xavier Glorot, Yoshua Bengio
#####################################################################
insp_mean_list = []
insp_std_list = []
insp_mean_list.extend(dbn.out_mean)
insp_mean_list.extend(video_cnn.insp_mean)
insp_std_list.extend(dbn.out_std)
insp_std_list.extend(video_cnn.insp_std)
######################################################################
#MLP layer
self.layers.append(
HiddenLayer(out,
n_in=net.hidden,
n_out=net.hidden,
rng=tr.rng,
W_scale=net.W_scale[-1],
b_scale=net.b_scale[-1],
activation=net.activation))
out = self.layers[-1].output
if tr.inspect:
insp_mean_list.extend([T.mean(out)])
insp_std_list.extend([T.std(out)])
self.insp_mean = T.stacklists(insp_mean_list)
self.insp_std = T.stacklists(insp_std_list)
if use.drop:
out = DropoutLayer(out, rng=tr.rng, p=drop.p_hidden).output
######################################################################
# softmax layer
self.layers.append(
LogRegr(out,
rng=tr.rng,
n_in=net.hidden,
W_scale=net.W_scale[-1],
b_scale=net.b_scale[-1],
n_out=net.n_class))
self.p_y_given_x = self.layers[-1].p_y_given_x
######################################################################
# cost function
self.cost = self.layers[-1].negative_log_likelihood(self.y)
# function computing the number of errors
self.errors = self.layers[-1].errors(self.y)
# parameter list
for layer in video_cnn.layers:
self.params.extend(layer.params)
# pre-trained dbn parameter last layer (W, b) doesn't need to incorporate into the params
# for calculating the gradient
self.params.extend(dbn.params[:-2])
# MLP hidden layer params
self.params.extend(self.layers[-2].params)
# softmax layer params
self.params.extend(self.layers[-1].params)
# number of inputs for MLP = (# maps last stage)*(# convnets)*(resulting video shape) + trajectory size
print 'MLP:', video_cnn.n_in_MLP, "->", net.hidden_penultimate, "+", net.hidden_traj, '->', \
net.hidden, '->', net.hidden, '->', net.n_class, ""
return