# load the skeleton normalisation --Lio didn't normalise video input, but should we?
import cPickle
f = open('CNN_normalization.pkl','rb')
CNN_normalization = cPickle.load(f)
Mean_CNN = CNN_normalization ['Mean_CNN']
Std_CNN = CNN_normalization['Std_CNN']
# customized data loader for both video module and skeleton module
loader = DataLoader_with_skeleton_normalisation(src, tr.batch_size, Mean_CNN, Std_CNN) # Lio changed it to read from HDF5 files
####################################################################
# 3DCNN for video module
####################################################################
# we load the CNN parameteres here
video_cnn = conv3d_chalearn(x, use, lr, batch, net, reg, drop, mom, tr, res_dir)
#####################################################################
# fuse the ConvNet output with skeleton output -- need to change here
######################################################################
out = video_cnn.out
# some activation inspection
insp = []
for insp_temp in video_cnn.insp: insp.append(insp_temp)
insp = T.stack(insp)
# softmax layer
layers.append(LogRegr(out, rng=tr.rng, n_in=net.hidden_vid,
W_scale=net.W_scale[-1], b_scale=net.b_scale[-1], n_out=net.n_class))
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
Std_CNN = CNN_normalization['Std_CNN']
# customized data loader for both video module and skeleton module
loader = DataLoader_with_skeleton_normalisation(
src, tr.batch_size, Mean_CNN,
Std_CNN) # Lio changed it to read from HDF5 files
####################################################################
# 3DCNN for video module
####################################################################
#if u have already trained the network before, the might pause for some reason, u can load the updated paramters
# we load the CNN parameteres here
#use.load = True
#load_path = '/home/zhiquan/fancy/meterials/chalearn2014_fancy_data/result_temp/3dcnn/try/70.9% 2018.05.05.02.28.21/'
#video_cnn = conv3d_chalearn(x, use, lr, batch, net, reg, drop, mom, tr, res_dir,load_path)
video_cnn = conv3d_chalearn(x, use, lr, batch, net, reg, drop, mom, tr,
res_dir)
#####################################################################
# fuse the ConvNet output with skeleton output -- need to change here
######################################################################
out = video_cnn.out
# some activation inspection
insp = []
for insp_temp in video_cnn.insp_mean:
insp.append(insp_temp)
insp = T.stack(insp)
# softmax layer
layers.append(
LogRegr(out,
rng=tr.rng,
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
