update = []
# shared variables
learning_rate = shared(float32(lr.init))
if use.mom:
momentum = shared(float32(mom.momentum))
drop.p_vid = shared(float32(drop.p_vid_val) )
drop.p_hidden = shared(float32(drop.p_hidden_val))
idx_mini = T.lscalar(name="idx_mini") # minibatch index
idx_micro = T.lscalar(name="idx_micro") # microbatch index
x = ndtensor(len(tr.in_shape))(name = 'x') # video input
y = T.ivector(name = 'y') # labels
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty(tr.batch_size))
y_int32 = T.cast(y_,'int32')
L1 = _shared(0)
L2 = _shared(0)
### useless fake, but DataLoader_with_skeleton_normalisation would require that
x_skeleton = ndtensor(len(tr._skeleon_in_shape))(name = 'x_skeleton') # video input
x_skeleton_ = _shared(empty(tr._skeleon_in_shape))
# load the skeleton normalisation --Lio didn't normalise video input, but should we?
import cPickle
f = open('CNN_normalization.pkl','rb')
src = "/mnt/wd/chalearn/preproc"
res_dir_ = "/home/lpigou/chalearn_wudi/try"
loader = DataLoader(src,
tr.batch_size) # Lio changed it to read from HDF5 files
####################################################################
####################################################################
print "\n%s\n\tbuilding\n%s" % (('-' * 30, ) * 2)
####################################################################
####################################################################
idx_mini = T.lscalar(name="idx_mini") # minibatch index
idx_micro = T.lscalar(name="idx_micro") # microbatch index
x = ndtensor(len(tr.in_shape))(name='x') # video input
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty((tr.batch_size, )))
y_int32 = T.cast(y_, 'int32')
y = T.ivector(name='y') # labels
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]
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_
micro_updates = []
last_upd = []
update = []
# shared variables
learning_rate = shared(float32(lr.init))
if use.mom:
momentum = shared(float32(mom.momentum))
drop.p_vid = shared(float32(drop.p_vid_val))
drop.p_hidden = shared(float32(drop.p_hidden_val))
idx_mini = T.lscalar(name="idx_mini") # minibatch index
idx_micro = T.lscalar(name="idx_micro") # microbatch index
x = ndtensor(len(tr.in_shape))(name='x') # video input
y = T.ivector(name='y') # labels
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty(tr.batch_size))
y_int32 = T.cast(y_, 'int32')
L1 = _shared(0)
L2 = _shared(0)
### useless fake, but DataLoader_with_skeleton_normalisation would require that
x_skeleton = ndtensor(len(tr._skeleon_in_shape))(
name='x_skeleton') # video input
x_skeleton_ = _shared(empty(tr._skeleon_in_shape))
# 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)
lt = localtime()
res_dir = res_dir_+"/try/"+str(lt.tm_year)+"."+str(lt.tm_mon).zfill(2)+"." \
+str(lt.tm_mday).zfill(2)+"."+str(lt.tm_hour).zfill(2)+"."\
+str(lt.tm_min).zfill(2)+"."+str(lt.tm_sec).zfill(2)
os.makedirs(res_dir)
# we need to parse an absolute path for HPC to load
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('path')
args = parser.parse_args()
load_path = args.path
######################################################################
net_convnet3d_grbm_early_fusion = convnet3d_grbm_early_fusion(res_dir, load_path)
net_convnet3d_grbm_early_fusion.load_params(os.path.join(load_path,'paramsbest.zip'))
x_ = _shared(empty(tr.in_shape))
x_skeleton_ = _shared(empty(tr._skeleon_in_shape))
p_y_given_x = net_convnet3d_grbm_early_fusion.prediction_function(x_, x_skeleton_)
#############################
# load normalisation constant given load_path
Mean_skel, Std_skel, Mean_CNN, Std_CNN = net_convnet3d_grbm_early_fusion.load_normalisation_constant(load_path)
for file_count, file in enumerate(samples):
condition = (file_count > -1)
if condition: #wudi only used first 650 for validation !!! Lio be careful!
save_path= os.path.join(data, file)
print file
time_start = time()
# we load precomputed feature set or recompute the whole feature set
if os.path.isfile(save_path):
def build_finetune_functions(self, x_, y_int32, x_skeleton_,
learning_rate):
'''
This is used to fine tune the network
'''
# compute the gradients with respect to the model parameters
gparams = T.grad(self.cost, self.params)
# compute list of fine-tuning updates
mini_updates = []
micro_updates = []
update = []
last_upd = []
# shared variables
if use.mom: momentum = shared(float32(mom.momentum))
def get_update(i):
return update[i] / (batch.mini / batch.micro)
for i, (param, gparam) in enumerate(zip(self.params, gparams)):
# shape of the parameters
shape = param.get_value(borrow=True).shape
# init updates := zeros
update.append(_shared(zeros(shape, dtype=config.floatX)))
# micro_updates: sum of lr*grad
micro_updates.append(
(update[i], update[i] + learning_rate * gparam))
# re-init updates to zeros
mini_updates.append((update[i], zeros(shape, dtype=config.floatX)))
if use.mom:
last_upd.append(_shared(zeros(shape, dtype=config.floatX)))
v = momentum * last_upd[i] - get_update(i)
mini_updates.append((last_upd[i], v))
if mom.nag: # nesterov momentum
mini_updates.append(
(param, param + momentum * v - get_update(i)))
else:
mini_updates.append((param, param + v))
else:
mini_updates.append((param, param - get_update(i)))
def get_batch(_data):
pos_mini = self.idx_mini * batch.mini
idx1 = pos_mini + self.idx_micro * batch.micro
idx2 = pos_mini + (self.idx_micro + 1) * batch.micro
return _data[idx1:idx2]
def givens(dataset_):
return {
self.x: get_batch(dataset_[0]),
self.y: get_batch(dataset_[1]),
self.x_skeleton: get_batch(dataset_[2])
}
print 'compiling apply_updates'
apply_updates = function([],
updates=mini_updates,
on_unused_input='ignore')
print 'compiling train_model'
train_model = function(
[self.idx_mini, self.idx_micro],
[self.cost, self.errors, self.insp_mean, self.insp_std],
updates=micro_updates,
givens=givens((x_, y_int32, x_skeleton_)),
on_unused_input='ignore')
print 'compiling test_model'
test_model = function(
[self.idx_mini, self.idx_micro],
[self.cost, self.errors, self.insp_mean, self.insp_std],
givens=givens((x_, y_int32, x_skeleton_)),
on_unused_input='ignore')
return apply_updates, train_model, test_model
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_
#############################
# load normalisation constant given load_path
#Mean_skel, Std_skel, Mean_CNN, Std_CNN = net_convnet3d_grbm_early_fusion.load_normalisation_constant(load_path)
f = open('/home/zhiquan/fancy/meterials/temp/SK_normalization.pkl', 'rb')
SK_normalization = cPickle.load(f)
Mean_skel = SK_normalization['Mean1']
Std_skel = SK_normalization['Std1']
####################################################################
# DBN for skeleton modules
####################################################################
# ------------------------------------------------------------------------------
# symbolic variables
x_skeleton = ndtensor(len(tr._skeleon_in_shape))(
name='x_skeleton') # video input
x_skeleton_ = _shared(empty(tr._skeleon_in_shape))
dbn = GRBM_DBN(numpy_rng=random.RandomState(123), n_ins=891, \
hidden_layers_sizes=[2000, 2000, 1000], n_outs=101, input_x=x_skeleton, label=y )
# we load the pretrained DBN skeleton parameteres here
load_path = '/idiap/user/dwu/chalearn/result/try/36.7% 2015.07.09.17.53.10'
dbn.load_params_DBN(os.path.join(load_path, 'paramsbest.zip'))
test_model = function([],
dbn.logLayer.p_y_given_x,
givens={x_skeleton: x_skeleton_},
on_unused_input='ignore')
for file_count, file in enumerate(samples):
condition = (file_count > -1)
if condition: #wudi only used first 650 for validation !!! Lio be careful!
#args = parser.parse_args()
#load_path = args.path
load_path = '/remote/idiap.svm/user.active/dwu/chalearn/result/try/CNN_normalisation_53.0% 2015.06.23.12.17.31/'
######################################################################
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
# we load the CNN parameteres here
x = ndtensor(len(tr.in_shape))(name='x') # video input
x_ = _shared(empty(tr.in_shape))
use.load = True
use.fast_conv = True
video_cnn = conv3d_chalearn(x, use, lr, batch, net, reg, drop, mom, tr,
res_dir, load_path)
out = video_cnn.out
layers = [] # all architecture layers
# softmax layer
if use.load:
W, b = load_params(use, load_path)
print W.shape, b.shape
layers.append(
LogRegr(out,
rng=tr.rng,
def build_finetune_functions(self, x_, y_int32, x_skeleton_, learning_rate):
'''
This is used to fine tune the network
'''
# compute the gradients with respect to the model parameters
gparams = T.grad(self.cost, self.params)
# compute list of fine-tuning updates
mini_updates = []
micro_updates = []
update = []
last_upd = []
# shared variables
if use.mom: momentum = shared(float32(mom.momentum))
def get_update(i): return update[i]/(batch.mini/batch.micro)
for i, (param, gparam) in enumerate(zip(self.params, gparams)):
# shape of the parameters
shape = param.get_value(borrow=True).shape
# init updates := zeros
update.append(_shared(zeros(shape, dtype=config.floatX)))
# micro_updates: sum of lr*grad
micro_updates.append((update[i], update[i] + learning_rate*gparam))
# re-init updates to zeros
mini_updates.append((update[i], zeros(shape, dtype=config.floatX)))
if use.mom:
last_upd.append(_shared(zeros(shape, dtype=config.floatX)))
v = momentum * last_upd[i] - get_update(i)
mini_updates.append((last_upd[i], v))
if mom.nag: # nesterov momentum
mini_updates.append((param, param + momentum*v - get_update(i)))
else:
mini_updates.append((param, param + v))
else:
mini_updates.append((param, param - get_update(i)))
def get_batch(_data):
pos_mini = self.idx_mini*batch.mini
idx1 = pos_mini + self.idx_micro*batch.micro
idx2 = pos_mini + (self.idx_micro+1)*batch.micro
return _data[idx1:idx2]
def givens(dataset_):
return {self.x: get_batch(dataset_[0]),
self.y: get_batch(dataset_[1]),
self.x_skeleton: get_batch(dataset_[2])}
print 'compiling apply_updates'
apply_updates = function([],
updates=mini_updates,
on_unused_input='ignore')
print 'compiling train_model'
train_model = function([self.idx_mini, self.idx_micro], [self.cost, self.errors, self.insp_mean, self.insp_std],
updates=micro_updates,
givens=givens((x_, y_int32, x_skeleton_)),
on_unused_input='ignore')
print 'compiling test_model'
test_model = function([self.idx_mini, self.idx_micro], [self.cost, self.errors, self.insp_mean, self.insp_std],
givens=givens((x_, y_int32, x_skeleton_)),
on_unused_input='ignore')
return apply_updates, train_model, test_model
#args = parser.parse_args()
#load_path = args.path
load_path='/remote/idiap.svm/user.active/dwu/chalearn/result/try/CNN_normalisation_53.0% 2015.06.23.12.17.31/'
######################################################################
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
# we load the CNN parameteres here
x = ndtensor(len(tr.in_shape))(name = 'x') # video input
x_ = _shared(empty(tr.in_shape))
use.load=True
use.fast_conv=True
video_cnn = conv3d_chalearn(x, use, lr, batch, net, reg, drop, mom, tr, res_dir, load_path)
out = video_cnn.out
layers = [] # all architecture layers
# softmax layer
if use.load:
W, b = load_params(use, load_path)
print W.shape, b.shape
layers.append(LogRegr(out, rng=tr.rng, n_in=net.hidden_vid, W=W, b=b,
W_scale=net.W_scale[-1], b_scale=net.b_scale[-1], n_out=net.n_class))
else:
elif pc=="lio":
src = "/mnt/wd/chalearn/preproc"
res_dir_ = "/home/lpigou/chalearn_wudi/try"
lt = localtime()
res_dir = res_dir_+"/try/"+str(lt.tm_year)+"."+str(lt.tm_mon).zfill(2)+"." \
+str(lt.tm_mday).zfill(2)+"."+str(lt.tm_hour).zfill(2)+"."\
+str(lt.tm_min).zfill(2)+"."+str(lt.tm_sec).zfill(2)
os.makedirs(res_dir)
######################################################################
net_convnet3d_grbm_early_fusion = convnet3d_grbm_early_fusion(res_dir, load_path)
net_convnet3d_grbm_early_fusion.load_params(os.path.join(load_path,'paramsbest.zip'))
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty(tr.batch_size))
y_int32 = T.cast(y_,'int32')
x_skeleton_ = _shared(empty(tr._skeleon_in_shape))
#############################
# load normalisation constant given load_path
Mean_skel, Std_skel, Mean_CNN, Std_CNN = net_convnet3d_grbm_early_fusion.load_normalisation_constant(load_path)
loader = DataLoader_with_skeleton_normalisation(src, tr.batch_size, \
Mean_CNN, Std_CNN, Mean_skel, Std_skel) # Lio changed it to read from HDF5 files
######################################################################
print "\n%s\n\tcompiling\n%s"%(('-'*30,)*2)
learning_rate = shared(float32(lr.init))
apply_updates, train_model, test_model = net_convnet3d_grbm_early_fusion.build_finetune_functions(x_, y_int32, x_skeleton_,learning_rate)
# shared variables
learning_rate = shared(float32(lr.init))
if use.mom:
momentum = shared(float32(mom.momentum))
drop.p_vid = shared(float32(drop.p_vid_val) )
drop.p_hidden = shared(float32(drop.p_hidden_val))
# symbolic variables
# in shape: #frames * gray/depth * body/hand * 4 maps
x = ndtensor(len(tr.in_shape))(name = 'x') # video input
# x = T.TensorVariable(CudaNdarrayType([False] * len(in_shape))) # video input
y = T.ivector(name = 'y') # labels
idx_mini = T.lscalar(name="idx_mini") # minibatch index
idx_micro = T.lscalar(name="idx_micro") # microbatch index
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty((tr.batch_size,)))
y_int32 = T.cast(y_,'int32')
# print parameters
# ------------------------------------------------------------------------------
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)
res_dir_ = "/home/lpigou/chalearn_wudi/try"
lt = localtime()
res_dir = res_dir_+"/try/"+str(lt.tm_year)+"."+str(lt.tm_mon).zfill(2)+"." \
+str(lt.tm_mday).zfill(2)+"."+str(lt.tm_hour).zfill(2)+"."\
+str(lt.tm_min).zfill(2)+"."+str(lt.tm_sec).zfill(2)
os.makedirs(res_dir)
######################################################################
net_convnet3d_grbm_early_fusion = convnet3d_grbm_early_fusion(
res_dir, load_path)
net_convnet3d_grbm_early_fusion.load_params(
os.path.join(load_path, 'paramsbest.zip'))
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty(tr.batch_size))
y_int32 = T.cast(y_, 'int32')
x_skeleton_ = _shared(empty(tr._skeleon_in_shape))
#############################
# load normalisation constant given load_path
Mean_skel, Std_skel, Mean_CNN, Std_CNN = net_convnet3d_grbm_early_fusion.load_normalisation_constant(
load_path)
loader = DataLoader_with_skeleton_normalisation(src, tr.batch_size, \
Mean_CNN, Std_CNN, Mean_skel, Std_skel) # Lio changed it to read from HDF5 files
######################################################################
print "\n%s\n\tcompiling\n%s" % (('-' * 30, ) * 2)
learning_rate = shared(float32(lr.init))
apply_updates, train_model, test_model = net_convnet3d_grbm_early_fusion.build_finetune_functions(
x_, y_int32, x_skeleton_, learning_rate)
src = "/mnt/wd/chalearn/preproc"
res_dir_ = "/home/lpigou/chalearn_wudi/try"
loader = DataLoader(src, tr.batch_size) # Lio changed it to read from HDF5 files
####################################################################
####################################################################
print "\n%s\n\tbuilding\n%s"%(('-'*30,)*2)
####################################################################
####################################################################
idx_mini = T.lscalar(name="idx_mini") # minibatch index
idx_micro = T.lscalar(name="idx_micro") # microbatch index
x = ndtensor(len(tr.in_shape))(name = 'x') # video input
x_ = _shared(empty(tr.in_shape))
y_ = _shared(empty((tr.batch_size,)))
y_int32 = T.cast(y_,'int32')
y = T.ivector(name = 'y') # labels
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
