recfield = 2 nfilter1 = 32 x = T.matrix(name='x', dtype=theano.config.floatX) y = T.matrix(name='y', dtype=theano.config.floatX) layer0 = x.reshape((bs, 3, imL, imL)) conv1 = ConvLayer(input = layer0, image_shape = (bs, 3, imL, imL), filter_shape =(nfilter1, 3, filterL, filterL), pool_shape = (recfield, recfield), flatten = True, actfun=tanh, tag='_conv1') outL = np.floor((imL-filterL+1.)/recfield).astype(np.int) fc2 = FCLayer(input=conv1.output(), n_in=nfilter1*outL*outL, n_out=imL*imL, actfun=None, tag='_fc2') params_cmb = conv1.params + fc2.params #params_cmb = fc0.params + fc2.params #ypred = fc2.output().reshape((bs, imL, imL)) ypred = fc2.output() model = GeneralModel(input=x, output=ypred, target=y, params=params_cmb, regularizers = 0, cost_func=mean_cross_entropy_map, error_func=mean_sqr_map) sgd = sgd_optimizer(data = msra, model = model, batch_size=bs, learning_rate=0.5,
filter_shape =(nfilter1, 3, filterL, filterL), pool_shape = (recfield, recfield), flatten = False, actfun=tanh, tag='_conv1') outL1 = np.floor((imL-filterL+1.)/recfield).astype(np.int) conv2 = ConvLayer(input = conv1.output(), image_shape=(bs, nfilter1, outL1, outL1), filter_shape = (nfilter2, nfilter1, filterL2, filterL2), pool_shape=(recfield, recfield), flatten=True, actfun=tanh, tag='_conv2') outL2 = np.floor((outL1-filterL2+1.)/recfield).astype(np.int) fc3 = FCLayer(input=conv2.output(), n_in=nfilter2*outL2*outL2, n_out=imL*imL, tag='_fc3') fc4 = FCLayer(input=fc3.output(), n_in=imL*imL, n_out=imL*imL, actfun=sigmoid, tag='_fc4') params_cmb = conv1.params + conv2.params + fc3.params + fc4.params #params_cmb = fc0.params + fc2.params #ypred = fc2.output().reshape((bs, imL, imL)) ypred = fc4.output() model = GeneralModel(input=x, data = msra, output=ypred, target=y, params=params_cmb, cost_func=mean_nneq_cross_entropy, error_func=mean_sqr, regularizers = 0, batch_size=bs) sgd = sgd_optimizer(data = msra,
def train_DNN(viper, lr=1e-3):
bs = 50
n_train, n_feat = viper.train_feat.shape[:2]
x = T.matrix(name='x', dtype=theano.config.floatX)
y = T.vector(name='y', dtype=theano.config.floatX)
# y = T.ivector(name='y')
# layer0 = x.reshape(viper.train_feat.shape)
if viper.tag is 'headsholder':
lr = 1e-5
fc1 = FCLayer(input=x, n_in=n_feat, n_out=256, actfun=tanh, tag='_fc1')
fc2 = FCLayer(input=fc1.output(), n_in=256, n_out=128, actfun=tanh, tag='_fc2')
fc3 = FCLayer(input=fc2.output(), n_in=128, n_out=1, actfun=sigmoid, tag='_fc3')
params_cmb = fc1.params + fc2.params + fc3.params
# ypred = fc3.output()
ypred = fc3.output().flatten()
# regularization
reg = 0
# reg = 0.001*((fc1.W**2).sum() \
# + (fc2.W**2).sum() \
# + (fc3.W**2).sum())
elif viper.tag is 'upperbody':
lr = 1e-5
fc1 = FCLayer(input=x, n_in=n_feat, n_out=4096, actfun=tanh, tag='_fc1')
fc2 = FCLayer(input=fc1.output(), n_in=4096, n_out=1024, actfun=tanh, tag='_fc2')
fc3 = FCLayer(input=fc2.output(), n_in=1024, n_out=256, actfun=tanh, tag='_fc3')
fc4 = FCLayer(input=fc3.output(), n_in=256, n_out=1, actfun=sigmoid, tag='_fc4')
params_cmb = fc1.params + fc2.params + fc3.params + fc4.params
# ypred = fc3.output()
ypred = fc4.output().flatten()
reg = 0
model = GeneralModel(input=x, output=ypred,
target=y, params=params_cmb,
regularizers=reg,
cost_func=cross_entropy,
error_func=abs_error,
batch_size=bs)
sgd = sgd_optimizer(data=viper,
model=model,
batch_size=bs,
learning_rate=lr,
n_epochs=200)
train_loss, valid_loss, test_loss = sgd.fit_viper()
pl.figure(1)
pl.plot(train_loss, color='b', label='train erorr')
pl.plot(valid_loss, color='g', label='valid error')
pl.plot(test_loss, color='r', label='test error')
pl.legend()
pl.savefig('../data_viper/dnn_train.jpg')
return model
def main():
''' pipeline for supervised salience training '''
if os.path.isdir('../data_viper/'):
datapath = '../data_viper/'
else:
datapath = '../data/'
# prepare training data for supervised salience training
datafile_viper = datapath + 'viper.pkl'
if not os.path.isfile(datafile_viper):
viper = DataMan_viper_small()
viper.make_data()
savefile(datafile_viper, viper)
viper = loadfile(datafile_viper)
viper = change_label(viper)
viper.train_feat = viper.get_pixeldata('train')
viper.valid_feat = viper.get_pixeldata('valid')
viper.test_feat = viper.get_pixeldata('test')
bs = 100
imL = 10
nfilter1 = 16
filterL = 3
x = T.tensor4(name='x', dtype=theano.config.floatX)
y = T.ivector(name='y')
# layer0 = x.reshape((bs, 3, imL, imL))
conv1 = ConvLayer(input=x, image_shape=(bs, 3, imL, imL),
filter_shape=(nfilter1, 3, filterL, filterL),
flatten=True,
actfun=tanh,
tag='_convpool1')
# outL = np.floor((imL-filterL+1.)/recfield).astype(np.int)
outL = imL-filterL+1
# nfilter3 = 16
# filterL3 = 3
# conv3 = ConvLayer(input=conv2.output(), image_shape=(bs, nfilter2, outL2, outL2),
# filter_shape=(nfilter3, nfilter2, filterL3, filterL3),
# flatten=True,
# actfun=tanh,
# tag='_conv3')
#
# outL3 = outL2-filterL3+1
fc2 = FCLayer(input=conv1.output(), n_in=nfilter1*outL*outL, n_out=256, actfun=tanh, tag='_fc2')
fc3 = FCLayer(input=fc2.output(), n_in=256, n_out=10, actfun=sigmoid, tag='_fc3')
params_cmb = conv1.params + fc2.params + fc3.params
# ypred = fc3.output().flatten()
ypred = fc3.output()
model = GeneralModel(input=x, data=viper, output=ypred,
target=y, params=params_cmb,
regularizers=0,
cost_func=negative_log_likelihood,
error_func=sqr_error,
batch_size=bs)
sgd = sgd_optimizer(data=viper,
model=model,
batch_size=bs,
learning_rate=0.001,
n_epochs=500)
sgd.fit_viper()
filepath = datapath + 'model/model.pkl'
savefile(filepath, model)
os.system('ls -lh ' + filepath)