def train_model(dataset):
X = T.tensor4()
A = T.matrix()
layers = []
layers.append(ConvPoolLayer(
input=X,
input_shape=(3, 160, 80),
filter_shape=(32, 3, 5, 5),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
))
layers.append(ConvPoolLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
filter_shape=(64, 32, 5, 5),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=True,
))
"""
layers.append(ConvPoolLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
filter_shape=(128, 64, 3, 3),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
b=0.0
))
"""
layers.append(FullConnLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
output_shape=32,
dropout_ratio=0.1,
active_func=actfuncs.tanh
))
layers.append(FullConnLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
output_shape=11,
dropout_input=layers[-1].dropout_output,
active_func=actfuncs.sigmoid
))
model = NeuralNet(layers, X, layers[-1].output)
model.target = A
model.cost = costfuncs.binxent(layers[-1].dropout_output, A) + \
1e-3 * model.get_norm(2)
model.error = costfuncs.binerr(layers[-1].output, A)
sgd.train(model, dataset, lr=1e-2, momentum=0.9,
batch_size=100, n_epochs=300,
epoch_waiting=10)
return model
def train_model(dataset, use_scpool):
"""
def shape_constrained_pooling(fmaps):
beta = 100.0
s = fmaps.sum(axis=[2, 3])
# Z = abs(actfuncs.tanh(beta * fmaps)).sum(axis=[2, 3])
Z = T.neq(fmaps, 0).sum(axis=[2, 3])
return s / Z
"""
def shape_constrained_pooling(F, S):
masked_F = T.switch(T.eq(S, 1), F, 0)
#masked_F = F * S
s = masked_F.sum(axis=[2, 3])
Z = T.neq(masked_F, 0).sum(axis=[2, 3])
return s / Z
X = T.tensor4()
A = T.matrix()
S = T.tensor3()
layers = []
layers.append(ConvPoolLayer(
input=X,
input_shape=(3, 160, 80),
filter_shape=(32, 3, 5, 5),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
b=0.0
))
layers.append(ConvPoolLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
filter_shape=(64, 32, 5, 5),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
b=0.0
))
"""
layers.append(ConvPoolLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
filter_shape=(128, 64, 3, 3),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
b=0.0
))
"""
"""
F = layers[-1].output * S.dimshuffle(0, 'x', 1, 2)
F = shape_constrained_pooling(F) if use_scpool else \
F.flatten(2)
"""
F = layers[-1].output
F = shape_constrained_pooling(F, S.dimshuffle(0, 'x', 1, 2)) if use_scpool else F.flatten(2)
layers.append(FullConnLayer(
input=F,
input_shape=layers[-1].output_shape[0],
output_shape=32,
dropout_ratio=0.1,
active_func=actfuncs.tanh
))
layers.append(FullConnLayer(
input=layers[-1].output,
input_shape=layers[-1].output_shape,
output_shape=11,
dropout_input=layers[-1].dropout_output,
active_func=actfuncs.sigmoid
))
model = NeuralNet(layers, [X, S], layers[-1].output)
model.target = A
model.cost = costfuncs.binxent(layers[-1].dropout_output, A) + \
1e-3 * model.get_norm(2)
model.error = costfuncs.binerr(layers[-1].output, A)
sgd.train(model, dataset, lr=1e-2, momentum=0.9,
batch_size=100, n_epochs=300,
epoch_waiting=10)
return model
def train_model(dataset, attr_model, seg_model):
def shape_constrained_pooling(fmaps):
s = fmaps.sum(axis=[2, 3])
Z = abs(actfuncs.tanh(fmaps)).sum(axis=[2, 3])
return s / Z
X = T.tensor4()
A = T.matrix()
feature_layers = []
feature_layers.append(
ConvPoolLayer(
input=X,
input_shape=(3, 160, 80),
filter_shape=(32, 3, 5, 5),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
W=attr_model.blocks[0]._W,
b=0.0,
)
)
feature_layers.append(
ConvPoolLayer(
input=feature_layers[-1].output,
input_shape=feature_layers[-1].output_shape,
filter_shape=(64, 32, 5, 5),
pool_shape=(2, 2),
active_func=actfuncs.tanh,
flatten=False,
W=attr_model.blocks[1]._W,
b=0.0,
)
)
seg_layers = []
seg_layers.append(
FullConnLayer(
input=feature_layers[-1].output.flatten(2),
input_shape=np.prod(feature_layers[-1].output_shape),
output_shape=1024,
dropout_ratio=0.1,
active_func=actfuncs.tanh,
W=seg_model.blocks[2]._W,
b=seg_model.blocks[2]._b,
)
)
seg_layers.append(
FullConnLayer(
input=seg_layers[-1].output,
input_shape=seg_layers[-1].output_shape,
output_shape=37 * 17,
dropout_input=seg_layers[-1].dropout_output,
active_func=actfuncs.sigmoid,
W=seg_model.blocks[3]._W,
b=seg_model.blocks[3]._b,
)
)
S = seg_layers[-1].output
S = S * (S >= 0.1)
S = S.reshape((S.shape[0], 37, 17))
S = S.dimshuffle(0, "x", 1, 2)
S_dropout = seg_layers[-1].dropout_output
S_dropout = S_dropout * (S_dropout >= 0.1)
S_dropout = S_dropout.reshape((S_dropout.shape[0], 37, 17))
S_dropout = S_dropout.dimshuffle(0, "x", 1, 2)
attr_layers = []
"""
attr_layers.append(ConvPoolLayer(
input=feature_layers[-1].output * S,
input_shape=feature_layers[-1].output_shape,
filter_shape=(128, 64, 3, 3),
pool_shape=(2, 2),
dropout_input=feature_layers[-1].output * S_dropout,
active_func=actfuncs.tanh,
flatten=False,
W=attr_model.blocks[2]._W,
b=0.0
))
"""
attr_layers.append(
FullConnLayer(
input=shape_constrained_pooling(feature_layers[-1].output * S),
input_shape=feature_layers[-1].output_shape,
output_shape=64,
dropout_input=shape_constrained_pooling(feature_layers[-1].dropout_output * S_dropout),
dropout_ratio=0.1,
active_func=actfuncs.tanh,
W=attr_model.blocks[2]._W,
b=attr_model.blocks[2]._b,
)
)
attr_layers.append(
FullConnLayer(
input=attr_layers[-1].output,
input_shape=attr_layers[-1].output_shape,
output_shape=11,
dropout_input=attr_layers[-1].dropout_output,
active_func=actfuncs.sigmoid,
W=attr_model.blocks[3]._W,
b=attr_model.blocks[3]._b,
)
)
model = NeuralNet(feature_layers + seg_layers + attr_layers, X, attr_layers[-1].output)
model.target = A
model.cost = costfuncs.binxent(attr_layers[-1].dropout_output, A) + 1e-3 * model.get_norm(2)
model.error = costfuncs.binerr(attr_layers[-1].output, A)
sgd.train(model, dataset, lr=1e-3, momentum=0.9, batch_size=100, n_epochs=300, epoch_waiting=10)
return model
