return T.maximum(X, 0.)
def dropout(X, p=0.):
if p > 0:
retain_prob = 1 - p
X *= srng.binomial(X.shape, p=retain_prob, dtype=theano.config.floatX)
X /= retain_prob
return X
def softmax(X):
e_x = T.exp(X - X.max(axis=1).dimshuffle(0, 'x'))
return e_x / e_x.sum(axis=1).dimshuffle(0, 'x')
def get_hyper_params(learning_rate=0.10):
kern_one=8#32
kern_two=12#64
kern_third=16#128
n_hidden=400
kern_params=[(kern_one, 1, 3, 3),(kern_two, kern_one, 3, 3),(kern_third, kern_two, 3, 3),
(kern_third * 6 * 6, n_hidden),(n_hidden, 7)]
params={'learning_rate': learning_rate,
'kern_params':kern_params}
return params
if __name__ == '__main__':
dataset_path="/home/user/cf/conv_frames/cls/images/"
dataset=load.get_images(dataset_path)
out_path="/home/user/cf/exp1/conv_net"
cls=learning.create_classifer(dataset_path,out_path,built_conv_cls,flat=False)
learning.evaluate_cls(dataset_path,out_path,flat=False)
def create_cls_fun(free_vars,model):
learning_rate=0.15
py_x=get_px_y(free_vars,model)
loss=get_loss_function(free_vars,py_x)
input_vars=free_vars.get_vars()
g_W = T.grad(cost=loss, wrt=model.W)
g_b = T.grad(cost=loss, wrt=model.b)
update = [(model.W, model.W - learning_rate * g_W),
(model.b,model.b - learning_rate * g_b)]
train = theano.function(inputs=input_vars,
outputs=loss, updates=update,
allow_input_downcast=True)
y_pred = T.argmax(py_x, axis=1)
test=theano.function(inputs=[free_vars.X], outputs=y_pred,
allow_input_downcast=True)
return train,test
def get_px_y(free_vars,model):
equation=T.dot(free_vars.X, model.W) + model.b
return T.nnet.softmax(equation)
def get_loss_function(free_vars,py_x):
return T.mean(T.nnet.categorical_crossentropy(py_x,free_vars.y))
if __name__ == "__main__":
dataset_path="/home/user/cf/conv_frames/cls/images/"
#cls=built_classifer()
out_path="/home/user/cf/exp1/logit"
#ml_tools.evaluate_cls(dataset_path,cls)
learning.create_classifer(dataset_path,out_path,built_logit_cls)
def get_px_y(free_vars,model):
hidden=model.hidden
output_layer=model.logistic
h = T.nnet.sigmoid(T.dot(free_vars.X, hidden.W) + hidden.b)
pyx = T.nnet.softmax(T.dot(h, output_layer.W) + output_layer.b)
return pyx
def get_loss_function(free_vars,py_x):
return T.mean(T.nnet.categorical_crossentropy(py_x,free_vars.y))
def sgd(loss, params, learning_rate=0.05):
gparams = [T.grad(loss, param) for param in params]
updates = [
(param, param - learning_rate * gparam)
for param, gparam in zip(params, gparams)
]
return updates
def get_hyper_params(learning_rate=0.05):
params={'learning_rate': learning_rate,
'n_in':3200,'n_out':7,'n_hidden':800}
return params
if __name__ == "__main__":
dataset_path="/home/user/cf/conv_frames/cls/images/"
dataset=load.get_images(dataset_path)
out_path="/home/user/cf/exp1/nn"
cls=learning.create_classifer(dataset_path,out_path,built_nn_cls)
learning.evaluate_cls(dataset_path,out_path)