__author__ = 'carlxie'

from layer.util import *
from layer.Layer import Layer
from layer.LayerType import LayerType
from layer.FullConLayer import FullConLayer

class Network(object):

    def __init__(self,neurons_per_layer,eta):
        self.neurons_per_layer = neurons_per_layer
        self.eta = eta

        inputLayer = Layer(None,LayerType.INPUT_TYPE,neurons_per_layer[0],1,1)
        self.layers = [inputLayer]
        preLayer = inputLayer
        for i in range(1,len(neurons_per_layer),1):
            layer = FullConLayer(preLayer,neurons_per_layer[i])
            self.layers.append(layer)
            preLayer = layer

    def feed_forward(self,x):
        inputLayer = self.layers[0]
        inputLayer.output[:,0,0] = x
        for layer in self.layers:
            layer.forward_prop()

    def back_forward(self,x,y):
        outputLayer = self.layers[-1]
        outputLayer.delta = self.get_last_delta(x,y)
        for i in range(len(self.layers)-1,0,-1):
            self.layers[i].back_prop()

    def get_last_delta(self,x,y):
        self.feed_forward(x)
        outputLayer = self.layers[-1]
        return -(vec_output(y,self.neurons_per_layer[-1]) - outputLayer.output[:,:,0])* 2 * vec_der_tanh(outputLayer.score)

    def train(self,training_data,T):
        for t in range(T):
            data = training_data[int(np.random.random()*len(training_data))]
            self.back_forward(data[:-1],data[-1])
            for layer in self.layers:
                layer.update_weights(self.eta)

    def cost(self,x,y):
        self.feed_forward(x)
        outputLayer = self.layers[-1]
        loss = sum((outputLayer.output[:,:,0] - vec_output(y,self.neurons_per_layer[-1])) ** 2)
        return loss

    def get_analytic_grads(self,x,y):
        self.back_forward(x,y)
        grads = []
        for i in range(1,len(self.layers),1):
            grads.append(self.layers[i].w_grads[0])
        for i in range(1,len(self.layers),1):
            grads.append(self.layers[i].b_grads[0])
        return flat_list(grads)

    def set_weights_biases(self,w,b):
        for i in range(1,len(self.layers),1):
            self.layers[i].biases = b[i-1]
            self.layers[i].weights = w[i-1]

    def cal_loss(self,x,y,w,b):
        self.set_weights_biases(w,b)
        return self.cost(x,y)

    def predict(self,x):
        self.feed_forward(x)
        return self.layers[-1].output

    def real_output(self,x):
        if x > 0:
            return 1
        else:
            return -1

    def evaluate(self, X, Y):
        error = 0.0
        for i in range(len(X)):
            if self.real_output(self.predict(X[i])) != Y[i]:
                error += 1
            print self.predict(X[i])
        return error / len(X)


def grad_check():
    train_data = np.loadtxt("t_train.dat")
    nn = Network([2,3,4,1],0.01)
    shapes = get_shapes(nn)
    ws = get_weights(nn)
    bs = get_biases(nn)
    weights = np.append(ws,bs)
    x = train_data[0][:-1]
    y = train_data[0][-1]
    num_grads = compute_num_grads(nn,x,y,shapes,weights)
    w11,b11 = reconstruct(weights,shapes)
    nn.set_weights_biases(w11,b11)
    analytic_grads = nn.get_analytic_grads(x,y)
    print num_grads
    print analytic_grads
    diff = num_grads - analytic_grads
    num = abs(sum(diff))
    den = abs(sum(num_grads)) + abs(sum(analytic_grads))
    print num / den

def real_test():
    train_data = np.loadtxt("t_train.dat")
    nn = Network([2,3,4,1],0.01)
    nn.train(train_data,100000)
    test_data = np.loadtxt("t_test.dat")
    print nn.evaluate(test_data[:,:-1],test_data[:,-1])

if __name__ == "__main__":
    real_test()
    #grad_check()