dqn.py 文件源码

def build_critic(self):
        """Build critic network

        recieve transformed tensor: raw_data, smooted_data, and downsampled_data
        """
        nf = self.n_feature
        # layer1
        # smoothed input
        sm_model = [Sequential() for _ in range(self.n_smooth)]
        for m in sm_model:
            m.add(Lambda(lambda x: x,  input_shape=(self.history_length, self.n_stock, 1)))
            m.add(Convolution2D(nb_filter=nf, nb_row=self.k_w, nb_col=1, border_mode='same'))
            m.add(BatchNormalization(mode=2, axis=-1))
            m.add(PReLU())
        # down sampled input
        dw_model = [Sequential() for _ in range(self.n_down)]
        for m in dw_model:
            m.add(Lambda(lambda x: x,  input_shape=(self.history_length, self.n_stock, 1)))
            m.add(Convolution2D(nb_filter=nf, nb_row=self.k_w, nb_col=1, border_mode='same'))
            m.add(BatchNormalization(mode=2, axis=-1))
            m.add(PReLU())
        # raw input
        state = Sequential()
        nf = self.n_feature
        state.add(Lambda(lambda x: x,  input_shape=(self.history_length, self.n_stock, 1)))
        state.add(Convolution2D(nb_filter=nf, nb_row=self.k_w, nb_col=1, border_mode='same'))
        state.add(BatchNormalization(mode=2, axis=-1))
        state.add(PReLU())
        merged = Merge([state,] + sm_model + dw_model, mode='concat', concat_axis=-1)
        # layer2
        nf = nf * 2
        model = Sequential()
        model.add(merged)
        model.add(Convolution2D(nb_filter=nf, nb_row=self.k_w, nb_col=1, border_mode='same'))
        model.add(BatchNormalization(mode=2, axis=-1))
        model.add(PReLU())
        model.add(Flatten())
        # layer3
        model.add(Dense(self.n_hidden))
        model.add(BatchNormalization(mode=1, axis=-1))
        model.add(PReLU())
        # layer4
        model.add(Dense(int(np.sqrt(self.n_hidden))))
        model.add(PReLU())
        # output
        model.add(Dense(2 * self.n_stock))
        model.add(Reshape((self.n_stock, 2)))
        return model