modular_neural_network.py 文件源码

def run_experiment(self, input_shape, labels, comb):
        # comb holds values like (32, (2,2), optimizers-Adam()). We need to use self.keys_mapper
        # which maps a name ("units", "kernel_sizes", "optimizers") to the position where it is
        # in comb. I wonder if it would be more comprehensible with a function like
        # get_element_from_comb(self, comb, key) { return comb[self.keys_mapper[key]] }
        opt = comb[self.keys_mapper["optimizers1"]]
        loss = comb[self.keys_mapper["losses1"]]
        f1 = comb[self.keys_mapper["filters1"]]
        f2 = comb[self.keys_mapper["filters2"]]
        u1 = comb[self.keys_mapper["units1"]]
        ks = comb[self.keys_mapper["kernel_sizes1"]]
        ps = comb[self.keys_mapper["pool_sizes1"]]
        d1 = comb[self.keys_mapper["dropouts1"]]
        d2 = comb[self.keys_mapper["dropouts2"]]
        return (opt, loss,
                Conv2D(f1, kernel_size=ks, activation='relu', input_shape=input_shape),
                Conv2D(f2, kernel_size=ks, activation='relu'),
                MaxPooling2D(pool_size=ps),
                Dropout(d1),
                Flatten(),
                Dense(u1, activation='relu'),
                Dropout(d2),
                Dense(len(labels), activation='softmax'))