def mutate_arc()

in nasrec/base_searcher.py [0:0]

• 76 lines of code
• 15 McCabe index (conditional complexity)

    def mutate_arc(self, parent):
        child = deepcopy(parent)
        # 1. choose block to mutate
        block_id = np.random.choice(self.num_blocks, 1)[0]
        # 2. choose one token of a block to mutate (e.g., block_type, dense_feat)
        token_name = np.random.choice(self.token_names, 1)[0]
        while token_name == "skip_connect" and block_id == 0:
            block_id = np.random.choice(self.num_blocks, 1)[0]
            token_name = np.random.choice(self.token_names, 1)[0]
        while (
            token_name == "cin"
            and len(self.micro_cin_option.arc) == 1
            and len(self.micro_cin_option.num_of_layers) == 1
        ) or (
            token_name == "attention"
            and self.att_num_tokens["head"] == 1
            and self.att_num_tokens["layer"] == 1
            and self.att_num_tokens["emb"] == 1
            and self.att_num_tokens["drop"] == 1
        ):
            token_name = np.random.choice(self.token_names, 1)[0]
        # 3. mutate the corresponding token
        new_action = child[block_id][token_name]
        while self._action_equal(new_action, child[block_id][token_name]):
            if token_name in ["block_type", "mlp_dense", "mlp_emb"]:
                new_action_vec = np.random.multinomial(
                    1, [1.0 / self.num_tokens[token_name]] * self.num_tokens[token_name]
                )
                new_action = np.argmax(new_action_vec)
            elif token_name == "cin":
                cin_width = np.argmax(
                    np.random.multinomial(
                        1,
                        [1.0 / len(self.micro_cin_option.arc)]
                        * len(self.micro_cin_option.arc),
                    )
                )
                cin_depth = np.argmax(
                    np.random.multinomial(
                        1,
                        [1.0 / len(self.micro_cin_option.num_of_layers)]
                        * len(self.micro_cin_option.num_of_layers),
                    )
                )
                new_action = {"width": cin_width, "depth": cin_depth}
            elif token_name == "attention":
                head = np.argmax(
                    np.random.multinomial(
                        1,
                        [1.0 / self.att_num_tokens["head"]]
                        * self.att_num_tokens["head"],
                    )
                )
                layer = np.argmax(
                    np.random.multinomial(
                        1,
                        [1.0 / self.att_num_tokens["layer"]]
                        * self.att_num_tokens["layer"],
                    )
                )
                emb = np.argmax(
                    np.random.multinomial(
                        1,
                        [1.0 / self.att_num_tokens["emb"]] * self.att_num_tokens["emb"],
                    )
                )
                drop = np.argmax(
                    np.random.multinomial(
                        1,
                        [1.0 / self.att_num_tokens["drop"]]
                        * self.att_num_tokens["drop"],
                    )
                )
                new_action = {"head": head, "layer": layer, "emb": emb, "drop": drop}
            else:
                new_action = np.random.binomial(1, 0.5, self.num_tokens[token_name])
        child[block_id][token_name] = new_action
        vecs = self.dicts_to_vecs(child)
        return vecs, child