decoder_edge

[xiaolinpan]

Thank you for sharing your code! I have a question about the _decoder_edge function. ` def _decoder_edge(vec):

        vec = tf.layers.dense(vec, (self.n_node + self.n_dummy) * (self.n_node + self.n_dummy) * (self.dim_edge+1))
        vec = tf.reshape(vec, [batch_size, self.n_node + self.n_dummy, self.n_node + self.n_dummy, self.dim_edge+1])

        logit = (vec + tf.transpose(vec, perm = [0, 2, 1, 3])) / 2

        probs = tf.nn.softmax(logit)[:,:,:,:-1] * tf.reshape(1-tf.eye(self.n_node + self.n_dummy), [1, self.n_node + self.n_dummy, self.n_node + self.n_dummy, 1])

        return probs

` Could you tell me why logit and probs is calculated like this? I want to convert this code to pytorch, thank you!

[seokhokang]

Thanks for your interest in our code.

1. The edge matrix (and its prediction) must be symmetric because when it comes to undirected graphs. For the reason, "logit" is the average of "vec" and the transpose of "vec" to make it symmetric.
2. "probs" is the softmax transformation of "logit" for edge (categorical) prediction. The diagonal elements of "probs" have no meaning because they do not correspond to edges. They are therefore forced to zero.

[xiaolinpan]

Thanks for your explanation！ ｀ def _reward(nodes, edges):

        def _preference(smi):
            val = 0
            # mol = Chem.MolFromSmiles(smi)
            # set val = 1 if mol is preferred 

            return val

        R_t = np.zeros((self.batch_size, self.dim_R))
        for j in range(self.batch_size):
            try:
                R_smi = self._vec_to_mol(nodes[j], edges[j], atom_list, train=True)
                R_t[j, 0] = 1
                if self.dim_R == 2: R_t[j, 1] = _preference(R_smi)
            except:
                pass

        return R_t

｀　 the function _reward only return 0, Are you missing something, your paper said it will return 0 or 1.

[seokhokang]

No. it returns 1 when it succeeds processing the line "R_smi = self._vec_to_mol(nodes[j], edges[j], atom_list, train=True)"

[xiaolinpan]

thank you!!!

[YeolYao]

Thank you for sharing your code! I have a question about the _decoder_edge function. ` def _decoder_edge(vec):

        vec = tf.layers.dense(vec, (self.n_node + self.n_dummy) * (self.n_node + self.n_dummy) * (self.dim_edge+1))
        vec = tf.reshape(vec, [batch_size, self.n_node + self.n_dummy, self.n_node + self.n_dummy, self.dim_edge+1])

        logit = (vec + tf.transpose(vec, perm = [0, 2, 1, 3])) / 2

        probs = tf.nn.softmax(logit)[:,:,:,:-1] * tf.reshape(1-tf.eye(self.n_node + self.n_dummy), [1, self.n_node + self.n_dummy, self.n_node + self.n_dummy, 1])

        return probs

` Could you tell me why logit and probs is calculated like this? I want to convert this code to pytorch, thank you!

Hello! Have you successfully converted this code to pytorch?

[xiaolinpan]

Sorry, two years have passed and I have forgotten the project and lost the code.

[YeolYao]

OK, Thanks a lot. I'll try it myself.