minihack/agent/polybeast/models/base.py [276:342]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
            )
        elif self.crop_model == "cnn":
            conv_extract_crop = [
                nn.Conv2d(
                    in_channels=in_channels[i],
                    out_channels=out_channels[i],
                    kernel_size=(F, F),
                    stride=S,
                    padding=P,
                )
                for i in range(L)
            ]

            self.extract_crop_representation = nn.Sequential(
                *interleave(conv_extract_crop, [nn.ELU()] * len(conv_extract))
            )

        # MESSAGING MODEL
        if "msg" not in flags:
            self.msg_model = "none"
        else:
            self.msg_model = flags.msg.model
            self.msg_hdim = flags.msg.hidden_dim
            self.msg_edim = flags.msg.embedding_dim
        if self.msg_model in ("gru", "lstm", "lt_cnn"):
            # character-based embeddings
            self.char_lt = nn.Embedding(
                NUM_CHARS, self.msg_edim, padding_idx=PAD_CHAR
            )
        else:
            # forward will set up one-hot inputs for the cnn, no lt needed
            pass

        if self.msg_model.endswith("cnn"):
            # from Zhang et al, 2016
            # Character-level Convolutional Networks for Text Classification
            # https://arxiv.org/abs/1509.01626
            if self.msg_model == "cnn":
                # inputs will be one-hot vectors, as done in paper
                self.conv1 = nn.Conv1d(NUM_CHARS, self.msg_hdim, kernel_size=7)
            elif self.msg_model == "lt_cnn":
                # replace one-hot inputs with learned embeddings
                self.conv1 = nn.Conv1d(
                    self.msg_edim, self.msg_hdim, kernel_size=7
                )
            else:
                raise NotImplementedError("msg.model == %s", flags.msg.model)

            # remaining convolutions, relus, pools, and a small FC network
            self.conv2_6_fc = nn.Sequential(
                nn.ReLU(),
                nn.MaxPool1d(kernel_size=3, stride=3),
                # conv2
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=7),
                nn.ReLU(),
                nn.MaxPool1d(kernel_size=3, stride=3),
                # conv3
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
                nn.ReLU(),
                # conv4
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
                nn.ReLU(),
                # conv5
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
                nn.ReLU(),
                # conv6
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



minihack/agent/rllib/models.py [270:336]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
            )
        elif self.crop_model == "cnn":
            conv_extract_crop = [
                nn.Conv2d(
                    in_channels=in_channels[i],
                    out_channels=out_channels[i],
                    kernel_size=(F, F),
                    stride=S,
                    padding=P,
                )
                for i in range(L)
            ]

            self.extract_crop_representation = nn.Sequential(
                *interleave(conv_extract_crop, [nn.ELU()] * len(conv_extract))
            )

        # MESSAGING MODEL
        if "msg" not in flags:
            self.msg_model = "none"
        else:
            self.msg_model = flags.msg.model
            self.msg_hdim = flags.msg.hidden_dim
            self.msg_edim = flags.msg.embedding_dim
        if self.msg_model in ("gru", "lstm", "lt_cnn"):
            # character-based embeddings
            self.char_lt = nn.Embedding(
                NUM_CHARS, self.msg_edim, padding_idx=PAD_CHAR
            )
        else:
            # forward will set up one-hot inputs for the cnn, no lt needed
            pass

        if self.msg_model.endswith("cnn"):
            # from Zhang et al, 2016
            # Character-level Convolutional Networks for Text Classification
            # https://arxiv.org/abs/1509.01626
            if self.msg_model == "cnn":
                # inputs will be one-hot vectors, as done in paper
                self.conv1 = nn.Conv1d(NUM_CHARS, self.msg_hdim, kernel_size=7)
            elif self.msg_model == "lt_cnn":
                # replace one-hot inputs with learned embeddings
                self.conv1 = nn.Conv1d(
                    self.msg_edim, self.msg_hdim, kernel_size=7
                )
            else:
                raise NotImplementedError("msg.model == %s", flags.msg.model)

            # remaining convolutions, relus, pools, and a small FC network
            self.conv2_6_fc = nn.Sequential(
                nn.ReLU(),
                nn.MaxPool1d(kernel_size=3, stride=3),
                # conv2
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=7),
                nn.ReLU(),
                nn.MaxPool1d(kernel_size=3, stride=3),
                # conv3
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
                nn.ReLU(),
                # conv4
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
                nn.ReLU(),
                # conv5
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
                nn.ReLU(),
                # conv6
                nn.Conv1d(self.msg_hdim, self.msg_hdim, kernel_size=3),
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -