Upload sCT

config.json

@@ -6,8 +6,8 @@
   "attention_heads": 16,
   "attention_maps_to_save": [],
   "auto_map": {
-    "AutoConfig": "sct.sCTConfig",
-    "AutoModel": "sct.sCT"
+    "AutoConfig": "pytorch_sct.sCTConfig",
+    "AutoModel": "pytorch_sct.sCT"
   },
   "cell_len": 19968,
   "embed_dim": 1024,

pytorch_sct.py

@@ -0,0 +1,756 @@
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from transformers import PretrainedConfig, PreTrainedModel
+
+
+class GeLU(nn.Module):
+    def __init__(self) -> None:
+        """
+        This is the gelu implementation from the original ESM repo.
+        Using F.gelu yields subtly wrong results.
+        """
+        super().__init__()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x * 0.5 * (1.0 + torch.erf(x / math.sqrt(2.0)))
+
+
+@dataclass
+class RotaryEmbeddingConfig:
+    """
+    Parameters to initialize the RotaryEmbedding layer. The rescaling factor allows
+    to adapt the rotary embeddings to larger lengths than what was used for training.
+    One of this strategy is presented in the Yarn paper: https://arxiv.org/pdf/2309.00071.pdf. # noqa
+    Args:
+    """
+
+    rescaling_factor: Optional[float]
+
+
+class RotaryEmbedding(torch.nn.Module):
+    """
+    Rotary position embeddings based on those in
+    [RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer).
+    Query and keys are transformed by rotation
+    matrices which depend on their relative positions.
+    """
+
+    def __init__(self, dim: int, rotary_embedding_config: RotaryEmbeddingConfig):
+        super().__init__()
+
+        # Extract argument from the config
+        self.rescaling_factor = rotary_embedding_config.rescaling_factor
+        self.upper_freq = 10000
+        self.dim = dim
+
+        self._seq_len_cached = None
+        self._cos_cached = None
+        self._sin_cached = None
+
+    def _apply_rotary_pos_emb(
+        self,
+        heads: torch.Tensor,
+        cos: torch.Tensor,
+        sin: torch.Tensor,
+    ) -> torch.Tensor:
+        """ """
+        x_first, x_second = (
+            heads[..., : heads.shape[-1] // 2],
+            heads[..., heads.shape[-1] // 2 :],
+        )
+
+        first_part = x_first * cos - x_second * sin
+        second_part = x_second * cos + x_first * sin
+
+        return torch.cat((first_part, second_part), dim=-1)
+
+    def _compute_cos_sin_tables(
+        self, x: torch.Tensor, inv_freq: torch.Tensor, seq_dimension: int = 2
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        seq_len = x.shape[seq_dimension]
+        # Reset the tables if the sequence length has changed,
+        # or if we're on a new device (possibly due to tracing for instance)
+        self._seq_len_cached = seq_len
+        t = torch.arange(x.shape[seq_dimension], device=x.device).type_as(inv_freq)
+        # freqs = torch.outer(t, inv_freq)
+        freqs = torch.einsum("i, j -> ij", t, inv_freq)
+
+        self._cos_cached = torch.cos(freqs)[None, :, None, :]
+        self._sin_cached = torch.sin(freqs)[None, :, None, :]
+        # emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+
+        # self._cos_cached = emb.cos()[None, None, :, :]
+        # self._sin_cached = emb.sin()[None, None, :, :]
+
+        return self._cos_cached, self._sin_cached
+
+    def forward(
+        self, q: torch.Tensor, k: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        if self.rescaling_factor is None:
+            inv_freq = 1.0 / (
+                self.upper_freq ** (torch.arange(0, self.dim, 2).float() / self.dim)
+            )
+        else:
+            updated_base = self.upper_freq * (
+                self.rescaling_factor ** (self.dim / (self.dim - 2))
+            )
+            inv_freq = 1.0 / (
+                updated_base ** (torch.arange(0, self.dim, 2).float() / self.dim)
+            )
+
+        self._cos_cached, self._sin_cached = self._compute_cos_sin_tables(
+            q,
+            inv_freq,
+            seq_dimension=-3,
+        )
+
+        return (
+            self._apply_rotary_pos_emb(q, self._cos_cached, self._sin_cached),
+            self._apply_rotary_pos_emb(k, self._cos_cached, self._sin_cached),
+        )
+
+
+class ResidualConvBlock(nn.Module):
+    """
+    Conv Block with Residual connection.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, seq_len: int, kernel_size: int = 1):
+        super().__init__()
+        self.conv_block = ConvBlock(
+            dim_in=dim_in, dim_out=dim_out, seq_len=seq_len, kernel_size=kernel_size
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        y = self.conv_block(x)
+        return x.reshape(y.shape) + y
+
+
+class ConvBlock(nn.Module):
+    """
+    Conv Block.
+    """
+
+    def __init__(self, dim_in: int, dim_out: int, seq_len: int, kernel_size: int = 1):
+        super().__init__()
+        self.conv = nn.Conv1d(
+            in_channels=dim_in,
+            out_channels=dim_out,
+            kernel_size=kernel_size,
+            padding="same",
+        )
+        self.layer_norm = nn.LayerNorm(seq_len, eps=1e-5)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.layer_norm(x)
+        x = x.reshape(x.shape[0], x.shape[1], -1)
+        x = self.conv(x)
+        x = F.gelu(x, approximate="tanh")
+        return x
+
+
+class ResidualDeConvBlock(nn.Module):
+    """
+    Conv Block with Residual connection.
+    """
+
+    def __init__(
+        self,
+        dim_in: int,
+        dim_out: int,
+        seq_len: int,
+        kernel_size: int = 1,
+        stride: int = 1,
+    ):
+        super().__init__()
+        self.deconv_block = DeConvBlock(
+            dim_in=dim_in,
+            dim_out=dim_out,
+            seq_len=seq_len,
+            kernel_size=kernel_size,
+            stride=stride,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        y = self.deconv_block(x)
+        return x.reshape(y.shape) + y
+
+
+class DeConvBlock(nn.Module):
+    """
+    DeConv Block.
+    """
+
+    def __init__(
+        self,
+        dim_in: int,
+        dim_out: int,
+        seq_len: int,
+        kernel_size: int = 1,
+        stride: int = 1,
+    ):
+        super().__init__()
+        self.deconv = nn.ConvTranspose1d(
+            in_channels=dim_in,
+            out_channels=dim_out,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=0,
+        )
+        self.layer_norm = nn.LayerNorm(seq_len)
+        self.kernel_size = kernel_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.layer_norm(x)
+        x = x.reshape(x.shape[0], x.shape[1], -1)
+        x = self.deconv(x)
+        if self.kernel_size == 5:
+            # handle the special case where haiku
+            # deconv removes padding automatically
+            x = x[:, :, 1:-2]
+        x = F.gelu(x, approximate="tanh")
+        return x
+
+
+class SpatialEncoding(nn.Module):
+    """
+    Spatial coordinates encoding module
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_scales: int = 10,
+        sigma_min: float = 1.0,
+        sigma_max: float = 10.0,
+    ):
+        super().__init__()
+        self.num_scales = num_scales
+        self.sigma_min = sigma_min
+        self.sigma_max = sigma_max
+        self.g = sigma_max / sigma_min
+        self.scales = torch.linspace(sigma_min, sigma_max, num_scales)
+        self.fc_layer = nn.Linear(embed_dim, embed_dim)
+
+    def scale_specific_encoder(
+        self, coordinates: torch.Tensor, scale: float
+    ) -> torch.Tensor:
+        x, y = coordinates[..., 0], coordinates[..., 1]
+        constant = self.sigma_min * (self.g ** (scale / (self.num_scales - 1)))
+        x_transform = torch.cos(x / constant)
+        y_transform = torch.sin(y / constant)
+        transformed_coordinates = torch.stack([x_transform, y_transform], dim=-1)
+        return transformed_coordinates
+
+    def forward(self, coordinates: torch.Tensor) -> torch.Tensor:
+        transformed_coordinates = [
+            self.scale_specific_encoder(coordinates, scale) for scale in self.scales
+        ]
+        transformed_coordinates = torch.cat(transformed_coordinates, dim=-1)
+        return self.fc_layer(transformed_coordinates)
+
+
+class ConvTowerBlock(nn.Module):
+    def __init__(
+        self, dim_in: int, dim_out: int, seq_len: int, kernel_size: int, num_cells: int
+    ) -> None:
+        super().__init__()
+        self.conv_layer = ConvBlock(
+            dim_in=dim_in, dim_out=dim_out, seq_len=seq_len, kernel_size=kernel_size
+        )
+        self.res_conv = ResidualConvBlock(
+            dim_in=dim_out, dim_out=dim_out, seq_len=seq_len, kernel_size=1
+        )
+        self.avg_pool = nn.AvgPool1d(kernel_size=2, stride=2)
+        self.num_cells = num_cells
+
+    def forward(self, x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
+        residual = x
+        x = x.reshape(x.shape[0], x.shape[1], self.num_cells, -1)  # noqa: FKA100
+        x = self.conv_layer(x)
+        x = x.reshape((x.shape[0], x.shape[1], self.num_cells, -1))
+        x = self.res_conv(x)
+        x = self.avg_pool(x)
+        return x, residual
+
+
+class DeConvTowerBlock(nn.Module):
+    def __init__(
+        self,
+        dim_in: int,
+        dim_out: int,
+        kernel_size: int,
+        seq_len: int,
+        stride: int = 2,
+        num_cells: int = 1,
+    ):
+        super().__init__()
+        self.deconv_block = DeConvBlock(
+            dim_in=dim_in,
+            dim_out=dim_out,
+            seq_len=seq_len,
+            kernel_size=kernel_size,
+            stride=stride,
+        )
+        self.res_deconv_block = ResidualDeConvBlock(
+            dim_in=dim_out, dim_out=dim_out, seq_len=seq_len * 2, kernel_size=1
+        )
+        self.num_cells = num_cells
+
+    def forward(self, x: torch.Tensor, res: torch.Tensor) -> torch.Tensor:
+        x = x.reshape((x.shape[0], x.shape[1], self.num_cells, -1))
+        x = self.deconv_block(x)
+        x = x.reshape((x.shape[0], x.shape[1], self.num_cells, -1))
+        x = self.res_deconv_block(x)
+
+        x = x + res
+        return x
+
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        num_heads: int,
+        key_size: int,
+        rotary_embedding_config: Optional[RotaryEmbeddingConfig] = None,
+        add_bias_kv: bool = False,
+        value_size: Optional[int] = None,
+        model_size: Optional[int] = None,
+        name: Optional[str] = None,
+    ):
+        super().__init__()
+        if not model_size:
+            model_size = key_size
+        if not value_size:
+            value_size = key_size
+        self.model_size = model_size
+        self.key_size = key_size
+        self.value_size = value_size
+        self.add_bias_kv = add_bias_kv
+        self.name = name
+        self.num_heads = num_heads
+        self._rotary_embedding_config = rotary_embedding_config
+
+        self.w_k = nn.Linear(self.model_size, self.num_heads * self.key_size)
+        self.w_q = nn.Linear(self.model_size, self.num_heads * self.key_size)
+        self.w_v = nn.Linear(self.model_size, self.num_heads * self.value_size)
+        self.output = nn.Linear(self.num_heads * self.value_size, self.model_size)
+        if self._rotary_embedding_config:
+            self._rotary_embedding = RotaryEmbedding(
+                self.key_size, self._rotary_embedding_config
+            )
+
+    def apply_rotary_embeddings(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        """ """
+        query, key = self._rotary_embedding(query, key)
+        return query, key
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_weight_bias: Optional[torch.Tensor] = None,
+    ) -> dict[str, torch.Tensor]:
+        """
+        Returns:
+            dictionary containing attention weights
+            and outputs.
+        """
+        key_heads = self.w_k(key).reshape(
+            (*key.shape[:-1], self.num_heads, self.key_size)
+        )
+        query_heads = self.w_q(query).reshape(
+            (*query.shape[:-1], self.num_heads, self.key_size)
+        )
+        value_heads = self.w_v(value).reshape(
+            (*value.shape[:-1], self.num_heads, self.value_size)
+        )
+        if self._rotary_embedding_config:
+            query_heads, key_heads = self.apply_rotary_embeddings(
+                query_heads, key_heads
+            )
+        attention_weights = torch.einsum(
+            "...thd, ...Thd -> ...htT", query_heads, key_heads
+        )
+        sqrt_key_size = np.sqrt(self.key_size)
+        attention_weights = attention_weights / sqrt_key_size
+        if attention_mask:
+            attention_weights = torch.where(attention_mask, attention_weights, -1e30)
+        if attention_weight_bias:
+            attention_weights = F.softmax(
+                attention_weights + attention_weight_bias, dim=-1
+            )
+        else:
+            attention_weights = F.softmax(attention_weights, dim=-1)
+        value_out = torch.einsum(
+            "...htT, ...Thd->...thd", attention_weights, value_heads
+        )
+        value_out = value_out.reshape((*value_out.shape[:-2], -1))
+        embeddings = self.output(value_out)
+
+        return {"attention_weights": attention_weights, "embeddings": embeddings}
+
+
+class SelfAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        num_heads: int,
+        embed_dim: int,
+        ffn_embed_dim: int,
+        key_size: Optional[int] = None,
+        add_bias_kv: bool = False,
+        add_bias_fnn: bool = True,
+        ffn_activation_name: str = "gelu-no-approx",
+        use_glu_in_ffn: bool = False,
+        layer_norm_eps: float = 1e-5,  # this is the default haiku value
+        pre_layer_norm: bool = True,
+        name: Optional[str] = None,
+        rotary_embedding_config: Optional[RotaryEmbeddingConfig] = None,
+    ):
+        super().__init__()
+        if key_size is None:
+            if embed_dim % num_heads != 0:
+                raise ValueError(
+                    f"The embedding dimension should be divisible by the number of "
+                    f"heads, however provided embedding dimension is {embed_dim} and "
+                    f"the number of heads is {num_heads}."
+                )
+            else:
+                key_size = embed_dim // num_heads
+
+        # Get ffn activation function
+        self._pre_layer_norm = pre_layer_norm
+        self._use_glu_in_fnn = use_glu_in_ffn
+        # Define layers
+        if use_glu_in_ffn:
+            # user should multiply ffn_embed_dim by 2/3 when using GLU
+            # to keep total number of parameters equal
+            # see https://arxiv.org/pdf/2002.05202.pdf. for more details
+            # we multiply by 2 here as the output will be split in 2 for GLU
+            self.fc1 = nn.Linear(embed_dim, int(2 * ffn_embed_dim), bias=add_bias_fnn)
+        else:
+            self.fc1 = nn.Linear(embed_dim, ffn_embed_dim, bias=add_bias_fnn)
+
+        self.fc2 = nn.Linear(ffn_embed_dim, embed_dim, bias=add_bias_fnn)
+
+        self.layer_norm_self_attention = nn.LayerNorm(
+            embed_dim,
+        )
+        self.layer_norm_mlp = nn.LayerNorm(embed_dim)
+        if ffn_activation_name == "swish":
+            self._ffn_activation_fn = nn.SiLU()
+        elif ffn_activation_name == "gelu-no-approx":
+            self._ffn_activation_fn = nn.GeLU(approximate="tanh")
+        else:
+            self._ffn_activation_fn = getattr(torch.nn, ffn_activation_name)
+
+        self.mha = MultiHeadAttention(
+            num_heads=num_heads,
+            key_size=key_size,
+            add_bias_kv=add_bias_kv,
+            model_size=embed_dim,
+            name="self_attention",
+            rotary_embedding_config=rotary_embedding_config,
+        )
+
+    def mlp(self, embed: torch.Tensor) -> torch.Tensor:
+
+        if self._pre_layer_norm:
+            x = self.layer_norm_mlp(embed)
+        else:
+            x = embed
+
+        if self._use_glu_in_fnn:
+            x = self.fc1(x)
+            x1, x2 = torch.split(x, split_size_or_sections=x.shape[-1] // 2, dim=-1)
+            x = self._ffn_activation_fn(x1) * x2
+        else:
+            x = self._ffn_activation_fn(self.fc1(x))
+        x = self.fc2(x)
+
+        if not self._pre_layer_norm:
+            x = self.layer_norm_mlp(x + embed)
+        return x
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        attention_weight_bias: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        res = x
+        if self._pre_layer_norm:
+            x = self.layer_norm_self_attention(x)
+
+        output = self.mha(
+            x,
+            x,
+            x,
+            attention_mask=attention_mask,
+            attention_weight_bias=attention_weight_bias,
+        )
+
+        if not self._pre_layer_norm:
+            output["embeddings"] = self.layer_norm_self_attention(
+                output["embeddings"] + res
+            )
+
+            x = output["embeddings"]
+        else:
+            x = output["embeddings"]
+            x = res + x
+
+        # MLP
+        if not self._pre_layer_norm:
+            x = self.mlp(x)
+        else:
+            x = x + self.mlp(x)
+
+        output["embeddings"] = x
+        return output
+
+
+class LMHead(nn.Module):
+    def __init__(
+        self, dim_in: int, embed_dim: int, dim_out: int, num_hidden_layers: int
+    ) -> None:
+        """ """
+        super().__init__()
+        self.num_hidden_layers = num_hidden_layers
+        self.linear_layers = nn.ModuleList([nn.Linear(dim_in, embed_dim)])
+        self.linear_layers.extend(
+            nn.ModuleList(
+                [nn.Linear(embed_dim, embed_dim)] for _ in range(num_hidden_layers - 1)
+            )
+        )
+        self.linear_out = nn.Linear(embed_dim, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        res = x  # noqa: F841
+        x = F.gelu(x, approximate="tanh")
+        for layer in self.linear_layers:
+            x = layer(x)
+            x = F.gelu(x, approximate="tanh")
+        out = self.linear_out(x)
+        return out
+
+
+@dataclass
+class sCTConfig(PretrainedConfig):  # noqa: N801
+    model_type = "sCT"
+
+    def __init__(self, **kwargs):  # type: ignore
+        self.alphabet_size = kwargs.get("alphabet_size", 7)
+        self.pad_token_id = kwargs.get("pad_token_id", 5)
+        self.mask_token_id = kwargs.get("mask_token_id", 6)
+        self.cell_len = kwargs.get("cell_len", 19968)
+
+        self.num_downsamples = kwargs.get("num_downsamples", 8)
+        self.attention_heads = kwargs.get("attention_heads", 16)
+        self.key_size = kwargs.get("key_size", None)
+        self.token_embed_dim = kwargs.get("token_embed_dim", 16)
+
+        self.embed_dim = kwargs.get("embed_dim", 1024)
+        self.ffn_embed_dim = kwargs.get("ffn_embed_dim", 2048)
+        self.num_layers = kwargs.get("num_layers", 4)
+        self.layer_norm_eps = kwargs.get("layer_norm_eps", 1e-5)
+        self.interpolation_method = kwargs.get("interpolation_method", "nearest")
+
+        # bad hack to satisfy cellnt_celltype_annotation.py:312
+        self.max_positions: int = kwargs.get("max_positions", 20480)
+        self.num_cells: int = kwargs.get("num_cells", 50)
+        self.num_hidden_layers_head: int = kwargs.get("num_hidden_layers_head", 1)
+
+        self.use_skip_connection: bool = kwargs.get("use_skip_connection", True)
+
+        # logging
+        self.use_gradient_checkpointing: bool = False
+
+        # return
+        self.embeddings_layers_to_save: Tuple[int, ...] = kwargs.get(
+            "embeddings_layers_to_save", ()
+        )
+        self.attention_maps_to_save: list[tuple[int, int]] = kwargs.get(
+            "attention_maps_to_save", []
+        )
+
+        # Spatial info configuration
+        self.use_spatial_information: bool = kwargs.get(
+            "use_spatial_information", False
+        )
+        self.num_scales: int = kwargs.get("num_scales", 10)
+        self.sigma_min: float = kwargs.get("sigma_min", 1.0)
+        self.sigma_max: float = kwargs.get("sigma_max", 10.0)
+
+        super().__init__(**kwargs)
+
+        def __post_init__(self) -> None:  # type: ignore # noqa: N807
+            """
+            Checks that the given values are compatible.
+            """
+            if self.key_size is None:
+                if not self.embed_dim % self.attention_heads == 0:
+                    raise ValueError(
+                        f"When no key size is provided, the embedding dimension"
+                        f"should be divisible by the number of heads, however "
+                        f"provided embedding dimension is {self.embed_dim} and "
+                        f"the number of heads is {self.attention_heads}."
+                    )
+                self.key_size = self.embed_dim // self.attention_heads
+
+
+class sCT(PreTrainedModel):  # noqa: N801
+    config_class = sCTConfig
+
+    def __init__(self, config: sCTConfig):
+        # super().__init__(config)
+        super().__init__(config=config)
+        if config.use_spatial_information:
+            self.spatial_embed_layer = SpatialEncoding(
+                embed_dim=config.token_embed_dim,
+                num_scales=config.num_scales,
+                sigma_min=config.sigma_min,
+                sigma_max=config.sigma_max,
+            )
+        self.cell_len = config.cell_len
+
+        self.token_embed = nn.Embedding(config.alphabet_size, config.token_embed_dim)
+
+        attention_maps_to_save = config.attention_maps_to_save
+        self._attention_layers_to_save = list({t[0] for t in attention_maps_to_save})
+
+        self._attention_maps_per_layer_to_save = {
+            layer: [t[1] for t in attention_maps_to_save if t[0] == layer]
+            for layer in self._attention_layers_to_save
+        }
+
+        max_layer = max(self._attention_layers_to_save + [0])
+        if max_layer > config.num_layers:
+            raise ValueError(
+                f"You are requiring attention maps for layer {max_layer}, "
+                f"while the model has {config.num_layers} layers only."
+            )
+
+        filter_list = np.linspace(
+            config.token_embed_dim,
+            config.embed_dim,
+            config.num_downsamples + 1,
+        )
+
+        filter_list = np.ceil(filter_list / 32) * 32
+        filter_list = filter_list.astype(int).tolist()
+
+        self._filter_list = filter_list
+        self._rotary_embedding_config = RotaryEmbeddingConfig(rescaling_factor=None)
+
+        self.stem_conv = nn.Sequential(
+            nn.Conv1d(
+                in_channels=config.token_embed_dim,
+                out_channels=config.token_embed_dim,
+                kernel_size=15,
+                padding="same",
+            ),
+            nn.GELU(approximate="tanh"),
+        )
+        downsampled_seq_lens = [
+            self.cell_len // (2**i) for i in range(len(filter_list) - 1)
+        ]
+
+        self.conv_tower = nn.ModuleList(
+            [
+                ConvTowerBlock(
+                    dim_in=self._filter_list[i],
+                    dim_out=self._filter_list[i + 1],
+                    kernel_size=5,
+                    seq_len=seq_len,
+                    num_cells=config.num_cells,
+                )
+                for i, seq_len in zip(range(len(filter_list) - 1), downsampled_seq_lens)
+            ]
+        )
+
+        self.deconv_tower = nn.ModuleList(
+            [
+                DeConvTowerBlock(
+                    dim_in=filter_list[-1 - i],
+                    dim_out=filter_list[-1 - i - 1],
+                    kernel_size=5,
+                    stride=2,
+                    seq_len=seq_len // 2,
+                    num_cells=config.num_cells,
+                )
+                for i, seq_len in zip(
+                    range(len(filter_list) - 1), downsampled_seq_lens[::-1]
+                )
+            ]
+        )
+        self.transformer_layers = nn.ModuleList(
+            [
+                SelfAttentionBlock(
+                    num_heads=config.attention_heads,
+                    embed_dim=config.embed_dim,
+                    ffn_embed_dim=config.ffn_embed_dim,
+                    key_size=config.key_size,
+                    add_bias_kv=False,
+                    add_bias_fnn=False,
+                    ffn_activation_name="swish",
+                    use_glu_in_ffn=True,
+                    layer_norm_eps=1e-5,  # this is the default haiku value
+                    pre_layer_norm=True,
+                    name=f"attention_layer_{layer_idx}",
+                    rotary_embedding_config=self._rotary_embedding_config,
+                )
+                for layer_idx in range(config.num_layers)
+            ]
+        )
+
+        self.lm_head = LMHead(
+            dim_in=config.token_embed_dim,
+            embed_dim=config.embed_dim,
+            dim_out=config.alphabet_size,
+            num_hidden_layers=config.num_hidden_layers_head,
+        )
+
+    def forward(self, input_ids: torch.Tensor) -> dict[str, torch.Tensor]:
+        outs = {}
+        embeddings = self.token_embed(input_ids)
+        x = embeddings.permute(0, 2, 1)
+        x = self.stem_conv(x)
+        residuals = []
+        for _idx, conv_block in enumerate(self.conv_tower):
+            x, res = conv_block(x)
+            residuals.append(res)
+        residuals = residuals[::-1]
+        x = x.permute(0, 2, 1)
+
+        for layer_idx, transformer in enumerate(self.transformer_layers):
+            output = transformer(x)
+            x = output["embeddings"]
+            if (layer_idx + 1) in self.config.embeddings_layers_to_save:
+                outs[f"embeddings_{(layer_idx + 1)}"] = output["embeddings"]
+            if (layer_idx + 1) in self._attention_layers_to_save:
+                for map_number in self._attention_maps_per_layer_to_save[layer_idx + 1]:
+                    dkey = f"attention_map_layer_{layer_idx + 1}_number_{map_number}"
+                    outs[dkey] = output["attention_weights"][:, map_number + 1]
+        x = x.permute(0, 2, 1)
+        for deconv_block, res in zip(self.deconv_tower, residuals):
+            x = deconv_block(x, res)
+        x = x.permute(0, 2, 1)
+        logits = self.lm_head(x)
+        outs["logits"] = logits
+
+        return outs