minicpmo-utils 0.1.0__py3-none-any.whl

stepaudio2/flashcosyvoice/modules/flow_components/estimator.py

@@ -0,0 +1,974 @@
+import math
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from diffusers.models.attention import (GEGLU, GELU, AdaLayerNorm,
+                                        AdaLayerNormZero, ApproximateGELU)
+from diffusers.models.attention_processor import Attention
+from diffusers.models.lora import LoRACompatibleLinear
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from einops import pack, rearrange, repeat
+
+from stepaudio2.flashcosyvoice.modules.flow_components.upsample_encoder import \
+    add_optional_chunk_mask
+
+
+def mask_to_bias(mask: torch.Tensor, dtype: torch.dtype) -> torch.Tensor:
+    assert mask.dtype == torch.bool
+    assert dtype in [torch.float32, torch.bfloat16, torch.float16]
+    mask = mask.to(dtype)
+    # attention mask bias
+    # NOTE(Mddct): torch.finfo jit issues
+    #     chunk_masks = (1.0 - chunk_masks) * torch.finfo(dtype).min
+    mask = (1.0 - mask) * -1.0e+10
+    return mask
+
+
+class SnakeBeta(nn.Module):
+    """
+    A modified Snake function which uses separate parameters for the magnitude of the periodic components
+    Shape:
+        - Input: (B, C, T)
+        - Output: (B, C, T), same shape as the input
+    Parameters:
+        - alpha - trainable parameter that controls frequency
+        - beta - trainable parameter that controls magnitude
+    References:
+        - This activation function is a modified version based on this paper by Liu Ziyin, Tilman Hartwig, Masahito Ueda:
+        https://arxiv.org/abs/2006.08195
+    Examples:
+        >>> a1 = snakebeta(256)
+        >>> x = torch.randn(256)
+        >>> x = a1(x)
+
+    Args:
+        in_features: shape of the input
+        out_features: shape of the output
+        alpha: trainable parameter that controls frequency
+        alpha_trainable: whether alpha is trainable
+        alpha_logscale: whether to use log scale for alpha
+            alpha is initialized to 1 by default, higher values = higher-frequency.
+            beta is initialized to 1 by default, higher values = higher-magnitude.
+            alpha will be trained along with the rest of your model.
+    """
+
+    def __init__(self, in_features, out_features, alpha=1.0, alpha_trainable=True, alpha_logscale=True):
+        super().__init__()
+        self.in_features = out_features if isinstance(out_features, list) else [out_features]
+        self.proj = LoRACompatibleLinear(in_features, out_features)
+
+        # initialize alpha
+        self.alpha_logscale = alpha_logscale
+        if self.alpha_logscale:  # log scale alphas initialized to zeros
+            self.alpha = nn.Parameter(torch.zeros(self.in_features) * alpha)
+            self.beta = nn.Parameter(torch.zeros(self.in_features) * alpha)
+        else:  # linear scale alphas initialized to ones
+            self.alpha = nn.Parameter(torch.ones(self.in_features) * alpha)
+            self.beta = nn.Parameter(torch.ones(self.in_features) * alpha)
+
+        self.alpha.requires_grad = alpha_trainable
+        self.beta.requires_grad = alpha_trainable
+
+        self.no_div_by_zero = 0.000000001
+
+    def forward(self, x):
+        """
+        Forward pass of the function.
+        Applies the function to the input elementwise.
+        SnakeBeta ∶= x + 1/b * sin^2 (xa)
+        """
+        x = self.proj(x)
+        if self.alpha_logscale:
+            alpha = torch.exp(self.alpha)
+            beta = torch.exp(self.beta)
+        else:
+            alpha = self.alpha
+            beta = self.beta
+
+        x = x + (1.0 / (beta + self.no_div_by_zero)) * torch.pow(torch.sin(x * alpha), 2)
+
+        return x
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh")
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim)
+        elif activation_fn == "snakebeta":
+            act_fn = SnakeBeta(dim, inner_dim)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(LoRACompatibleLinear(inner_dim, dim_out))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states):
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",
+        final_dropout: bool = False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+            )
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                # scale_qk=False, # uncomment this to not to use flash attention
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+    ):
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 1. Self-Attention
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=encoder_attention_mask if self.only_cross_attention else attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+
+        # 2. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+            )
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 3. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [self.ff(hid_slice) for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = ff_output + hidden_states
+
+        return hidden_states
+
+
+class SinusoidalPosEmb(torch.nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+        assert self.dim % 2 == 0, "SinusoidalPosEmb requires dim to be even"
+
+    def forward(self, x, scale=1000):
+        if x.ndim < 1:
+            x = x.unsqueeze(0)
+        device = x.device
+        half_dim = self.dim // 2
+        emb = math.log(10000) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, device=device).float() * -emb)
+        emb = scale * x.unsqueeze(1) * emb.unsqueeze(0)
+        emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+        return emb
+
+
+class Block1D(torch.nn.Module):
+    def __init__(self, dim, dim_out, groups=8):
+        super().__init__()
+        self.block = torch.nn.Sequential(
+            torch.nn.Conv1d(dim, dim_out, 3, padding=1),
+            torch.nn.GroupNorm(groups, dim_out),
+            nn.Mish(),
+        )
+
+    def forward(self, x, mask):
+        output = self.block(x * mask)
+        return output * mask
+
+
+class ResnetBlock1D(torch.nn.Module):
+    def __init__(self, dim, dim_out, time_emb_dim, groups=8):
+        super().__init__()
+        self.mlp = torch.nn.Sequential(nn.Mish(), torch.nn.Linear(time_emb_dim, dim_out))
+
+        self.block1 = Block1D(dim, dim_out, groups=groups)
+        self.block2 = Block1D(dim_out, dim_out, groups=groups)
+
+        self.res_conv = torch.nn.Conv1d(dim, dim_out, 1)
+
+    def forward(self, x, mask, time_emb):
+        h = self.block1(x, mask)
+        h += self.mlp(time_emb).unsqueeze(-1)
+        h = self.block2(h, mask)
+        output = h + self.res_conv(x * mask)
+        return output
+
+
+class Downsample1D(nn.Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.conv = torch.nn.Conv1d(dim, dim, 3, 2, 1)
+
+    def forward(self, x):
+        return self.conv(x)
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        time_embed_dim: int,
+        act_fn: str = "silu",
+        out_dim: int = None,
+        post_act_fn: Optional[str] = None,
+        cond_proj_dim=None,
+    ):
+        super().__init__()
+        assert act_fn == "silu", "act_fn must be silu"
+
+        self.linear_1 = nn.Linear(in_channels, time_embed_dim)
+
+        if cond_proj_dim is not None:
+            self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
+        else:
+            self.cond_proj = None
+
+        self.act = nn.SiLU()
+
+        if out_dim is not None:
+            time_embed_dim_out = out_dim
+        else:
+            time_embed_dim_out = time_embed_dim
+        self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out)
+
+        if post_act_fn is None:
+            self.post_act = None
+        else:
+            self.post_act = nn.SiLU()
+
+    def forward(self, sample, condition=None):
+        if condition is not None:
+            sample = sample + self.cond_proj(condition)
+        sample = self.linear_1(sample)
+
+        if self.act is not None:
+            sample = self.act(sample)
+
+        sample = self.linear_2(sample)
+
+        if self.post_act is not None:
+            sample = self.post_act(sample)
+        return sample
+
+
+class Upsample1D(nn.Module):
+    """A 1D upsampling layer with an optional convolution.
+
+    Parameters:
+        channels (`int`):
+            number of channels in the inputs and outputs.
+        use_conv (`bool`, default `False`):
+            option to use a convolution.
+        use_conv_transpose (`bool`, default `False`):
+            option to use a convolution transpose.
+        out_channels (`int`, optional):
+            number of output channels. Defaults to `channels`.
+    """
+
+    def __init__(self, channels, use_conv=False, use_conv_transpose=True, out_channels=None, name="conv"):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        self.conv = None
+        if use_conv_transpose:
+            self.conv = nn.ConvTranspose1d(channels, self.out_channels, 4, 2, 1)
+        elif use_conv:
+            self.conv = nn.Conv1d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, inputs):
+        assert inputs.shape[1] == self.channels
+        if self.use_conv_transpose:
+            return self.conv(inputs)
+
+        outputs = F.interpolate(inputs, scale_factor=2.0, mode="nearest")
+
+        if self.use_conv:
+            outputs = self.conv(outputs)
+
+        return outputs
+
+
+class Transpose(torch.nn.Module):
+    def __init__(self, dim0: int, dim1: int):
+        super().__init__()
+        self.dim0 = dim0
+        self.dim1 = dim1
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = torch.transpose(x, self.dim0, self.dim1)
+        return x
+
+
+class CausalConv1d(torch.nn.Conv1d):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        kernel_size: int,
+        stride: int = 1,
+        dilation: int = 1,
+        groups: int = 1,
+        bias: bool = True,
+        padding_mode: str = 'zeros',
+        device=None,
+        dtype=None
+    ) -> None:
+        super(CausalConv1d, self).__init__(in_channels, out_channels,
+                                           kernel_size, stride,
+                                           padding=0, dilation=dilation,
+                                           groups=groups, bias=bias,
+                                           padding_mode=padding_mode,
+                                           device=device, dtype=dtype)
+        assert stride == 1
+        self.causal_padding = kernel_size - 1
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = F.pad(x, (self.causal_padding, 0), value=0.0)
+        x = super(CausalConv1d, self).forward(x)
+        return x
+
+
+class CausalBlock1D(Block1D):
+    def __init__(self, dim: int, dim_out: int):
+        super(CausalBlock1D, self).__init__(dim, dim_out)
+        self.block = torch.nn.Sequential(
+            CausalConv1d(dim, dim_out, 3),
+            Transpose(1, 2),
+            nn.LayerNorm(dim_out),
+            Transpose(1, 2),
+            nn.Mish(),
+        )
+
+    def forward(self, x: torch.Tensor, mask: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        output = self.block(x * mask)
+        return output * mask
+
+
+class CausalResnetBlock1D(ResnetBlock1D):
+    def __init__(self, dim: int, dim_out: int, time_emb_dim: int, groups: int = 8):
+        super(CausalResnetBlock1D, self).__init__(dim, dim_out, time_emb_dim, groups)
+        self.block1 = CausalBlock1D(dim, dim_out)
+        self.block2 = CausalBlock1D(dim_out, dim_out)
+
+
+class ConditionalDecoder(nn.Module):
+    """
+    This decoder requires an input with the same shape of the target. So, if your text content
+    is shorter or longer than the outputs, please re-sampling it before feeding to the decoder.
+
+    Args:
+        in_channels: number of input channels
+        out_channels: number of output channels
+        channels: tuple of channel dimensions
+        dropout: dropout rate
+        attention_head_dim: dimension of attention heads
+        n_blocks: number of transformer blocks
+        num_mid_blocks: number of middle blocks
+        num_heads: number of attention heads
+        act_fn: activation function name
+    """
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        channels=(256, 256),
+        dropout=0.05,
+        attention_head_dim=64,
+        n_blocks=1,
+        num_mid_blocks=2,
+        num_heads=4,
+        act_fn="snake",
+    ):
+        super().__init__()
+        channels = tuple(channels)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.time_embeddings = SinusoidalPosEmb(in_channels)
+        time_embed_dim = channels[0] * 4
+        self.time_mlp = TimestepEmbedding(
+            in_channels=in_channels,
+            time_embed_dim=time_embed_dim,
+            act_fn="silu",
+        )
+        self.down_blocks = nn.ModuleList([])
+        self.mid_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = in_channels
+        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
+            input_channel = output_channel
+            output_channel = channels[i]
+            is_last = i == len(channels) - 1
+            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            downsample = (
+                Downsample1D(output_channel) if not is_last else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+            self.down_blocks.append(nn.ModuleList([resnet, transformer_blocks, downsample]))
+
+        for _ in range(num_mid_blocks):
+            input_channel = channels[-1]
+            out_channels = channels[-1]
+            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
+
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+
+            self.mid_blocks.append(nn.ModuleList([resnet, transformer_blocks]))
+
+        channels = channels[::-1] + (channels[0],)
+        for i in range(len(channels) - 1):
+            input_channel = channels[i] * 2
+            output_channel = channels[i + 1]
+            is_last = i == len(channels) - 2
+            resnet = ResnetBlock1D(
+                dim=input_channel,
+                dim_out=output_channel,
+                time_emb_dim=time_embed_dim,
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            upsample = (
+                Upsample1D(output_channel, use_conv_transpose=True)
+                if not is_last
+                else nn.Conv1d(output_channel, output_channel, 3, padding=1)
+            )
+            self.up_blocks.append(nn.ModuleList([resnet, transformer_blocks, upsample]))
+        self.final_block = Block1D(channels[-1], channels[-1])
+        self.final_proj = nn.Conv1d(channels[-1], self.out_channels, 1)
+        self.initialize_weights()
+
+    def initialize_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv1d):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.GroupNorm):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
+                if m.bias is not None:
+                    nn.init.constant_(m.bias, 0)
+
+    def forward(self, x, mask, mu, t, spks=None, cond=None, streaming=False):
+        """Forward pass of the UNet1DConditional model.
+
+        Args:
+            x (torch.Tensor): shape (batch_size, in_channels, time)
+            mask (_type_): shape (batch_size, 1, time)
+            t (_type_): shape (batch_size)
+            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
+            cond (_type_, optional): placeholder for future use. Defaults to None.
+
+        Raises:
+            ValueError: _description_
+            ValueError: _description_
+
+        Returns:
+            _type_: _description_
+        """
+
+        t = self.time_embeddings(t).to(t.dtype)
+        t = self.time_mlp(t)
+
+        x = pack([x, mu], "b * t")[0]
+
+        if spks is not None:
+            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
+            x = pack([x, spks], "b * t")[0]
+        if cond is not None:
+            x = pack([x, cond], "b * t")[0]
+
+        hiddens = []
+        masks = [mask]
+        for resnet, transformer_blocks, downsample in self.down_blocks:
+            mask_down = masks[-1]
+            x = resnet(x, mask_down, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            attn_mask = add_optional_chunk_mask(x, mask_down.bool(), False, False, 0, 0, -1).repeat(1, x.size(1), 1)
+            attn_mask = mask_to_bias(attn_mask, x.dtype)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            hiddens.append(x)  # Save hidden states for skip connections
+            x = downsample(x * mask_down)
+            masks.append(mask_down[:, :, ::2])
+        masks = masks[:-1]
+        mask_mid = masks[-1]
+
+        for resnet, transformer_blocks in self.mid_blocks:
+            x = resnet(x, mask_mid, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            attn_mask = add_optional_chunk_mask(x, mask_mid.bool(), False, False, 0, 0, -1).repeat(1, x.size(1), 1)
+            attn_mask = mask_to_bias(attn_mask, x.dtype)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+
+        for resnet, transformer_blocks, upsample in self.up_blocks:
+            mask_up = masks.pop()
+            skip = hiddens.pop()
+            x = pack([x[:, :, :skip.shape[-1]], skip], "b * t")[0]
+            x = resnet(x, mask_up, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            attn_mask = add_optional_chunk_mask(x, mask_up.bool(), False, False, 0, 0, -1).repeat(1, x.size(1), 1)
+            attn_mask = mask_to_bias(attn_mask, x.dtype)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            x = upsample(x * mask_up)
+        x = self.final_block(x, mask_up)
+        output = self.final_proj(x * mask_up)
+        return output * mask
+
+
+class CausalConditionalDecoder(ConditionalDecoder):
+    """
+    This decoder requires an input with the same shape of the target. So, if your text content
+    is shorter or longer than the outputs, please re-sampling it before feeding to the decoder.
+
+    Args:
+        in_channels: number of input channels
+        out_channels: number of output channels
+        channels: list of channel dimensions
+        dropout: dropout rate
+        attention_head_dim: dimension of attention heads
+        n_blocks: number of transformer blocks
+        num_mid_blocks: number of middle blocks
+        num_heads: number of attention heads
+        act_fn: activation function name
+        static_chunk_size: size of static chunks
+        num_decoding_left_chunks: number of left chunks for decoding
+    """
+
+    def __init__(
+        self,
+        in_channels=320,
+        out_channels=80,
+        channels=[256],  # noqa
+        dropout=0.0,
+        attention_head_dim=64,
+        n_blocks=4,
+        num_mid_blocks=12,
+        num_heads=8,
+        act_fn="gelu",
+        static_chunk_size=50,
+        num_decoding_left_chunks=-1,
+    ):
+        torch.nn.Module.__init__(self)
+        channels = tuple(channels)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.time_embeddings = SinusoidalPosEmb(in_channels)
+        time_embed_dim = channels[0] * 4
+        self.time_mlp = TimestepEmbedding(
+            in_channels=in_channels,
+            time_embed_dim=time_embed_dim,
+            act_fn="silu",
+        )
+        self.static_chunk_size = static_chunk_size
+        self.num_decoding_left_chunks = num_decoding_left_chunks
+        self.down_blocks = nn.ModuleList([])
+        self.mid_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        output_channel = in_channels
+        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
+            input_channel = output_channel
+            output_channel = channels[i]
+            is_last = i == len(channels) - 1
+            resnet = CausalResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            downsample = (
+                Downsample1D(output_channel) if not is_last else CausalConv1d(output_channel, output_channel, 3)
+            )
+            self.down_blocks.append(nn.ModuleList([resnet, transformer_blocks, downsample]))
+
+        for _ in range(num_mid_blocks):
+            input_channel = channels[-1]
+            out_channels = channels[-1]
+            resnet = CausalResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
+
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+
+            self.mid_blocks.append(nn.ModuleList([resnet, transformer_blocks]))
+
+        channels = channels[::-1] + (channels[0],)
+        for i in range(len(channels) - 1):
+            input_channel = channels[i] * 2
+            output_channel = channels[i + 1]
+            is_last = i == len(channels) - 2
+            resnet = CausalResnetBlock1D(
+                dim=input_channel,
+                dim_out=output_channel,
+                time_emb_dim=time_embed_dim,
+            )
+            transformer_blocks = nn.ModuleList(
+                [
+                    BasicTransformerBlock(
+                        dim=output_channel,
+                        num_attention_heads=num_heads,
+                        attention_head_dim=attention_head_dim,
+                        dropout=dropout,
+                        activation_fn=act_fn,
+                    )
+                    for _ in range(n_blocks)
+                ]
+            )
+            upsample = (
+                Upsample1D(output_channel, use_conv_transpose=True)
+                if not is_last
+                else CausalConv1d(output_channel, output_channel, 3)
+            )
+            self.up_blocks.append(nn.ModuleList([resnet, transformer_blocks, upsample]))
+        self.final_block = CausalBlock1D(channels[-1], channels[-1])
+        self.final_proj = nn.Conv1d(channels[-1], self.out_channels, 1)
+        self.initialize_weights()
+
+    def forward(self, x, mask, mu, t, spks=None, cond=None, streaming=False):
+        """Forward pass of the UNet1DConditional model.
+
+        Args:
+            x (torch.Tensor): shape (batch_size, in_channels, time)
+            mask (_type_): shape (batch_size, 1, time)
+            t (_type_): shape (batch_size)
+            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
+            cond (_type_, optional): placeholder for future use. Defaults to None.
+
+        Raises:
+            ValueError: _description_
+            ValueError: _description_
+
+        Returns:
+            _type_: _description_
+        """
+        t = self.time_embeddings(t).to(t.dtype)
+        t = self.time_mlp(t)
+
+        x = pack([x, mu], "b * t")[0]
+
+        if spks is not None:
+            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
+            x = pack([x, spks], "b * t")[0]
+        if cond is not None:
+            x = pack([x, cond], "b * t")[0]
+
+        hiddens = []
+        masks = [mask]
+        for resnet, transformer_blocks, downsample in self.down_blocks:
+            mask_down = masks[-1]
+            x = resnet(x, mask_down, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            if streaming is True:
+                attn_mask = add_optional_chunk_mask(x, mask_down.bool(), False, False, 0, self.static_chunk_size, -1)
+            else:
+                attn_mask = add_optional_chunk_mask(x, mask_down.bool(), False, False, 0, 0, -1).repeat(1, x.size(1), 1)
+            attn_mask = mask_to_bias(attn_mask, x.dtype)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            hiddens.append(x)  # Save hidden states for skip connections
+            x = downsample(x * mask_down)
+            masks.append(mask_down[:, :, ::2])
+        masks = masks[:-1]
+        mask_mid = masks[-1]
+
+        for resnet, transformer_blocks in self.mid_blocks:
+            x = resnet(x, mask_mid, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            if streaming is True:
+                attn_mask = add_optional_chunk_mask(x, mask_mid.bool(), False, False, 0, self.static_chunk_size, -1)
+            else:
+                attn_mask = add_optional_chunk_mask(x, mask_mid.bool(), False, False, 0, 0, -1).repeat(1, x.size(1), 1)
+            attn_mask = mask_to_bias(attn_mask, x.dtype)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+
+        for resnet, transformer_blocks, upsample in self.up_blocks:
+            mask_up = masks.pop()
+            skip = hiddens.pop()
+            x = pack([x[:, :, :skip.shape[-1]], skip], "b * t")[0]
+            x = resnet(x, mask_up, t)
+            x = rearrange(x, "b c t -> b t c").contiguous()
+            if streaming is True:
+                attn_mask = add_optional_chunk_mask(x, mask_up.bool(), False, False, 0, self.static_chunk_size, -1)
+            else:
+                attn_mask = add_optional_chunk_mask(x, mask_up.bool(), False, False, 0, 0, -1).repeat(1, x.size(1), 1)
+            attn_mask = mask_to_bias(attn_mask, x.dtype)
+            for transformer_block in transformer_blocks:
+                x = transformer_block(
+                    hidden_states=x,
+                    attention_mask=attn_mask,
+                    timestep=t,
+                )
+            x = rearrange(x, "b t c -> b c t").contiguous()
+            x = upsample(x * mask_up)
+        x = self.final_block(x, mask_up)
+        output = self.final_proj(x * mask_up)
+        return output * mask