mindflow.cell.DiffusionTransformer

class mindflow.cell.DiffusionTransformer(in_channels, out_channels, hidden_channels, layers, heads, time_token_cond=True, compute_dtype=mstype.float32)

Diffusion model with Transformer backbone implementation.

Parameters

• in_channels (int) – The number of input channel.

• out_channels (int) – The number of output channel.

• hidden_channels (int) – The number of hidden channel.

• layers (int) – The number of transformer block layers.

• heads (int) – The number of transformer heads.

• time_token_cond (bool) – Whether to use timestep as condition token. Default: True.

• compute_dtype (mindspore.dtype) – The dtype of compute, it can be mstype.float32 or mstype.float16. Default: mstype.float32, indicates mindspore.float32.

Inputs:

• x (Tensor) - The input has a shape of \((batch\_size, sequence\_len, in\_channels)\).

• timestep (Tensor) - The timestep input has a shape of \((batch\_size,)\).

Outputs:

• output (Tensor) - The output has a shape of \((batch\_size, sequence\_len, out\_channels)\).

Supported Platforms:

Ascend

Examples

>>> from mindspore import ops
>>> from mindflow.cell import DiffusionTransformer
>>> in_channels, out_channels, hidden_channels, layers, heads, batch_size, seq_len = 16, 16, 256, 3, 4, 8, 256
>>> model = DiffusionTransformer(in_channels=in_channels,
...                              out_channels=out_channels,
...                              hidden_channels=hidden_channels,
...                              layers=layers,
...                              heads=heads)
>>> x = ops.rand((batch_size, seq_len, in_channels))
>>> timestep = ops.randint(0, 1000, (batch_size,))
>>> output = model(x, timestep)
>>> print(output.shape)
(8, 256, 16)