# SwinIR
### *class* deepinv.models.SwinIR(img_size=128, patch_size=1, in_chans=3, embed_dim=180, depths=(6, 6, 6, 6, 6, 6), num_heads=(6, 6, 6, 6, 6, 6), window_size=8, mlp_ratio=2, qkv_bias=True, qk_scale=None, drop_rate=0.0, attn_drop_rate=0.0, drop_path_rate=0.1, norm_layer=nn.LayerNorm, ape=False, patch_norm=True, use_checkpoint=False, upscale=1, img_range=1.0, upsampler='', resi_connection='1conv', pretrained='download', pretrained_noise_level=15, \*\*kwargs)
Bases: [`Denoiser`](https://deepinv.org/api/stubs/deepinv.models.Denoiser.html.md#deepinv.models.Denoiser)
SwinIR denoising network.
The Swin Image Restoration (SwinIR) denoising network was introduced by Liang *et al.*[1](#footcite-liang2021swinir). This code is adapted from the official implementation by the
authors.
* **Parameters:**
* **img_size** ([*int*](https://docs.python.org/3.12/library/functions.html#int) *,* [*tuple*](https://docs.python.org/3.12/library/stdtypes.html#tuple)) – Input image size. Default 128.
* **patch_size** ([*int*](https://docs.python.org/3.12/library/functions.html#int) *,* [*tuple*](https://docs.python.org/3.12/library/stdtypes.html#tuple)) – Patch size. Default: 1.
* **in_chans** ([*int*](https://docs.python.org/3.12/library/functions.html#int)) – Number of input image channels. Default: 3.
* **embed_dim** ([*int*](https://docs.python.org/3.12/library/functions.html#int)) – Patch embedding dimension. Default: 180.
* **depths** (*Sequence*) – Depth of each Swin Transformer layer.
* **num_heads** (*Sequence*) – Number of attention heads in different layers.
* **window_size** ([*int*](https://docs.python.org/3.12/library/functions.html#int)) – Window size. Default: 8.
* **mlp_ratio** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – Ratio of mlp hidden dim to embedding dim. Default: 2.
* **qkv_bias** ([*bool*](https://docs.python.org/3.12/library/functions.html#bool)) – If True, add a learnable bias to query, key, value. Default: True.
* **qk_scale** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – Override default qk scale of head_dim \*\* -0.5 if set. Default: None.
* **drop_rate** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – Dropout rate. Default: 0.
* **attn_drop_rate** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – Attention dropout rate. Default: 0.
* **drop_path_rate** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – Stochastic depth rate. Default: 0.1.
* **norm_layer** ([*torch.nn.Module*](https://docs.pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module)) – Normalization layer. Default: nn.LayerNorm.
* **ape** ([*bool*](https://docs.python.org/3.12/library/functions.html#bool)) – If True, add absolute position embedding to the patch embedding. Default: False.
* **patch_norm** ([*bool*](https://docs.python.org/3.12/library/functions.html#bool)) – If True, add normalization after patch embedding. Default: True.
* **use_checkpoint** ([*bool*](https://docs.python.org/3.12/library/functions.html#bool)) – Whether to use checkpointing to save memory. Default: False.
* **upscale** ([*int*](https://docs.python.org/3.12/library/functions.html#int)) – Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
* **img_range** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – Image range. 1. or 255. Default: 1.
* **upsampler** ([*str*](https://docs.python.org/3.12/library/stdtypes.html#str) *,* *None*) – The reconstruction module. ‘’/’pixelshuffle’/’pixelshuffledirect’/’nearest+conv’/None.
Default: ‘’.
* **resi_connection** ([*str*](https://docs.python.org/3.12/library/stdtypes.html#str)) – The convolutional block before residual connection. Should be either ‘1conv’ or ‘3conv’.
Default: ‘1conv’.
* **pretrained** ([*str*](https://docs.python.org/3.12/library/stdtypes.html#str) *,* *None*) –
Use a pretrained network. If `pretrained=None`, the weights will be initialized at
random using PyTorch’s default initialization. If `pretrained='download'`, the weights will be downloaded from
the authors’ online repository [https://github.com/JingyunLiang/SwinIR/releases/tag/v0.0](https://github.com/JingyunLiang/SwinIR/releases/tag/v0.0). Downloading is only
available for the following three architectures:
- **Denoising net**: `upscale=1`, `in_chans` 1 or 3
- **Lightweight super-resolution net**: `upscale=2`, `in_chans=3`, `img_size=64`, `embed_dim=60`,
`depths=(6, 6, 6, 6)`, `num_heads=(6, 6, 6, 6)`, `upsampler='pixelshuffledirect'`
- **Classical super-resolution net**: `upscale=2`, `in_chans=3`, `img_size=64`, `embed_dim=180`,
`depths=(6, 6, 6, 6, 6, 6)`, `num_heads=(6, 6, 6, 6, 6, 6)`, `upsampler='pixelshuffle'`
Finally, `pretrained` can also be set as a path to the user’s own pretrained weights. Default: ‘download’.
See [pretrained-weights](https://deepinv.org/user_guide/reconstruction/pretrained-models.html.md#pretrained-weights) for more details.
* **pretrained_noise_level** ([*int*](https://docs.python.org/3.12/library/functions.html#int)) – The noise level of the pretrained denoising model to be downloaded (in 0-255
scale). Must be one of {15, 25, 50}.
Only used when `upscale=1`. Default: 15.
* **References:**
* **[1]** Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: image restoration using swin transformer. In *Proceedings of the IEEE/CVF international conference on computer vision*, 1833–1844. 2021.
#### forward(x, sigma=None, \*\*kwargs)
Run the denoiser on noisy image. The noise level is not used in this denoiser.
* **Parameters:**
* **x** ([*torch.Tensor*](https://docs.pytorch.org/docs/stable/tensors.html#torch.Tensor)) – noisy image, of shape B, C, W, H.
* **sigma** ([*float*](https://docs.python.org/3.12/library/functions.html#float)) – noise level (not used).
## Examples using `SwinIR`: