`mmagic.models.editors.gca`¶

Package Contents¶

Classes¶

`GCA`	Guided Contextual Attention image matting model.
`GCAModule`	Guided Contextual Attention Module.
`ResGCADecoder`	ResNet decoder with shortcut connection and gca module.
`ResNetDec`	ResNet decoder for image matting.
`ResShortcutDec`	ResNet decoder for image matting with shortcut connection.
`ResGCAEncoder`	ResNet backbone with shortcut connection and gca module.
`ResNetEnc`	ResNet encoder for image matting.
`ResShortcutEnc`	ResNet backbone for image matting with shortcut connection.

class mmagic.models.editors.gca.GCA(data_preprocessor, backbone, loss_alpha=None, init_cfg: Optional[dict] = None, train_cfg=None, test_cfg=None)[source]¶

Bases: mmagic.models.base_models.BaseMattor

Guided Contextual Attention image matting model.

https://arxiv.org/abs/2001.04069

Parameters

data_preprocessor (dict, optional) – The pre-process config of BaseDataPreprocessor.
backbone (dict) – Config of backbone.
loss_alpha (dict) – Config of the alpha prediction loss. Default: None.
init_cfg (dict, optional) – Initialization config dict. Default: None.
train_cfg (dict) – Config of training. In train_cfg, train_backbone should be specified. If the model has a refiner, train_refiner should be specified.
test_cfg (dict) – Config of testing. In test_cfg, If the model has a refiner, train_refiner should be specified.

_forward(inputs)[source]¶

Forward function.

Parameters: inputs (torch.Tensor) – Input tensor.
Returns: Output tensor.
Return type: Tensor

_forward_test(inputs)[source]¶

Forward function for testing GCA model.

Parameters: inputs (torch.Tensor) – batch input tensor.
Returns: Output tensor of model.
Return type: Tensor

_forward_train(inputs, data_samples)[source]¶

Forward function for training GCA model.

Parameters

inputs (torch.Tensor) – batch input tensor collated by data_preprocessor.
data_samples (List[BaseDataElement]) – data samples collated by data_preprocessor.

Returns

Contains the loss items and batch information.

Return type

dict

class mmagic.models.editors.gca.GCAModule(in_channels, out_channels, kernel_size=3, stride=1, rate=2, pad_args=dict(mode='reflect'), interpolation='nearest', penalty=- 10000.0, eps=0.0001)[source]¶

Bases: torch.nn.Module

Guided Contextual Attention Module.

From https://arxiv.org/pdf/2001.04069.pdf. Based on https://github.com/nbei/Deep-Flow-Guided-Video-Inpainting. This module use image feature map to augment the alpha feature map with guided contextual attention score.

Image feature and alpha feature are unfolded to small patches and later used as conv kernel. Thus, we refer the unfolding size as kernel size. Image feature patches have a default kernel size 3 while the kernel size of alpha feature patches could be specified by rate (see rate below). The image feature patches are used to convolve with the image feature itself to calculate the contextual attention. Then the attention feature map is convolved by alpha feature patches to obtain the attention alpha feature. At last, the attention alpha feature is added to the input alpha feature.

Parameters

in_channels (int) – Input channels of the guided contextual attention module.
out_channels (int) – Output channels of the guided contextual attention module.
kernel_size (int) – Kernel size of image feature patches. Default 3.
stride (int) – Stride when unfolding the image feature. Default 1.
rate (int) – The downsample rate of image feature map. The corresponding kernel size and stride of alpha feature patches will be rate x 2 and rate. It could be regarded as the granularity of the gca module. Default: 2.
pad_args (dict) – Parameters of padding when convolve image feature with image feature patches or alpha feature patches. Allowed keys are mode and value. See torch.nn.functional.pad() for more information. Default: dict(mode=’reflect’).
interpolation (str) – Interpolation method in upsampling and downsampling.
penalty (float) – Punishment hyperparameter to avoid a large correlation between each unknown patch and itself. Default: -1e4.
eps (float) – A small number to avoid dividing by 0 when calculating the normed image feature patch. Default: 1e-4.

init_weights()[source]¶: Init weights for the model.

forward(img_feat, alpha_feat, unknown=None, softmax_scale=1.0)[source]¶

Forward function of GCAModule.

Parameters

img_feat (Tensor) – Image feature map of shape (N, ori_c, ori_h, ori_w).
alpha_feat (Tensor) – Alpha feature map of shape (N, alpha_c, ori_h, ori_w).
unknown (Tensor, optional) – Unknown area map generated by trimap. If specified, this tensor should have shape (N, 1, ori_h, ori_w).
softmax_scale (float, optional) – The softmax scale of the attention if unknown area is not provided in forward. Default: 1.

Returns

The augmented alpha feature.

Return type

Tensor

extract_feature_maps_patches(img_feat, alpha_feat, unknown)[source]¶

Extract image feature, alpha feature unknown patches.

Parameters

img_feat (Tensor) – Image feature map of shape (N, img_c, img_h, img_w).
alpha_feat (Tensor) – Alpha feature map of shape (N, alpha_c, ori_h, ori_w).
unknown (Tensor, optional) – Unknown area map generated by trimap of shape (N, 1, img_h, img_w).

Returns

3-tuple of

Tensor: Image feature patches of shape (N, img_h*img_w, img_c, img_ks, img_ks).

Tensor: Guided contextual attention alpha feature map. (N, img_h*img_w, alpha_c, alpha_ks, alpha_ks).

Tensor: Unknown mask of shape (N, img_h*img_w, 1, 1).

Return type

tuple

compute_similarity_map(img_feat, img_ps)[source]¶

Compute similarity between image feature patches.

Parameters

img_feat (Tensor) – Image feature map of shape (1, img_c, img_h, img_w).
img_ps (Tensor) – Image feature patches tensor of shape (1, img_h*img_w, img_c, img_ks, img_ks).

Returns

Similarity map between image feature patches with shape (1, img_h*img_w, img_h, img_w).

Return type

Tensor

compute_guided_attention_score(similarity_map, unknown_ps, scale, self_mask)[source]¶

Compute guided attention score.

Parameters

similarity_map (Tensor) – Similarity map of image feature with shape (1, img_h*img_w, img_h, img_w).
unknown_ps (Tensor) – Unknown area patches tensor of shape (1, img_h*img_w, 1, 1).
scale (Tensor) – Softmax scale of known and unknown area: [unknown_scale, known_scale].
self_mask (Tensor) – Self correlation mask of shape (1, img_h*img_w, img_h, img_w). At (1, i*i, i, i) mask value equals -1e4 for i in [1, img_h*img_w] and other area is all zero.

Returns

Similarity map between image feature patches with shape (1, img_h*img_w, img_h, img_w).

Return type

Tensor

propagate_alpha_feature(gca_score, alpha_ps)[source]¶

Propagate alpha feature based on guided attention score.

Parameters

gca_score (Tensor) – Guided attention score map of shape (1, img_h*img_w, img_h, img_w).
alpha_ps (Tensor) – Alpha feature patches tensor of shape (1, img_h*img_w, alpha_c, alpha_ks, alpha_ks).

Returns

Propagated alpha feature map of shape (1, alpha_c, alpha_h, alpha_w).

Return type

Tensor

process_unknown_mask(unknown, img_feat, softmax_scale)[source]¶

Process unknown mask.

Parameters

unknown (Tensor, optional) – Unknown area map generated by trimap of shape (N, 1, ori_h, ori_w)
img_feat (Tensor) – The interpolated image feature map of shape (N, img_c, img_h, img_w).
softmax_scale (float, optional) – The softmax scale of the attention if unknown area is not provided in forward. Default: 1.

Returns

2-tuple of

Tensor: Interpolated unknown area map of shape (N, img_h*img_w, img_h, img_w).

Tensor: Softmax scale tensor of known and unknown area of shape (N, 2).

Return type

tuple

extract_patches(x, kernel_size, stride)[source]¶

Extract feature patches.

The feature map will be padded automatically to make sure the number of patches is equal to (H / stride) * (W / stride).

Parameters

x (Tensor) – Feature map of shape (N, C, H, W).
kernel_size (int) – Size of each patches.
stride (int) – Stride between patches.

Returns

Extracted patches of shape (N, (H / stride) * (W / stride) , C, kernel_size, kernel_size).

Return type

Tensor

pad(x, kernel_size, stride)[source]¶

Pad input tensor.

Parameters

x (Tensor) – Input tensor.
kernel_size (int) – Kernel size of conv layer.
stride (int) – Stride of conv layer.

Returns

Padded tensor

Return type

Tensor

get_self_correlation_mask(img_feat)[source]¶

Create self correlation mask.

Parameters: img_feat (Tensor) – Input tensor.
Returns: Mask tensor.
Return type: Tensor

static l2_norm(x)[source]¶

L2 normalization function.

Parameters: x (Tensor) – Input tensor.
Returns: L2 normalized output tensor.
Return type: Tensor

class mmagic.models.editors.gca.ResGCADecoder(block, layers, in_channels, kernel_size=3, conv_cfg=None, norm_cfg=dict(type='BN'), act_cfg=dict(type='LeakyReLU', negative_slope=0.2, inplace=True), with_spectral_norm=False, late_downsample=False)[source]¶

Bases: ResShortcutDec

ResNet decoder with shortcut connection and gca module.

feat1 ---------------------------------------- conv2 --- out
                                            |
feat2 ----------------------------------- conv1
                                       |
feat3 ------------------------------ layer4
                                  |
feat4, img_feat -- gca_module - layer3
                |
feat5 ------- layer2
           |
out ---  layer1

gca module also requires unknown tensor generated by trimap which is ignored in the above graph.

Parameters

block (str) – Type of residual block. Currently only BasicBlockDec is implemented.
layers (list[int]) – Number of layers in each block.
in_channels (int) – Channel number of input features.
kernel_size (int) – Kernel size of the conv layers in the decoder.
conv_cfg (dict) – Dictionary to construct convolution layer. If it is None, 2d convolution will be applied. Default: None.
norm_cfg (dict) – Config dict for normalization layer. “BN” by default.
act_cfg (dict) – Config dict for activation layer, “ReLU” by default.
with_spectral_norm (bool) – Whether use spectral norm. Default: False.
late_downsample (bool) – Whether to adopt late downsample strategy, Default: False.

forward(inputs)[source]¶

Forward function of resnet shortcut decoder.

Parameters

inputs (dict) –

Output dictionary of the ResGCAEncoder containing:

out (Tensor): Output of the ResGCAEncoder.
feat1 (Tensor): Shortcut connection from input image.
feat2 (Tensor): Shortcut connection from conv2 of ResGCAEncoder.
feat3 (Tensor): Shortcut connection from layer1 of ResGCAEncoder.
feat4 (Tensor): Shortcut connection from layer2 of ResGCAEncoder.
feat5 (Tensor): Shortcut connection from layer3 of ResGCAEncoder.
img_feat (Tensor): Image feature extracted by guidance head.
unknown (Tensor): Unknown tensor generated by trimap.

Returns

Output tensor.

Return type

Tensor

class mmagic.models.editors.gca.ResNetDec(block, layers, in_channels, kernel_size=3, conv_cfg=None, norm_cfg=dict(type='BN'), act_cfg=dict(type='LeakyReLU', negative_slope=0.2, inplace=True), with_spectral_norm=False, late_downsample=False, init_cfg: Optional[dict] = None)[source]¶

Bases: mmengine.model.BaseModule

ResNet decoder for image matting.

This class is adopted from https://github.com/Yaoyi-Li/GCA-Matting.

Parameters

block (str) – Type of residual block. Currently only BasicBlockDec is implemented.
layers (list[int]) – Number of layers in each block.
in_channels (int) – Channel num of input features.
kernel_size (int) – Kernel size of the conv layers in the decoder.
conv_cfg (dict) – dictionary to construct convolution layer. If it is None, 2d convolution will be applied. Default: None.
norm_cfg (dict) – Config dict for normalization layer. “BN” by default.
act_cfg (dict) – Config dict for activation layer, “ReLU” by default.
with_spectral_norm (bool) – Whether use spectral norm after conv. Default: False.
late_downsample (bool) – Whether to adopt late downsample strategy, Default: False.
init_cfg (dict, optional) – Initialization config dict. Default: None.

init_weights()[source]¶: Init weights for the module.

_make_layer(block, planes, num_blocks, conv_cfg, norm_cfg, act_cfg, with_spectral_norm)[source]¶

forward(x)[source]¶

Forward function.

Parameters: x (Tensor) – Input tensor with shape (N, C, H, W).
Returns: Output tensor.
Return type: Tensor

class mmagic.models.editors.gca.ResShortcutDec(block, layers, in_channels, kernel_size=3, conv_cfg=None, norm_cfg=dict(type='BN'), act_cfg=dict(type='LeakyReLU', negative_slope=0.2, inplace=True), with_spectral_norm=False, late_downsample=False, init_cfg: Optional[dict] = None)[source]¶

Bases: ResNetDec

ResNet decoder for image matting with shortcut connection.

feat1 --------------------------- conv2 --- out
                               |
feat2 ---------------------- conv1
                          |
feat3 ----------------- layer4
                     |
feat4 ------------ layer3
                |
feat5 ------- layer2
           |
out ---  layer1

Parameters

block (str) – Type of residual block. Currently only BasicBlockDec is implemented.
layers (list[int]) – Number of layers in each block.
in_channels (int) – Channel number of input features.
kernel_size (int) – Kernel size of the conv layers in the decoder.
conv_cfg (dict) – Dictionary to construct convolution layer. If it is None, 2d convolution will be applied. Default: None.
norm_cfg (dict) – Config dict for normalization layer. “BN” by default.
act_cfg (dict) – Config dict for activation layer, “ReLU” by default.
late_downsample (bool) – Whether to adopt late downsample strategy, Default: False.

forward(inputs)[source]¶

Forward function of resnet shortcut decoder.

Parameters

inputs (dict) –

Output dictionary of the ResNetEnc containing:

out (Tensor): Output of the ResNetEnc.
feat1 (Tensor): Shortcut connection from input image.
feat2 (Tensor): Shortcut connection from conv2 of ResNetEnc.
feat3 (Tensor): Shortcut connection from layer1 of ResNetEnc.
feat4 (Tensor): Shortcut connection from layer2 of ResNetEnc.
feat5 (Tensor): Shortcut connection from layer3 of ResNetEnc.

Returns

Output tensor.

Return type

Tensor

class mmagic.models.editors.gca.ResGCAEncoder(block, layers, in_channels, conv_cfg=None, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU'), with_spectral_norm=False, late_downsample=False, order=('conv', 'act', 'norm'), init_cfg: Optional[dict] = None)[source]¶

Bases: ResShortcutEnc

ResNet backbone with shortcut connection and gca module.

image ---------------- shortcut[0] -------------- feat1
 |
conv1-conv2 ---------- shortcut[1] -------------- feat2
       |
     conv3-layer1 ---- shortcut[2] -------------- feat3
             |
             | image - guidance_conv ------------ img_feat
             |             |
            layer2 --- gca_module - shortcut[4] - feat4
                            |
                          layer3 -- shortcut[5] - feat5
                             |
                           layer4 --------------- out

gca module also requires unknown tensor generated by trimap which is ignored in the above graph.

Implementation of Natural Image Matting via Guided Contextual Attention https://arxiv.org/pdf/2001.04069.pdf.

Parameters

block (str) – Type of residual block. Currently only BasicBlock is implemented.
layers (list[int]) – Number of layers in each block.
in_channels (int) – Number of input channels.
conv_cfg (dict) – Dictionary to construct convolution layer. If it is None, 2d convolution will be applied. Default: None.
norm_cfg (dict) – Config dict for normalization layer. “BN” by default.
act_cfg (dict) – Config dict for activation layer, “ReLU” by default.
late_downsample (bool) – Whether to adopt late downsample strategy. Default: False.
order (tuple[str]) – Order of conv, norm and act layer in shortcut convolution module. Default: (‘conv’, ‘act’, ‘norm’).
init_cfg (dict, optional) – Initialization config dict. Default: None.

forward(x)[source]¶

Forward function.

Parameters: x (Tensor) – Input tensor with shape (N, C, H, W).
Returns: Contains the output tensor, shortcut feature and intermediate feature.
Return type: dict

class mmagic.models.editors.gca.ResNetEnc(block, layers, in_channels, conv_cfg=None, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU'), with_spectral_norm=False, late_downsample=False, init_cfg: Optional[dict] = None)[source]¶

Bases: mmengine.model.BaseModule

ResNet encoder for image matting.

This class is adopted from https://github.com/Yaoyi-Li/GCA-Matting. Implement and pre-train on ImageNet with the tricks from https://arxiv.org/abs/1812.01187 without the mix-up part.

Parameters

block (str) – Type of residual block. Currently only BasicBlock is implemented.
layers (list[int]) – Number of layers in each block.
in_channels (int) – Number of input channels.
conv_cfg (dict) – dictionary to construct convolution layer. If it is None, 2d convolution will be applied. Default: None.
norm_cfg (dict) – Config dict for normalization layer. “BN” by default.
act_cfg (dict) – Config dict for activation layer, “ReLU” by default.
with_spectral_norm (bool) – Whether use spectral norm after conv. Default: False.
late_downsample (bool) – Whether to adopt late downsample strategy, Default: False.
init_cfg (dict, optional) – Initialization config dict. Default: None.

init_weights()[source]¶: Init weights for the module.

_make_layer(block, planes, num_blocks, stride, conv_cfg, norm_cfg, act_cfg, with_spectral_norm)[source]¶

forward(x)[source]¶

Forward function.

Parameters: x (Tensor) – Input tensor with shape (N, C, H, W).
Returns: Output tensor.
Return type: Tensor

class mmagic.models.editors.gca.ResShortcutEnc(block, layers, in_channels, conv_cfg=None, norm_cfg=dict(type='BN'), act_cfg=dict(type='ReLU'), with_spectral_norm=False, late_downsample=False, order=('conv', 'act', 'norm'), init_cfg: Optional[dict] = None)[source]¶

Bases: ResNetEnc

ResNet backbone for image matting with shortcut connection.

image ---------------- shortcut[0] --- feat1
  |
conv1-conv2 ---------- shortcut[1] --- feat2
       |
      conv3-layer1 --- shortcut[2] --- feat3
              |
             layer2 -- shortcut[4] --- feat4
               |
              layer3 - shortcut[5] --- feat5
                |
               layer4 ---------------- out

Baseline model of Natural Image Matting via Guided Contextual Attention https://arxiv.org/pdf/2001.04069.pdf.

Parameters

block (str) – Type of residual block. Currently only BasicBlock is implemented.
layers (list[int]) – Number of layers in each block.
in_channels (int) – Number of input channels.
conv_cfg (dict) – Dictionary to construct convolution layer. If it is None, 2d convolution will be applied. Default: None.
norm_cfg (dict) – Config dict for normalization layer. “BN” by default.
act_cfg (dict) – Config dict for activation layer, “ReLU” by default.
with_spectral_norm (bool) – Whether use spectral norm after conv. Default: False.
late_downsample (bool) – Whether to adopt late downsample strategy. Default: False.
order (tuple[str]) – Order of conv, norm and act layer in shortcut convolution module. Default: (‘conv’, ‘act’, ‘norm’).
init_cfg (dict, optional) – Initialization config dict. Default: None.

_make_shortcut(in_channels, out_channels, conv_cfg, norm_cfg, act_cfg, order, with_spectral_norm)[source]¶

forward(x)[source]¶

Forward function.

Parameters: x (Tensor) – Input tensor with shape (N, C, H, W).
Returns: Contains the output tensor and shortcut feature.
Return type: dict

mmagic.models.editors.gca¶

Package Contents¶

Classes¶

`mmagic.models.editors.gca`¶