Skip to content

deeplabcut.pose_estimation_pytorch.models.modules.gated_attention_unit

Gated Attention Unit.

Based on the building blocks used for the mmdetection CSPNeXt implementation. For more information, see https://github.com/open-mmlab/mmdetection.

Classes:

Name Description
GatedAttentionUnit

Gated Attention Unit (GAU) in RTMBlock.
Scale

Scale vector by element multiplications.

Functions:

Name Description
rope

Applies Rotary Position Embedding to input tensor.

GatedAttentionUnit

Bases: Module

Gated Attention Unit (GAU) in RTMBlock.

Methods:

Name Description
rel_pos_bias

Add relative position bias.
Source code in deeplabcut/pose_estimation_pytorch/models/modules/gated_attention_unit.py 
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
class GatedAttentionUnit(nn.Module):
    """Gated Attention Unit (GAU) in RTMBlock."""

    def __init__(
        self,
        num_token,
        in_token_dims,
        out_token_dims,
        expansion_factor=2,
        s=128,
        eps=1e-5,
        dropout_rate=0.0,
        drop_path=0.0,
        attn_type="self-attn",
        act_fn="SiLU",
        bias=False,
        use_rel_bias=True,
        pos_enc=False,
    ):
        super().__init__()
        self.s = s
        self.num_token = num_token
        self.use_rel_bias = use_rel_bias
        self.attn_type = attn_type
        self.pos_enc = pos_enc

        if drop_path > 0.0:
            self.drop_path = timm_layers.DropPath(drop_path)
        else:
            self.drop_path = nn.Identity()

        self.e = int(in_token_dims * expansion_factor)
        if use_rel_bias:
            if attn_type == "self-attn":
                self.w = nn.Parameter(torch.rand([2 * num_token - 1], dtype=torch.float))
            else:
                self.a = nn.Parameter(torch.rand([1, s], dtype=torch.float))
                self.b = nn.Parameter(torch.rand([1, s], dtype=torch.float))
        self.o = nn.Linear(self.e, out_token_dims, bias=bias)

        if attn_type == "self-attn":
            self.uv = nn.Linear(in_token_dims, 2 * self.e + self.s, bias=bias)
            self.gamma = nn.Parameter(torch.rand((2, self.s)))
            self.beta = nn.Parameter(torch.rand((2, self.s)))
        else:
            self.uv = nn.Linear(in_token_dims, self.e + self.s, bias=bias)
            self.k_fc = nn.Linear(in_token_dims, self.s, bias=bias)
            self.v_fc = nn.Linear(in_token_dims, self.e, bias=bias)
            nn.init.xavier_uniform_(self.k_fc.weight)
            nn.init.xavier_uniform_(self.v_fc.weight)

        self.ln = ScaleNorm(in_token_dims, eps=eps)

        nn.init.xavier_uniform_(self.uv.weight)

        if act_fn == "SiLU" or act_fn == nn.SiLU:
            self.act_fn = nn.SiLU(True)
        elif act_fn == "ReLU" or act_fn == nn.ReLU:
            self.act_fn = nn.ReLU(True)
        else:
            raise NotImplementedError

        if in_token_dims == out_token_dims:
            self.shortcut = True
            self.res_scale = Scale(in_token_dims)
        else:
            self.shortcut = False

        self.sqrt_s = math.sqrt(s)

        self.dropout_rate = dropout_rate

        if dropout_rate > 0.0:
            self.dropout = nn.Dropout(dropout_rate)

    def rel_pos_bias(self, seq_len, k_len=None):
        """Add relative position bias."""

        if self.attn_type == "self-attn":
            t = F.pad(self.w[: 2 * seq_len - 1], [0, seq_len]).repeat(seq_len)
            t = t[..., :-seq_len].reshape(-1, seq_len, 3 * seq_len - 2)
            r = (2 * seq_len - 1) // 2
            t = t[..., r:-r]
        else:
            a = rope(self.a.repeat(seq_len, 1), dim=0)
            b = rope(self.b.repeat(k_len, 1), dim=0)
            t = torch.bmm(a, b.permute(0, 2, 1))
        return t

    def _forward(self, inputs):
        """GAU Forward function."""

        if self.attn_type == "self-attn":
            x = inputs
        else:
            x, k, v = inputs

        x = self.ln(x)

        # [B, K, in_token_dims] -> [B, K, e + e + s]
        uv = self.uv(x)
        uv = self.act_fn(uv)

        if self.attn_type == "self-attn":
            # [B, K, e + e + s] -> [B, K, e], [B, K, e], [B, K, s]
            u, v, base = torch.split(uv, [self.e, self.e, self.s], dim=2)
            # [B, K, 1, s] * [1, 1, 2, s] + [2, s] -> [B, K, 2, s]
            base = base.unsqueeze(2) * self.gamma[None, None, :] + self.beta

            if self.pos_enc:
                base = rope(base, dim=1)
            # [B, K, 2, s] -> [B, K, s], [B, K, s]
            q, k = torch.unbind(base, dim=2)

        else:
            # [B, K, e + s] -> [B, K, e], [B, K, s]
            u, q = torch.split(uv, [self.e, self.s], dim=2)

            k = self.k_fc(k)  # -> [B, K, s]
            v = self.v_fc(v)  # -> [B, K, e]

            if self.pos_enc:
                q = rope(q, 1)
                k = rope(k, 1)

        # [B, K, s].permute() -> [B, s, K]
        # [B, K, s] x [B, s, K] -> [B, K, K]
        qk = torch.bmm(q, k.permute(0, 2, 1))

        if self.use_rel_bias:
            if self.attn_type == "self-attn":
                bias = self.rel_pos_bias(q.size(1))
            else:
                bias = self.rel_pos_bias(q.size(1), k.size(1))
            qk += bias[:, : q.size(1), : k.size(1)]
        # [B, K, K]
        kernel = torch.square(F.relu(qk / self.sqrt_s))

        if self.dropout_rate > 0.0:
            kernel = self.dropout(kernel)
        # [B, K, K] x [B, K, e] -> [B, K, e]
        x = u * torch.bmm(kernel, v)

        # [B, K, e] -> [B, K, out_token_dims]
        x = self.o(x)

        return x

    def forward(self, x):
        if self.shortcut:
            if self.attn_type == "cross-attn":
                res_shortcut = x[0]
            else:
                res_shortcut = x
            main_branch = self.drop_path(self._forward(x))
            return self.res_scale(res_shortcut) + main_branch
        else:
            return self.drop_path(self._forward(x))

rel_pos_bias 

rel_pos_bias(seq_len, k_len=None)

Add relative position bias.

Source code in deeplabcut/pose_estimation_pytorch/models/modules/gated_attention_unit.py 
150
151
152
153
154
155
156
157
158
159
160
161
162
def rel_pos_bias(self, seq_len, k_len=None):
    """Add relative position bias."""

    if self.attn_type == "self-attn":
        t = F.pad(self.w[: 2 * seq_len - 1], [0, seq_len]).repeat(seq_len)
        t = t[..., :-seq_len].reshape(-1, seq_len, 3 * seq_len - 2)
        r = (2 * seq_len - 1) // 2
        t = t[..., r:-r]
    else:
        a = rope(self.a.repeat(seq_len, 1), dim=0)
        b = rope(self.b.repeat(k_len, 1), dim=0)
        t = torch.bmm(a, b.permute(0, 2, 1))
    return t

Scale

Bases: Module

Scale vector by element multiplications.

Parameters:

Name Type Description Default

dim

The dimension of the scale vector.

 required

init_value

The initial value of the scale vector.

 1.0

trainable

Whether the scale vector is trainable.

 True
Source code in deeplabcut/pose_estimation_pytorch/models/modules/gated_attention_unit.py 
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
class Scale(nn.Module):
    """Scale vector by element multiplications.

    Args:
        dim: The dimension of the scale vector.
        init_value: The initial value of the scale vector.
        trainable: Whether the scale vector is trainable.
    """

    def __init__(self, dim, init_value=1.0, trainable=True):
        super().__init__()
        self.scale = nn.Parameter(init_value * torch.ones(dim), requires_grad=trainable)

    def forward(self, x):
        return x * self.scale

rope 

rope(x, dim)

Applies Rotary Position Embedding to input tensor.

Source code in deeplabcut/pose_estimation_pytorch/models/modules/gated_attention_unit.py 
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
def rope(x, dim):
    """Applies Rotary Position Embedding to input tensor."""
    shape = x.shape
    if isinstance(dim, int):
        dim = [dim]

    spatial_shape = [shape[i] for i in dim]
    total_len = 1
    for i in spatial_shape:
        total_len *= i

    position = torch.reshape(torch.arange(total_len, dtype=torch.int, device=x.device), spatial_shape)

    for i in range(dim[-1] + 1, len(shape) - 1, 1):
        position = torch.unsqueeze(position, dim=-1)

    half_size = shape[-1] // 2
    freq_seq = -torch.arange(half_size, dtype=torch.int, device=x.device) / float(half_size)
    inv_freq = 10000**-freq_seq

    sinusoid = position[..., None] * inv_freq[None, None, :]

    sin = torch.sin(sinusoid)
    cos = torch.cos(sinusoid)
    x1, x2 = torch.chunk(x, 2, dim=-1)

    return torch.cat([x1 * cos - x2 * sin, x2 * cos + x1 * sin], dim=-1)