(self,
embed_dims,
num_heads,
window_size,
qkv_bias=True,
qk_scale=None,
attn_drop_rate=0.,
proj_drop_rate=0.,
init_cfg=None)
| 37 | """ |
| 38 | |
| 39 | def __init__(self, |
| 40 | embed_dims, |
| 41 | num_heads, |
| 42 | window_size, |
| 43 | qkv_bias=True, |
| 44 | qk_scale=None, |
| 45 | attn_drop_rate=0., |
| 46 | proj_drop_rate=0., |
| 47 | init_cfg=None): |
| 48 | |
| 49 | super().__init__(init_cfg=init_cfg) |
| 50 | self.embed_dims = embed_dims |
| 51 | self.window_size = window_size # Wh, Ww |
| 52 | self.num_heads = num_heads |
| 53 | head_embed_dims = embed_dims // num_heads |
| 54 | self.scale = qk_scale or head_embed_dims**-0.5 |
| 55 | |
| 56 | # define a parameter table of relative position bias |
| 57 | self.relative_position_bias_table = nn.Parameter( |
| 58 | torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), |
| 59 | num_heads)) # 2*Wh-1 * 2*Ww-1, nH |
| 60 | |
| 61 | # About 2x faster than original impl |
| 62 | Wh, Ww = self.window_size |
| 63 | rel_index_coords = self.double_step_seq(2 * Ww - 1, Wh, 1, Ww) |
| 64 | rel_position_index = rel_index_coords + rel_index_coords.T |
| 65 | rel_position_index = rel_position_index.flip(1).contiguous() |
| 66 | self.register_buffer('relative_position_index', rel_position_index) |
| 67 | |
| 68 | self.qkv = nn.Linear(embed_dims, embed_dims * 3, bias=qkv_bias) |
| 69 | self.attn_drop = nn.Dropout(attn_drop_rate) |
| 70 | self.proj = nn.Linear(embed_dims, embed_dims) |
| 71 | self.proj_drop = nn.Dropout(proj_drop_rate) |
| 72 | |
| 73 | self.softmax = nn.Softmax(dim=-1) |
| 74 | |
| 75 | def init_weights(self): |
| 76 | trunc_normal_init(self.relative_position_bias_table, std=0.02) |
no test coverage detected