def get_swin_config(swin_name):
config = SwinConfig()
name_split = swin_name.split("_")
model_size = name_split[1]
img_size = int(name_split[4])
window_size = int(name_split[3][-1])
if model_size == "tiny":
embed_dim = 96
depths = (2, 2, 6, 2)
num_heads = (3, 6, 12, 24)
elif model_size == "small":
embed_dim = 96
depths = (2, 2, 18, 2)
num_heads = (3, 6, 12, 24)
elif model_size == "base":
embed_dim = 128
depths = (2, 2, 18, 2)
num_heads = (4, 8, 16, 32)
else:
embed_dim = 192
depths = (2, 2, 18, 2)
num_heads = (6, 12, 24, 48)
if "in22k" in swin_name:
num_classes = 21841
else:
num_classes = 1000
repo_id = "datasets/huggingface/label-files"
filename = "imagenet-1k-id2label.json"
id2label = json.load(
open(cached_download(hf_hub_url(repo_id, filename)), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
config.image_size = img_size
config.num_labels = num_classes
config.embed_dim = embed_dim
config.depths = depths
config.num_heads = num_heads
config.window_size = window_size
return config
def get_config(self):
return MaskFormerConfig.from_backbone_and_decoder_configs(
backbone_config=SwinConfig(depths=[1, 1, 1, 1], ),
decoder_config=DetrConfig(
decoder_ffn_dim=128,
num_queries=self.num_queries,
decoder_attention_heads=2,
d_model=self.mask_feature_size,
),
mask_feature_size=self.mask_feature_size,
fpn_feature_size=self.mask_feature_size,
num_channels=self.num_channels,
num_labels=self.num_labels,
)
def get_config(self):
return SwinConfig(
image_size=self.image_size,
patch_size=self.patch_size,
num_channels=self.num_channels,
embed_dim=self.embed_dim,
depths=self.depths,
num_heads=self.num_heads,
window_size=self.window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=self.qkv_bias,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
drop_path_rate=self.drop_path_rate,
hidden_act=self.hidden_act,
use_absolute_embeddings=self.use_absolute_embeddings,
path_norm=self.patch_norm,
layer_norm_eps=self.layer_norm_eps,
initializer_range=self.initializer_range,
)