def __init__(self,
inp_dim,
hidden_dim,
num_layers,
batch_norm=False,
dropout=0.):
super(MLP, self).__init__()
layer_list = OrderedDict()
in_dim = inp_dim
for l in range(num_layers):
layer_list['fc{}'.format(l)] = nn.Linear(in_dim, hidden_dim)
if l < num_layers - 1:
if batch_norm:
layer_list['norm{}'.format(l)] = nn.BatchNorm1D(
num_features=hidden_dim)
layer_list['relu{}'.format(l)] = nn.LeakyReLU()
if dropout > 0:
layer_list['drop{}'.format(l)] = nn.Dropout(p=dropout)
in_dim = hidden_dim
if num_layers > 0:
self.network = nn.Sequential()
for i in layer_list:
self.network.add_sublayer(i, layer_list[i])
else:
self.network = nn.Identity()
def __init__(self,
num_inputs,
input_size,
embedding_size,
width,
depth,
pairs,
stage,
use_bn=True):
super().__init__()
self.stage = stage
self.depth = depth
self.w_embeddings = nn.Embedding(input_size,
1,
weight_attr=xavier_init([input_size]))
self.v_embeddings = nn.Embedding(input_size,
embedding_size,
weight_attr=xavier_init(
[input_size, embedding_size]))
in_features = [num_inputs * embedding_size] + [width] * depth
out_features = [width] * depth + [1]
if use_bn:
self.bn = nn.LayerList([nn.BatchNorm(width) for _ in range(depth)])
else:
self.bn = nn.LayerList([nn.Identity() for _ in range(depth)])
self.linear = nn.LayerList([
nn.Linear(*_, weight_attr=xavier_init([_]))
for _ in zip(in_features, out_features)
])
self.comb_mask = None if stage == 0 else np.load('comb_mask.npy')
pairs = pairs if stage == 0 else sum(self.comb_mask)
self.mask = paddle.create_parameter(
[1, pairs],
'float32',
default_initializer=nn.initializer.Uniform(0.6 - 0.001,
0.6 + 0.001))
self.bn2 = nn.BatchNorm(pairs) if use_bn else nn.Identity()
def __init__(self,
c1,
c2,
k=1,
s=1,
p=None,
g=1,
act=True): # ch_in, ch_out, kernel, stride, padding, groups
super(Conv, self).__init__()
self.conv = nn.Conv2D(c1,
c2,
k,
s,
autopad(k, p),
groups=g,
bias_attr=False)
self.bn = nn.BatchNorm2D(c2)
self.act = nn.LeakyReLU(0.1) if act else nn.Identity()
def __init__(self, path=None, features=256, non_negative=True):
"""Init.
Args:
path (str, optional): Path to saved model. Defaults to None.
features (int, optional): Number of features. Defaults to 256.
backbone (str, optional): Backbone network for encoder. Defaults to resnet50
"""
print("Loading weights: ", path)
super(MidasNet, self).__init__()
use_pretrained = False if path is None else True
self.pretrained, self.scratch = _make_encoder(
backbone="resnext101_wsl",
features=features,
use_pretrained=use_pretrained)
self.scratch.refinenet4 = FeatureFusionBlock(features)
self.scratch.refinenet3 = FeatureFusionBlock(features)
self.scratch.refinenet2 = FeatureFusionBlock(features)
self.scratch.refinenet1 = FeatureFusionBlock(features)
output_conv = [
nn.Conv2D(features, 128, kernel_size=3, stride=1, padding=1),
nn.Upsample(scale_factor=2, mode="bilinear"),
nn.Conv2D(128, 32, kernel_size=3, stride=1, padding=1),
nn.ReLU(),
nn.Conv2D(32, 1, kernel_size=1, stride=1, padding=0),
nn.ReLU() if non_negative else nn.Identity(),
]
if non_negative:
output_conv.append(nn.ReLU())
self.scratch.output_conv = nn.Sequential(*output_conv)
if path:
self.load(path)
def reset_classifier(self, num_classes, global_pool=''):
self.num_classes = num_classes
self.head = nn.Linear(
self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
def __init__(self,
img_size=224,
patch_size=4,
in_chans=3,
class_num=1000,
embed_dim=96,
depths=[2, 2, 6, 2],
num_heads=[3, 6, 12, 24],
window_size=7,
mlp_ratio=4.,
qkv_bias=True,
qk_scale=None,
drop_rate=0.,
attn_drop_rate=0.,
drop_path_rate=0.1,
norm_layer=nn.LayerNorm,
ape=False,
patch_norm=True,
use_checkpoint=False,
**kwargs):
super(SwinTransformer, self).__init__()
self.num_classes = num_classes = class_num
self.num_layers = len(depths)
self.embed_dim = embed_dim
self.ape = ape
self.patch_norm = patch_norm
self.num_features = int(embed_dim * 2**(self.num_layers - 1))
self.mlp_ratio = mlp_ratio
# split image into non-overlapping patches
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
norm_layer=norm_layer if self.patch_norm else None)
num_patches = self.patch_embed.num_patches
patches_resolution = self.patch_embed.patches_resolution
self.patches_resolution = patches_resolution
# absolute position embedding
if self.ape:
self.absolute_pos_embed = self.create_parameter(
shape=(1, num_patches, embed_dim), default_initializer=zeros_)
self.add_parameter("absolute_pos_embed", self.absolute_pos_embed)
trunc_normal_(self.absolute_pos_embed)
self.pos_drop = nn.Dropout(p=drop_rate)
# stochastic depth
dpr = np.linspace(0, drop_path_rate,
sum(depths)).tolist() # stochastic depth decay rule
# build layers
self.layers = nn.LayerList()
for i_layer in range(self.num_layers):
layer = BasicLayer(
dim=int(embed_dim * 2**i_layer),
input_resolution=(patches_resolution[0] // (2**i_layer),
patches_resolution[1] // (2**i_layer)),
depth=depths[i_layer],
num_heads=num_heads[i_layer],
window_size=window_size,
mlp_ratio=self.mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
norm_layer=norm_layer,
downsample=PatchMerging if
(i_layer < self.num_layers - 1) else None,
use_checkpoint=use_checkpoint)
self.layers.append(layer)
self.norm = norm_layer(self.num_features)
self.avgpool = nn.AdaptiveAvgPool1D(1)
self.head = nn.Linear(
self.num_features,
num_classes) if self.num_classes > 0 else nn.Identity()
self.apply(self._init_weights)
