MCPcopy Index your code
hub / github.com/ActiveVisionLab/DFNet / Embedder

Class Embedder

script/models/nerfw.py:98–163  ·  view source on GitHub ↗

Source from the content-addressed store, hash-verified

96
97# Positional encoding (section 5.1 of NERF)
98class Embedder:
99 def __init__(self, **kwargs):
100 self.kwargs = kwargs
101 self.N_freqs = 0
102 self.N = -1 # epoch to max frequency, for Nerfie embedding only
103 self.create_embedding_fn()
104
105 def create_embedding_fn(self):
106 embed_fns = []
107 d = self.kwargs['input_dims']
108 out_dim = 0
109 if self.kwargs['include_input']:
110 embed_fns.append(lambda x : x)
111 out_dim += d
112
113 max_freq = self.kwargs['max_freq_log2']
114 self.N_freqs = self.kwargs['num_freqs']
115
116 if self.kwargs['log_sampling']:
117 freq_bands = 2.**torch.linspace(0., max_freq, steps=self.N_freqs) # tensor([ 1., 2., 4., 8., 16., 32., 64., 128., 256., 512.])
118 else:
119 freq_bands = torch.linspace(2.**0., 2.**max_freq, steps=self.N_freqs)
120
121 for freq in freq_bands: # 10 iters for 3D location, 4 iters for 2D direction
122 for p_fn in self.kwargs['periodic_fns']:
123 embed_fns.append(lambda x, p_fn=p_fn, freq=freq : p_fn(x * freq))
124 out_dim += d
125 self.embed_fns = embed_fns
126 self.out_dim = out_dim
127
128 def embed(self, inputs):
129 if self.kwargs['max_freq_log2'] != 0:
130 ret = torch.cat([fn(inputs) for fn in self.embed_fns], -1) # cos, sin embedding
131 else:
132 ret = inputs
133 return ret
134
135 def get_embed_weight(self, epoch, num_freqs, N):
136 ''' Nerfie Paper Eq.(8) '''
137 alpha = num_freqs * epoch / N
138 W_j = []
139 for i in range(num_freqs):
140 tmp = torch.clamp(torch.Tensor([alpha - i]), 0, 1)
141 tmp2 = (1 - torch.cos(torch.Tensor([np.pi]) * tmp)) / 2
142 W_j.append(tmp2)
143 return W_j
144
145 def embed_DNeRF(self, inputs, epoch):
146 ''' Nerfie paper section 3.5 Coarse-to-Fine Deformation Regularization '''
147 # get weight for each frequency band j
148 W_j = self.get_embed_weight(epoch, self.N_freqs, self.N) # W_j: [W_0, W_1, W_2, ..., W_{m-1}]
149
150 # Fourier embedding
151 out = []
152 for fn in self.embed_fns: # 17, embed_fns:[input, cos, sin, cos, sin, ..., cos, sin]
153 out.append(fn(inputs))
154
155 # apply weighted positional encoding, only to cos&sins

Callers 1

get_embedderFunction · 0.70

Calls

no outgoing calls

Tested by

no test coverage detected