MCPcopy Create free account
hub / github.com/openai/point-e / __init__

Method __init__

point_e/models/transformer.py:156–193  ·  view source on GitHub ↗
(
        self,
        *,
        device: torch.device,
        dtype: torch.dtype,
        input_channels: int = 3,
        output_channels: int = 3,
        n_ctx: int = 1024,
        width: int = 512,
        layers: int = 12,
        heads: int = 8,
        init_scale: float = 0.25,
        time_token_cond: bool = False,
    )

Source from the content-addressed store, hash-verified

154
155class PointDiffusionTransformer(nn.Module):
156 def __init__(
157 self,
158 *,
159 device: torch.device,
160 dtype: torch.dtype,
161 input_channels: int = 3,
162 output_channels: int = 3,
163 n_ctx: int = 1024,
164 width: int = 512,
165 layers: int = 12,
166 heads: int = 8,
167 init_scale: float = 0.25,
168 time_token_cond: bool = False,
169 ):
170 super().__init__()
171 self.input_channels = input_channels
172 self.output_channels = output_channels
173 self.n_ctx = n_ctx
174 self.time_token_cond = time_token_cond
175 self.time_embed = MLP(
176 device=device, dtype=dtype, width=width, init_scale=init_scale * math.sqrt(1.0 / width)
177 )
178 self.ln_pre = nn.LayerNorm(width, device=device, dtype=dtype)
179 self.backbone = Transformer(
180 device=device,
181 dtype=dtype,
182 n_ctx=n_ctx + int(time_token_cond),
183 width=width,
184 layers=layers,
185 heads=heads,
186 init_scale=init_scale,
187 )
188 self.ln_post = nn.LayerNorm(width, device=device, dtype=dtype)
189 self.input_proj = nn.Linear(input_channels, width, device=device, dtype=dtype)
190 self.output_proj = nn.Linear(width, output_channels, device=device, dtype=dtype)
191 with torch.no_grad():
192 self.output_proj.weight.zero_()
193 self.output_proj.bias.zero_()
194
195 def forward(self, x: torch.Tensor, t: torch.Tensor):
196 """

Callers

nothing calls this directly

Calls 3

MLPClass · 0.85
TransformerClass · 0.85
__init__Method · 0.45

Tested by

no test coverage detected