MCPcopy Create free account
hub / github.com/MiniMax-AI/VTP / _init_text_components

Method _init_text_components

vtp/models/vtp_hf/modeling_vtp.py:135–180  ·  view source on GitHub ↗

Initialize text encoder and related components.

(self)

Source from the content-addressed store, hash-verified

133 self.pixel_decoder = None
134
135 def _init_text_components(self):
136 """Initialize text encoder and related components."""
137 config = self.config
138
139 act_layer = QuickGELU if config.text_quick_gelu else nn.GELU
140 norm_layer = LayerNorm
141
142 text = TextTransformer(
143 context_length=config.text_context_length,
144 vocab_size=config.text_vocab_size,
145 width=config.text_embed_dim,
146 heads=config.text_num_heads,
147 layers=config.text_depth,
148 mlp_ratio=config.text_mlp_ratio,
149 ls_init_value=config.text_ls_init_value,
150 output_dim=config.text_embed_dim,
151 embed_cls=config.text_embed_cls,
152 no_causal_mask=config.text_no_causal_mask,
153 pad_id=config.text_pad_id,
154 pool_type=config.text_pool_type,
155 proj_type=config.text_proj_type,
156 proj_bias=config.text_proj_bias,
157 output_tokens=config.text_output_tokens,
158 act_layer=act_layer,
159 norm_layer=norm_layer,
160 )
161
162 self.text_transformer = text.transformer
163 self.context_length = text.context_length
164 self.vocab_size = text.vocab_size
165 self.token_embedding = text.token_embedding
166 self.positional_embedding = text.positional_embedding
167 self.ln_final = text.ln_final
168 self.text_projection = text.text_projection
169 self.text_pool_type = text.pool_type
170 self.register_buffer('attn_mask', text.attn_mask, persistent=False)
171
172 # Logit scale
173 init_logit_scale = config.init_logit_scale or np.log(1 / 0.07)
174 lshape = [1] if config.nonscalar_logit_scale else []
175 self.logit_scale = nn.Parameter(torch.ones(lshape) * init_logit_scale)
176
177 if config.init_logit_bias is not None:
178 self.logit_bias = nn.Parameter(torch.ones(lshape) * config.init_logit_bias)
179 else:
180 self.logit_bias = None
181
182 # ==================== Basic Feature Methods ====================
183

Callers 1

__init__Method · 0.95

Calls 1

TextTransformerClass · 0.85

Tested by

no test coverage detected