MCPcopy Create free account
hub / github.com/ChenWu98/agent-attack / load

Method load

agent_attack/models/instructblip.py:76–92  ·  view source on GitHub ↗

Loads model and processors (InstructBLIPProcessor contains Vicuna Tokenizer, Q-Former Tokenizer, and an ImageProcessor) using the HF `InstructBLIP*.from_pretrained()` functionality.

(self)

Source from the content-addressed store, hash-verified

74 self.string2indices[trigger_string] = token_idx_list
75
76 def load(self) -> Tuple[nn.Module, Tokenizer, ImageProcessor]:
77 """
78 Loads model and processors (InstructBLIPProcessor contains Vicuna Tokenizer, Q-Former Tokenizer, and an
79 ImageProcessor) using the HF `InstructBLIP*.from_pretrained()` functionality.
80 """
81 with self.distributed_state.main_process_first():
82 text_img_processor = InstructBlipProcessor.from_pretrained(self.hub_path)
83 model = InstructBlipForConditionalGeneration.from_pretrained(self.hub_path)
84
85 # Lift `image_processor` for use in evaluation harnesses
86 image_processor = text_img_processor.image_processor
87
88 # Place Model on Device
89 model = model.to(self.distributed_state.device, dtype=self.dtype)
90 model.eval()
91
92 return model, text_img_processor, image_processor
93
94 def freeze(self) -> None:
95 self.model.eval()

Callers 1

__init__Method · 0.95

Calls 1

toMethod · 0.80

Tested by

no test coverage detected