metadata
license: mit
This is the structure of the BLIPNet model. You can load the model with it, or you can create a bigger model for your task.
class BLIPNet(torch.nn.Module): def init(self, ): super().init() #Generation Model self.model = BlipForConditionalGeneration.from_pretrained(MODEL_NAME, cache_dir="model") #Same with https://huggingface.co/uf-aice-lab/BLIP-Math self.ebd_dim = ebd_dim= 443136
#Classification Model
fc_dim = 64 # You can choose a higher number for better performance, for example, 1024.
self.head = nn.Sequential(
nn.Linear(self.ebd_dim, fc_dim),
nn.ReLU(),
)
self.score = nn.Linear(fc_dim, 5) #5 classes
def forward(self, pixel_values, input_ids):
outputs = self.model(input_ids=input_ids, pixel_values=pixel_values, labels=input_ids)
image_text_embeds = self.model.vision_model(pixel_values, return_dict=True).last_hidden_state
image_text_embeds = self.head(image_embeds.view(-1, self.ebd_dim))
#A classification model is based on embeddings from a generative model to leverage BLIP's powerful image-text encoding capabilities.
logits = self.score(image_embeds)
#generated text, probabilities of classification
return outputs, logits
model = BLIPNet() model.load_state_dict(torch.load(best_model_wts_path) ,strict=False)
You need to input the sample in the same way as: https://huggingface.co/uf-aice-lab/BLIP-Math Then you can get the text and score at the same time.