Justifying parameters for reproduction

[herb1999]

Thank you for your help many times！I re-set the seeds, but it seems that the result is still not satisfactory. If it is convenient, can you tell me which specific seeds are suitable? I also used the GTX 3090 for the experiment

Originally posted by @guaguaxyf in #9

The results lower than that in paper may be caused by the parameters different from the reported Implementation Details: "We set 0.5 for both α and β", which is not in code.

You could try modify ViT_Network.py as

    def knowledge_boosting(self, lang_embed, word_embed, query_info, label):
        P_head = query_info['proto'].clone().cuda()
        T = 2.
        lang_logit = F.linear(lang_embed, P_head)
        loss_seman = F.cross_entropy(lang_logit, label)
        
        loss_kd = F.kl_div(F.log_softmax(lang_embed/T,dim=1), F.softmax(word_embed[label]/T,dim=1), reduction='batchmean')
        loss = loss_kd + 0.1*loss_seman
        return 0.5*loss
        # return 0.1*loss

and set

    parser.add_argument('-ED_hp', type=float, default=0.5)

But I wonder why there's still one point lower in acc (seed=1): [81.502, 80.115, 79.478, 76.566, 76.649, 74.65, 74.251, 74.718, 73.821, 73.85, 74.165]
Here's my configs:

python train.py \
    -gpu 7 \
    -project base \
    -dataset cub200 \
    -shot 5 \
    -base_mode ft_dot \
    -new_mode avg_cos \
    -gamma 0.1 \
    -lr_base 2e-4 \
    -lr_new 2e-4 \
    -decay 0.0005  \
    -epochs_base 5  \
    -epochs_new 3  \
    -schedule Cosine  \
    -milestones 20 30 45 \
    -temperature 16  \
    -start_session 0  \
    -batch_size_base 128  \
    -seed 1  \
    -vit  \
    -comp_out 1  \
    -ED  \
    -SKD  \
    -LT  \
    -out PriViLege  \