search

KaiyangZhou / deep-person-reid

Torchreid: Deep learning person re-identification in PyTorch.
https://kaiyangzhou.github.io/deep-person-reid/
MIT License
4.24k stars 1.14k forks source link

osnet with triplet loss #457

Open machlea opened 3 years ago

machlea commented 3 years ago

loss: name: 'triplet' softmax: label_smooth: True

train: optim: 'amsgrad' lr: 0.0015 max_epoch: 250 batch_size: 64 fixbase_epoch: 0 open_layers: ['classifier'] lr_scheduler: 'cosine'

cfg.loss.triplet.weight_t = 0.5 # weight to balance hard triplet loss
cfg.loss.triplet.weight_x = 1. # weight to balance cross entropy loss

This is my hyperparameters.i even tried loss.triplet.weight_t = 1.0.However, i tried several times, the result cannot be same as paper mAP:83.0%, rank-1:93.6% mAP:83.2%, rank-1:93.4%

maybe there are some important hyperparameters that i have ignored?

shayan-aqabarary commented 2 years ago

loss: name: 'triplet' softmax: label_smooth: True

train: optim: 'amsgrad' lr: 0.0015 max_epoch: 250 batch_size: 64 fixbase_epoch: 0 open_layers: ['classifier'] lr_scheduler: 'cosine'

cfg.loss.triplet.weight_t = 0.5 # weight to balance hard triplet loss
cfg.loss.triplet.weight_x = 1. # weight to balance cross entropy loss

This is my hyperparameters.i even tried loss.triplet.weight_t = 1.0.However, i tried several times, the result cannot be same as paper mAP:83.0%, rank-1:93.6% mAP:83.2%, rank-1:93.4%

maybe there are some important hyperparameters that i have ignored?

I have this problem too, My accuracy for all sections are different with article tables, is it normal? I want optimize this program, do you have any idea for increase the accuracy?