Closed daodaofr closed 6 years ago
Thanks for your attention.
In my script, model_weight_file
and ckpt_file
are two different things. The former only contains the state_dict
of the model; the latter also contains optimizer states. My uploaded model_weight_file
is just taken out of my trained ckpt_file
, thus the file size is about one half.
So if you want to test your own trained ckpt_file
, you can use this command:
python script/experiment/train_pcb.py \
-d '(0,)' \
--only_test true \
--dataset DATASET_NAME \
--trainset_part TRAINVAL_OR_TRAIN \
--exp_dir EXPERIMENT_DIRECTORY \
--steps_per_log 20 \
--epochs_per_val 1
That is the same command as your training command, except that only_test
is set to true
, meaning only testing the model without training.
If your problem is not solved, can you provide the full log file that shows the result of mAP 3%, cmc1 11%
? So that I can check if it is the problem I mentioned above.
Thank you for your reply, which perfectly solves my issue.
When I train a model on Market 1501, the training process seems fine. In the training log, I got mAP 70%, cmc1 89% on test dataset.
But when I run your test script and load the my trained model for test, I only get mAP 3%, cmc1 11%.
If I load your model, I can get your results.
Do you have any idea what cause this gap? On issue I notice is that my ckpt.pth model is ~200M, which is twice the size of your model (100M).