Keras-retinanet: How to load snapshots?

[mattignal]

Hi there,

I've tried to load my existing snapshots, but each time it says 'Creating Model..' instead of Loading Model. What do I input on the command line to load my snapshot?

Thanks!

[hgaiser]

Let's start with what the command is that you're executing?

[mattignal]

I've tried a few things, such as specifying path, but generally (working from memory here):

keras_retinanet/bin/train.py --snapshot resnet50_data_04.h5 csv annotations.csv classes.csv

[hgaiser]

Try printing the value of args.snapshot on this line.

It seems to work for me at least.

[JaosonMa]

i use this too ,
CUDA_VISIBLE_DEVICES=3 keras_retinanet/bin/train.py \
--snapshot-path /home/jasonma/dev_sdb/models/keras-retinanet/20180920-detect \
--tensorboard-dir /home/jasonma/dev_sdb/models/keras-retinanet/20180920-detect \
--snapshot /home/jasonma/dev_sdb/models/keras-retinanet/20180920-detect/resnet50_pascal_05.h5 \
--batch-size=2 \
--gpu=3 \
pascal /home/jasonma/dev_sdb/data/keras-retinanet-data/

resnet50_pascal_05.h5
but it start
Epoch 1/50
94/10000 [..............................
was not continue from epoch 05,
how should i set it ?
@mattignal @hgaiser

[hgaiser]

It will start from epoch 1, but with the state of your snapshot.

On Fri, 21 Sep 2018, 07:20 JaosonMa, [e-mail] wrote:

i use this too ,
CUDA_VISIBLE_DEVICES=3 keras_retinanet/bin/train.py
--snapshot-path
/home/jasonma/dev_sdb/models/keras-retinanet/20180920-detect
--tensorboard-dir
/home/jasonma/dev_sdb/models/keras-retinanet/20180920-detect
--snapshot
/home/jasonma/dev_sdb/models/keras-retinanet/20180920-detect/resnet50_pascal_05.h5

--batch-size=2
--gpu=3
pascal /home/jasonma/dev_sdb/data/keras-retinanet-data/

resnet50_pascal_05.h5
but it start
Epoch 1/50
94/10000 [..............................
was not continue from epoch 05,
how should i set it ?
@mattignal https://github.com/mattignal @hgaiser
https://github.com/hgaiser

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
https://github.com/fizyr/keras-retinanet/issues/269#issuecomment-423416470,
or mute the thread
https://github.com/notifications/unsubscribe-auth/AArtapwrdAtR3e1PvsGc0PLlz3efszU0ks5udHcngaJpZM4R56S6
.

[lakshmipriya04]

Incase my training gets stopped at some particular epoch (say 5). Is there any way to restart the training from same epoch 5 ?
TIA

[hgaiser]

You can use the --snapshot argument.

[rawat123]

@hgaiser so do we need to use --snapshot with --weights

[hgaiser]

@hgaiser so do we need to use --snapshot with --weights

Nope you use one or the other.