Retinanet model running error

[bovane]

Hello,I am a student. I have some doubt about running retinanet model . when I use google colab run retinanet model I couldn't understand the parameters meaning，could you explain for me and help me run correctly ？

By the way, I can execute the faster-rcnn model correctly ! Thank you

[bovane]

@delmalih

[delmalih]

Hello, your issue comes from the way you run the command with colab. Try running the command using one line.

Also, make sure to replace all parameters "<...>" with your own. For example : should be replaced by the batch size you want.

[bovane]

thank you ,but I can't find the file such as config-file,could you explain it to me ?

@delmalih

[bovane]

I just download these code and run it with colab , these peramters such as config-file and model.weight I can't find them. So I can't run these command correctly . Could you help me ? I'm stuck for a long time
I‘m a fresh-bird，alway stupied.

Such as for this command, I can't find the config-file ! Are they in your program ?
@delmalih

[delmalih]

If you're trying tu run an inference for he faster r-cnn, for the modeil weights, either you train on your own dataset or you can download some pretrained models on the model zoo : https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/MODEL_ZOO.md

For the config file check here : https://github.com/facebookresearch/maskrcnn-benchmark/tree/master/configs
There is plenty of them, try to create your own file or copy/paste an existing one

[bovane]

Thank you very much ，I will try it again ! Thank you !

[bovane]

With your help I succeed in training and runing an inference,Thank you ☺️ ,but when I am trying to train the retinanet there are some problem.

train.py: error: argument dataset_type: invalid choice: ' --tensorboard-dir' tensorboard-dir peramters and coco dataset error . I guess wehther I need download coco dataset .tar.gz file ? other than the path ? Could you explain it for me ? Thank you !

[bovane]

@delmalih

[delmalih]

For the tensorboard-dir parameter, you need to give any folder to store some files required for the tensorboard (for ex. /tmp/tensorboard)
After the coco parameter, you need to give the path to your dataset (in the coco format !important!). Take a look at the MSCoco dataset to get this format.

[bovane]

Thank you very much for your help, I will keep trying

[bovane]

I'm Sorry for my stupied ! I konw I need to give to my dataset path with coco format ! but I just download the all-mias.tar.gz and use the generate_COCO_annotations.py to create the coco dataset !

then I use the coco path for peramter . I don't understand the dataset should be looklike !

[delmalih]

Ok so normally, you should set your dataset_path as "../mias-db/COCO" and it should work. The generate_COCO_annotations.py script creates a COCO like database with the MIAS dataset.

[bovane]

uh,so sad 😭 😔 😢 🙁 I can't run the retinanet correctly ! for coco dataset path I use absolutely path .

[delmalih]

First, remove all "", we use them only if we want to write commands on multiple lines.
Then, put your paths between quotation marks "/content/drive/...."

[bovane]

I have tried use the quotation marks my coco dataset pathway,but still fail

[delmalih]

Could you send me the entire command you used to run the training process ? Along with its output

[bovane]

ok,I will !
this is your model command line
!python train.py --compute-val-loss \ # Computer val loss or not
--tensorboard-dir
--batch-size
--epochs
coco

then I use :
!python train.py --compute-val-loss \ # Computer val loss or not
--tensorboard-dir output/
--batch-size 20
--epochs 10
coco /content/drive/Shared drives/BoVane/mias-db/COCO

then it will occur the error IndentationError: unexpected indent !!!

when I use the single command line like this:
!python train.py --compute-val-loss \ --tensorboard-dir output/ \ --batch-size 20 \ --epochs 10 \ coco /content/drive/Shared drives/BoVane/mias-db/COCO

Using TensorFlow backend.
2020-04-10 18:49:44.153250: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1
usage: train.py [-h]
[--snapshot SNAPSHOT | --imagenet-weights | --weights WEIGHTS | --no-weights]
[--backbone BACKBONE] [--batch-size BATCH_SIZE] [--gpu GPU]
[--multi-gpu MULTI_GPU] [--multi-gpu-force] [--epochs EPOCHS]
[--steps STEPS] [--lr LR] [--snapshot-path SNAPSHOT_PATH]
[--tensorboard-dir TENSORBOARD_DIR] [--no-snapshots]
[--no-evaluation] [--freeze-backbone] [--random-transform]
[--image-min-side IMAGE_MIN_SIDE]
[--image-max-side IMAGE_MAX_SIDE] [--config CONFIG]
[--weighted-average] [--compute-val-loss] [--multiprocessing]
[--workers WORKERS] [--max-queue-size MAX_QUEUE_SIZE]
{coco,pascal,kitti,oid,csv} ...
train.py: error: argument dataset_type: invalid choice: ' ' (choose from 'coco', 'pascal', 'kitti', 'oid', 'csv')

[delmalih]

Try this command :

!python train.py --compute-val-loss --tensorboard-dir "/tmp/tensorboard-dir" --batch-size 20 --epochs 10 coco "/content/drive/Shared drives/BoVane/mias-db/COCO"

[bovane]

wow 🤩 😯 😲 ! I can run this command Thank you !
but there is another error !
Using TensorFlow backend.
2020-04-10 18:53:19.156283: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1
Traceback (most recent call last):
File "train.py", line 511, in
main()
File "train.py", line 440, in main
keras.backend.tensorflow_backend.set_session(get_session())
File "train.py", line 50, in get_session
config = tf.ConfigProto()
AttributeError: module 'tensorflow' has no attribute 'ConfigProto'

I am just finding the train.py to fix the bug ! Thank you 🙏 😊 ☺️

[delmalih]

Check your tf version and make sure that it is compatible with the code

[bovane]

yep ! I really appreciate your help, now I only have fcos model left ！