Training Implementation Issue.

[chrisakatibs]

Hello,

Thanks for sharing your work! I am trying to implement your code with Lyft's Autonomous Vehicle Dataset on Google Colab.

Unfortunately I am having trouble getting it into operation. All my values regress to -0.0000.

Now I will say I did not crop the data. This is obviously a glaring error in my procedure but I had wondered of the purpose of the data cropping and was reluctant to perform such an timely operation on data with cloud services(Google Drive, 9,996 training samples).

Is this most likely my issue? or do you believe that running crop will not fix this error?

Training for cars, bicycles, motorcycles, trucks, busses, pedestrians, other_vehicles

!python train.py --strategy="all" --n_epochs=120 --batch_size=2 --learning_rate=0.0001 --small_addon_for_BCE=1e-6 --max_gradient_norm=5 --alpha_bce=1.5 --beta_bce=1 --huber_delta=3 --dump_vis="no" --data_root_dir="/content/gdrive/MyDrive/voxel_data/" --model_dir="model" --model_name="model6" --dump_test_interval=40 --summary_interval=4998 --summary_val_interval=4998 --summary_flush_interval=9996 --ckpt_max_keep=10

[Manueljohnson063]

Sir ,Kindly tell how did you add muliple objects .I am really stuck with it ,please help.(I had suceeded in detecting single object if you want i will share my work(Colab))

[Manueljohnson063]

please help sir

[chrisakatibs]

I didn't unfortunately. I am also struggling to get a lidar repo working. That would be appreciated. Currently I am trying a Point Pillars Repo and attempting to get their vector shapes the same as my input. My plan prior had been if I couldn't get multi-object detection with voxelnet I would run different trained single object detectors in unison. Brute force and to roundabout but that was the plan but I couldnt get my cars only dataset working (Lyft kitti format).

[Manueljohnson063]

@chrisakatibs did you want my google colab link for car detection ?

[chrisakatibs]

@Manueljohnson063 yeah here or DM [email protected]. :)

I'm wondering if my problem has to do with my data.

[GoKerrChen]

Hi Sir, I'm facing an issue when doing the training processing. Here is the error message poped up: AttributeError: in user code:

/content/Voxelnet/model_helper/training_helper.py:44 distributed_train_step  *
    per_replica_losses = strategy.experimental_run_v2(model.train_step,

AttributeError: 'OneDeviceStrategy' object has no attribute 'experimental_run_v2'

Could you help me look at it?

[chrisakatibs]

Not sure if it was the right move, but the new call is just strategy.run not the deprecated experimental_run_v2

[GoKerrChen]

Hi Chrissakatibs, thanks for your quick response! I haven't change anything regarding the code and I met this problem. I wonder if you did some modification or you just follow the code here?

[Manueljohnson063]

I hope Its due to the version of tensor flow you are using 1.instll (Colab) !pip install tensorflow-gpu==2.2.0rc2 2.check import tensorflow as tf print(tf.version)

and output : 2.2.0-rc2

[chrisakatibs]

I had to edit the files after I git cloned it every time before running the script on Google Colab. And then tried to follow the code. I did not succeed yet.

[GoKerrChen]

I hope Its due to the version of tensor flow you are using 1.instll (Colab) !pip install tensorflow-gpu==2.2.0rc2 2.check import tensorflow as tf print(tf.version)

and output : 2.2.0-rc2

Hey Sir, thx for your kind response and I succeed in installing the install but failed in check as Colab keeps telling me that there are no version attributes for tf

[GoKerrChen]

@chrisakatibs I think your problem might due to the non-cropping issue, here is the training screenshot

[GoKerrChen]

@Manueljohnson063 Thank you, sir, for the idea! I have figured out the issue, instead of install TensorFlow 2.2, install TensorFlow 2.0 as the attribute causing me an error has been removed after the 2.0 version.

[Manueljohnson063]

@GoKerrChen Bro there exist also problems in between training or at the end of training ,i got this version as I mentioned after a long research, if the present version is not succeed please use the above. If not Congrats. One more VOXELNET only detect single object!!!!!!!!!!!!!!!!!!!!!!!!!!!!

[Manueljohnson063]

@GoKerrChen Could you please me help if you are able to detect multiple objects in using this method([email protected]).

[GoKerrChen]

@Manueljohnson063 Sorry bro, I'm still facing the issue of evaluating the model and I haven't think about the multiple object detection yet.

[chrisakatibs]

So I was having a lot of trouble preprocessing files with MIT/second repo so I am using tyagi and possibly updating files with tjanna.

I got tyagi to start but i need to find some velodyne point cloud bins with incorrect indices messing up the training.

So at this point I still haven't trained the model.

Still having an issue with training on calibration or issue with running evaluation.

In regards to the calibration, I can send you an example calibration file and you might have to zero pad generate the files(hopefully that works) or sift through his code, making edits where the calib calculations are which is a higher degree of difficulty.

In regards to multiple object detection, I edited the config file to this for me.

classes = {"car": 0, "pedestrian": 1, "bicycle": 2, "bus": 3, "truck": 3, "other_vehicle": 3, "motorcycle": 2, #"truck": 4, #"other_vehicle": 5, #"motorcycle": 6, }

I didn't have the correct amount set because the tensor wouldn't reshape correctly when nb_classes = 7 on the detection head for classification. So I conformed my code to match their structure.

If it does work I'll let you know, If I didn't quite understand your question let me know,

Christopher Thibodeau Cal Poly Software Engineering Senior (408) 394-7534 [email protected]

On Tue, Mar 2, 2021 at 6:18 PM GoKerrChen [email protected] wrote:

@Manueljohnson063 https://github.com/Manueljohnson063 Sorry bro, I'm still facing the issue of evaluating the model and I haven't think about the multiple object detection yet.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/steph1793/Voxelnet/issues/3#issuecomment-789374073, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADX6H4ZXN4JGQTAWTMSQZFLTBWL73ANCNFSM4XB4L6MA .

[steph1793]

Hello Everyone I am so sorry for the delay in the response. I had to stop working on this project for a long moment. I will have a closer look at all the concerns and then come back to you.

So sorry again.