How to set the weights of model.ckpt.data to random for custom dataset

[Petros626]

Hello Guys,

I use the ssd_mobilenet_v2_quantized_300x300_coco for traing on bird eye view images of LiDAR data. The sparse data aren't good enough for training, even fine tuning the parameter like learning rate, change optimizer helped.

Is there a way to open the file model.ckpt.data-00000-of-00001 and set its weights to random or reset to reload the new file in the config file for training the model again?

Please help me with detailed instructions I need it for my degree...

[seboz123]

Just remove:
fine_tune_checkpoint_type: "detection"
fine_tune_checkpoint_version: V2
fine_tune_checkpoint: "/data/pretrained_model/ckpt-0"
from your pipeline config file

[Petros626]

Just remove: fine_tune_checkpoint_type: "detection" fine_tune_checkpoint_version: V2 fine_tune_checkpoint: "/data/pretrained_model/ckpt-0" from your pipeline config file

Hey @seboz123 ,

thank you for the quick response.
The second message parameter is unknown for me, but the other two are in the config file:

fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED/model.ckpt"
fine_tune_checkpoint_type: "detection"

I've read some approaches (#8892), where people did this, but the opinion was divergent and official authors never confirm this.
So there's no need to load the model.ckpt.data-00000-of-00001 in an IDE and set it there, just to be 100% sure?

[sushreebarsa]

@Petros626
In order to expedite the trouble-shooting process, could you please provide the entire URL of the repository which you are using. Please provide more details on the issue reported here.Thank you!

[Petros626]

@Petros626 In order to expedite the trouble-shooting process, could you please provide the entire URL of the repository which you are using. Please provide more details on the issue reported here.Thank you!

Hey @sushreebarsa,

in my case I'm using the TensorFlow 1 Object Detection API (https://github.com/tensorflow/models/tree/master/research/object_detection) and requested to reset all the pretrained weights, so that you could train the model architecture from scratch with your dataset and the parameters, which the authors figured out.