search

tnikolla / robot-grasp-detection

Detecting robot grasping positions with deep neural networks. The model is trained on Cornell Grasping Dataset. This is an implementation mainly based on the paper 'Real-Time Grasp Detection Using Convolutional Neural Networks' from Redmon and Angelova.
Apache License 2.0
232 stars 84 forks source link

About the models #4

Closed ClaireTun closed 6 years ago

ClaireTun commented 6 years ago

Hi, I am wondering whether the grasp model is the one pretrained on the imagenet? Can I use it to train the trainset? I tried , but the loss is nan?

tnikolla commented 6 years ago

Yes to both of your first two questions.

I had also the loss nan problem and this was the most frustrating of all. It turend out to be coused from bad data. If you are using the dataset from Cornell than do a search for a 'nan' or 'NAN' string in the *cpos.txt file. Tell me if you find something!

tnikolla commented 6 years ago

Hi,

at least in folder 01 of the dataset you will find some 'NaN' values for the rectangles. Do a grep -ri --include="*cpos.txt" "nan" . to find if there are others and just delete the rectangle that they fall in, so the four corresponding lines.

I wrote to the Cornell Lab about the bad data. I hope they will fix this!

Pragya001 commented 4 years ago

@tnikolla Hey i am using the imagenet model i.e., --model_path=./models/imagenet/m2/m2.ckpt, which is giving the following error -> "Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint."

And when i use --model_path=./models/grasp/m4/m4.ckpt the loss is nan. I remove all the bad files with nan rectangle from dataset, still it shows loss as NAN.

Can you please help me solve this. Also if you have a trained model can you mail me on : goyal.pragya001@gmail.com

This is really urgent, if you could help.