Build an object classifier for chess pawns

[abhidxt299]

Select a suitable object detector and train it on a dataset of chess pawns. You can use any of the dataset in the links provided.

1. https://public.roboflow.ai/object-detection/chess-full
2. https://www.kaggle.com/koryakinp/chess-positions

[DevanshD3]

Bhaiya is this the same issue that was given in the Whatsapp group? Like we had to train the model with different classes of chess pawns.

[abhidxt299]

Yes, it is. But this time, I wish you'd try out a different method of classifying the object. You can first try to implement whatever is shown in the 1st link.

[afrajshaikh]

I want to work on this isue.

[afrajshaikh]

Is there any specific subtask I need to do or the whole task. And do we need to choose an object detector directly from Roboflow?

[abhidxt299]

You have to do the entire task. For the 2nd part of the question, it's totally up to you. But I'd suggest you to start with Roboflow.

[afrajshaikh]

Ok

[Raghav-Dhir]

hey @afrajshaikh anyway we can work on this together, I was also interested , was thinking of using rcnn

[afrajshaikh]

hi @Raghav-Dhir, what about YOLO v5 ?

[Raghav-Dhir]

i asked @abhidxt299 about that , he said it would be preferred if we go for R-CNN's or RFCN's.

[afrajshaikh]

Cool

[afrajshaikh]

@Raghav-Dhir have you decided which one to use?

[Raghav-Dhir]

https://towardsdatascience.com/step-by-step-r-cnn-implementation-from-scratch-in-python-e97101ccde55 i think this should be good starting point

[afrajshaikh]

Cool

[afrajshaikh]

@Raghav-Dhir in which format you downloaded the dataset?

[Raghav-Dhir]

in csv format and also i think we should have a common collab file

[Raghav-Dhir]

although i have still figure out some things https://colab.research.google.com/drive/1zeN3JqbS6jDEWxa35RMJ-5Z5OndJ2qM-?usp=sharing we can do changes in this

[afrajshaikh]

I have requested access

[afrajshaikh]

@Raghav-Dhir i talked to @abhidxt299 he said we dont need to implement a detector from scratch right now, we can directly select a detector for now as we will be making our own afterwards. There are many object detectors available on roboflow under models section you check that

[abhidxt299]

@Raghav-Dhir @afrajshaikh you can try using YOLOv3 model. It'd be a really good start.

[Raghav-Dhir]

Sure

[Raghav-Dhir]

have made some changes in the shared notebook. wanted you to check before training the model

[afrajshaikh]

@Raghav-Dhir which model are you training?

[afrajshaikh]

I just kept YOLO v3 for training

[afrajshaikh]

like half an hour ago

[Raghav-Dhir]

yeah it was around 500 epochs so i reduced that and added the _classes.txt and _annotations.txt

[Raghav-Dhir]

just see the file once

[Raghav-Dhir]

also tell me if you can see the keras-yolo3 folder because i am not sure if it will visible to you or not

[afrajshaikh]

@Raghav-Dhir i guess you first downloaded the data on desktop and then you are importing it, you could have just imported the dataset by copying and pasting the download link.

[afrajshaikh]

I am not sure about the import dataset (from dektop) code, i tried it previously but was unable to do it, I think for using it in collab you need to first upload it on google drive folder. Seniors plz confirm.

[Raghav-Dhir]

@afrajshaikh i have copied the download link only

[afrajshaikh]

when i imported the dateset annotation was in txt format only, in your case it is in csv format so you converted it into txt right?

[afrajshaikh]

It has to be in txt then you can train it.

[Raghav-Dhir]

yes i did

[afrajshaikh]

For how many epoch you are training and your videos.py code block took lot of time?

[afrajshaikh]

https://drive.google.com/file/d/1-8tZgdZrPOkF9nHhOQL7IoNpTVS--H_G/view?usp=sharing I trained the model by keeping epochs 200. For training batch size was 8 and while testing it was 4. Test set loss was around 43 and validation set loss was around 40.

[afrajshaikh]

validation set batch_size=4*