Binary Unet Segmentation

mr.ashutoshraj · January 22, 2020, 10:30am

I am trying to train Scene text using U-Net architecture as shown in camvid lesson.

I am getting my training and validation loss as 0 after some time. I am not getting any results also. Please let me know if anyone have suggestions.

Details:
Training Images: 5603
Mask Image: 5603
Image Size: 256X256

Original Image

Mask Image

mr.ashutoshraj · January 22, 2020, 10:35am

Notebook example:

mschmit5 · January 22, 2020, 10:56am

The notebook you linked doesn’t have 0 training loss and validation loss. Can you show us your the rest of your code? I’m not qualified to answer your question but I’m trying to help you get an answer quicker. Read How to ask for help

mr.ashutoshraj · January 22, 2020, 11:59am

I am following the same code. Just changed the dataset folder.

adrian · January 22, 2020, 1:41pm

When you run the code on the camvid dataset do you get decent result? If so check your dataset. In the images shown the mask is all black-is that just a remdering issue of the pic-do you see correct mask in fastai show method?

mr.ashutoshraj · January 23, 2020, 4:41am

Hi Adrian,

I have used the correct mask. If you open it through an image processing library. You will find only 1 and 0 pixels, as it is a binary classification problem. I have uploaded my notebook please check into this.

github.com

ashutoshraj/TextSegmentation/blob/master/detection.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Text segmentation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "%matplotlib inline"
   ]
  },

This file has been truncated. show original

This dataset is from ICDAR 2019 train_images.zip

adrian · January 23, 2020, 1:01pm

Looks all OK until fit_one_cycle() can you try fir_one_cycle with the default pct_start of 0.3 and see if that improves the result

mr.ashutoshraj · January 23, 2020, 1:27pm

No any improvement. Is it because the model is not finding any constant features across the dataset?

adrian · January 23, 2020, 1:33pm

Does it run OK if you use the camvid dataset?

How many images do you have? - if not many, instead of get_transforms() try adding a bunch of different transforms to boost your data sample ± get more images.

juvian · January 23, 2020, 9:21pm

You are doing open_mask(get_y_fn(img_f)) to show mask in jupyter notebook but you then use return open_mask(fn, div=True) in the actual dataset. The div=True does division by 255, so if your mask is already in 0-1 format which might be the case, you are then making it all 0. And that probably makes sense because I don’t notice the masks in show_batch

mr.ashutoshraj · January 24, 2020, 5:49am

Thanks. I have done the mistake here by creating a mask of 0 and 1 pixels. The problem got solved.
Thanks once again

mr.ashutoshraj · January 24, 2020, 5:49am

I will try this one also. It will defintely boost my accuracy.
Thanks adrian

Adarsh97 · July 28, 2020, 7:23am

hi. i am just starting off segmentation. i just want to know how heavy were your saved models? I am trying a similar task but my resnet50 model, without any fine-tuning, is occupying 3.2GB which is quite a lot in comparison to the 200-300Mb models i saved for classification tasks.

I understand that segmentation is kind of a per-pixel classification and thus is more resource-intensive so does that mean segmentation models would be 10 times more heavier in terms of space as is my case?

Thank you!

Adarsh97 · August 26, 2020, 10:56am

did you actually submit it in kaggle? i am having issues in adding the test files and it would be of great help if you could guide me to do the same.