Image classifier for Indian Men/Women — Fast.ai Deep Learning Course — Part 1 — V3 using Colab

( TEST YOUR MODEL WITH CUSTOM IMAGE)

I wanted to see how i can use smaller datasets with Deep learning techniques for solving practical problems. I trained a model with 60 urban indian men and women images .The validation set contained 10 images each of women and men. I wanted to see how accurate my model was by testing with rural indian men/women.

1. DATA COLLECTION

I downloaded 60 images of urban men and women images from Google images and organised the data as shown below. The valid folder had 10 images of men and women ( i tried downloading black and white rural men/women images)

Indian_faces_dataset
    |-- train
         |-- man (60 images)
         |-- woman (50 images) 
    |-- valid
         |-- man (10 images)
         |-- woman (10 images)

You can download the dataset at ….

2. Upload data to your Google drive

Sign in to your Google drive using your gmail id. Upload the data. I uploaded it to drive/Mydrive/Fastaidata/Indian_faces_dataset

2. Google Colab

Sign in using your gmail account to Google Colab. Open a new Python3 Notebook.

Set up your free GPU in Colab : Just follow Edit > Notebook settings or Runtime>Change runtime type and select GPU as Hardware accelerator.

Check whether GPU is enabled in your notebook -

3. Set up fast ai version 1 in Google Colab

!pip install torch_nightly -f https://download.pytorch.org/whl/nightly/cu92/torch_nightly.html
!pip install fastai

4. Importing Libraries

%reload_ext autoreload
%autoreload 2
%matplotlib inline
from fastai import *
from fastai.vision import *
import os
import matplotlib.pyplot as plt
import matplotlib.image as mpimg

5. Mounting Google Drive locally to access data / files from your Drive:

from google.colab import drive
drive.mount('/content/drive/')

When you run the code above, you will get a link. Click the link . Select ‘Allow’. Copy Authorisation code and paste into Enter Authorisation text box and press Enter.

6. Data preprocessing

path = '/content/drive/My Drive/FastaiData/Indian_faces_dataset'
path;
np.random.seed(2)
tfms = get_transforms(do_flip=False)
#default bs=64, set image size=100 to run successfully on colab
data = ImageDataBunch.from_folder(path,ds_tfms=tfms, size=100)
data.show_batch(rows=3, figsize=(10,10))

Few images from your training dataset will be shown on 3 rows

7. Building/Training/ Evaluating the model

learn = create_cnn(data, models.resnet34, metrics=error_rate)
learn.fit_one_cycle(6)

We get the following output. We can see the error decreases as the number of epoch rises.

Total time: 00:09 
epoch  train_loss  valid_loss  error_rate 
1      0.976905    1.010379    0.550000    (00:01)        
2      0.742076    0.861343    0.450000    (00:01) 
3      0.612918    0.711301    0.400000    (00:01) 
4      0.529260    0.636116    0.350000    (00:01) 
5      0.446259    0.614106    0.350000    (00:01) 
6      0.398564    0.598918    0.250000    (00:01)

Save the model .

learn.save('stage-a')
interp = ClassificationInterpretation.from_learner(learn)
interp.plot_top_losses(9, figsize=(15,11))
interp.most_confused(min_val=2)
interp.plot_confusion_matrix()

8. Unfreezing the model and training all layers .

Total time: 00:08 
epoch  train_loss  valid_loss  error_rate 
1      0.263781    1.136744    0.400000    (00:01) 
2      0.184603    1.689964    0.350000    (00:01) 
3      0.140850    4.449215    0.350000    (00:01) 
4      0.122480    5.085160    0.400000    (00:01) 
5      0.110812    3.682216    0.300000    (00:01)

We can see that the error is not improving. Hence we are going to load the model saved as ‘stage-a’ and tune it by changing the learning rates

9. Fine Tuning

learn.lr_find()
learn.recorder.plot()

We can see that the loss reduces from 1e-6 to 1e-3. Let’s use 1e-6 to 1e-3 as the learning rates and see how our model performs.

learn.fit_one_cycle(3, max_lr=slice(1e-6,1e-3))

The output

Total time: 00:05 
epoch  train_loss  valid_loss  error_rate 
1      0.009076    2.324049    0.300000    (00:01) 
2      0.032253    1.738601    0.250000    (00:01) 
3      0.026521    1.630889    0.250000    (00:01)

With three epochs we are able to achieve 75% accuracy. Lets check our model against custom images and see how well it predicts.

10. Uploading a custom image.

Download an image that you want to test on your model from google images. Save it in your computer. For example, i have saved an image as testing_image1.jpeg . Upload the image to your drive in the folder where you have the dataset. Then run the code below. This loads the image in filename.

filename = '/content/drive/My Drive/FastaiData/Indian_faces_dataset/testing_image1.jpg'

11. Testing your model with a custom image

img = open_image(filename)
losses = img.predict(learn)
learn.data.classes[losses.argmax()]

It prints out the prediction as

'woman'

12. Checking/ Verifying

You can print your image to verify if the prediction made by your model is right. To print your image type the code below.

img=mpimg.imread(filename)
imgplot = plt.imshow(img)
plt.show()

Quite surprised at the predictions. The accuracy of this model has to be improved. With very little effort, with a very small dataset using the resnet34 model and fastai library, we have got a reasonably good model.

My notebook is available at https://github.com/bhuvanakundumani/ruralvsurbanindianfaces.git