---
date: 2020-01-16 10:36
description: Tutorial on creating a custom Image Classifier using Turicreate and a dataset from Kaggle
tags: Tutorial, Colab, TuriCreate, AI
---
# Creating a Custom Image Classifier using Turicreate to detect Smoke and Fire
*For setting up Kaggle with Google Colab, please refer to <a href="/posts/2020-01-15-Setting-up-Kaggle-to-use-with-Colab"> my previous post</a>*
## Dataset
### Mounting Google Drive
```python
import os
from google.colab import drive
drive.mount('/content/drive')
```
### Downloading Dataset from Kaggle
```python
os.environ['KAGGLE_CONFIG_DIR'] = "/content/drive/My Drive/"
!kaggle datasets download ashutosh69/fire-and-smoke-dataset
!unzip "fire-and-smoke-dataset.zip"
```
## Pre-Processing
```Termcap
!mkdir default smoke fire
```
\
```Termcap
!ls data/data/img_data/train/default/*.jpg
```
\
```Termcap
img_1002.jpg img_20.jpg img_519.jpg img_604.jpg img_80.jpg
img_1003.jpg img_21.jpg img_51.jpg img_60.jpg img_8.jpg
img_1007.jpg img_22.jpg img_520.jpg img_61.jpg img_900.jpg
img_100.jpg img_23.jpg img_521.jpg 'img_62 (2).jpg' img_920.jpg
img_1014.jpg img_24.jpg 'img_52 (2).jpg' img_62.jpg img_921.jpg
img_1018.jpg img_29.jpg img_522.jpg 'img_63 (2).jpg' img_922.jpg
img_101.jpg img_3000.jpg img_523.jpg img_63.jpg img_923.jpg
img_1027.jpg img_335.jpg img_524.jpg img_66.jpg img_924.jpg
img_102.jpg img_336.jpg img_52.jpg img_67.jpg img_925.jpg
img_1042.jpg img_337.jpg img_530.jpg img_68.jpg img_926.jpg
img_1043.jpg img_338.jpg img_531.jpg img_700.jpg img_927.jpg
img_1046.jpg img_339.jpg 'img_53 (2).jpg' img_701.jpg img_928.jpg
img_1052.jpg img_340.jpg img_532.jpg img_702.jpg img_929.jpg
img_107.jpg img_341.jpg img_533.jpg img_703.jpg img_930.jpg
img_108.jpg img_3.jpg img_537.jpg img_704.jpg img_931.jpg
img_109.jpg img_400.jpg img_538.jpg img_705.jpg img_932.jpg
img_10.jpg img_471.jpg img_539.jpg img_706.jpg img_933.jpg
img_118.jpg img_472.jpg img_53.jpg img_707.jpg img_934.jpg
img_12.jpg img_473.jpg img_540.jpg img_708.jpg img_935.jpg
img_14.jpg img_488.jpg img_541.jpg img_709.jpg img_938.jpg
img_15.jpg img_489.jpg 'img_54 (2).jpg' img_70.jpg img_958.jpg
img_16.jpg img_490.jpg img_542.jpg img_710.jpg img_971.jpg
img_17.jpg img_491.jpg img_543.jpg 'img_71 (2).jpg' img_972.jpg
img_18.jpg img_492.jpg img_54.jpg img_71.jpg img_973.jpg
img_19.jpg img_493.jpg 'img_55 (2).jpg' img_72.jpg img_974.jpg
img_1.jpg img_494.jpg img_55.jpg img_73.jpg img_975.jpg
img_200.jpg img_495.jpg img_56.jpg img_74.jpg img_980.jpg
img_201.jpg img_496.jpg img_57.jpg img_75.jpg img_988.jpg
img_202.jpg img_497.jpg img_58.jpg img_76.jpg img_9.jpg
img_203.jpg img_4.jpg img_59.jpg img_77.jpg
img_204.jpg img_501.jpg img_601.jpg img_78.jpg
img_205.jpg img_502.jpg img_602.jpg img_79.jpg
img_206.jpg img_50.jpg img_603.jpg img_7.jpg
```
The image files are not actually JPEG, thus we first need to save them in the correct format for Turicreate
```python
from PIL import Image
import glob
folders = ["default","smoke","fire"]
for folder in folders:
n = 1
for file in glob.glob("./data/data/img_data/train/" + folder + "/*.jpg"):
im = Image.open(file)
rgb_im = im.convert('RGB')
rgb_im.save((folder + "/" + str(n) + ".jpg"), quality=100)
n +=1
for file in glob.glob("./data/data/img_data/train/" + folder + "/*.jpg"):
im = Image.open(file)
rgb_im = im.convert('RGB')
rgb_im.save((folder + "/" + str(n) + ".jpg"), quality=100)
n +=1
```
\
```Termcap
!mkdir train
!mv default ./train
!mv smoke ./train
!mv fire ./train
```
## Making the Image Classifier
### Making an SFrame
```Termcap
!pip install turicreate
```
\
```python
import turicreate as tc
import os
data = tc.image_analysis.load_images("./train", with_path=True)
data["label"] = data["path"].apply(lambda path: os.path.basename(os.path.dirname(path)))
print(data)
data.save('fire-smoke.sframe')
```
\
```Termcap
+-------------------------+------------------------+
| path | image |
+-------------------------+------------------------+
| ./train/default/1.jpg | Height: 224 Width: 224 |
| ./train/default/10.jpg | Height: 224 Width: 224 |
| ./train/default/100.jpg | Height: 224 Width: 224 |
| ./train/default/101.jpg | Height: 224 Width: 224 |
| ./train/default/102.jpg | Height: 224 Width: 224 |
| ./train/default/103.jpg | Height: 224 Width: 224 |
| ./train/default/104.jpg | Height: 224 Width: 224 |
| ./train/default/105.jpg | Height: 224 Width: 224 |
| ./train/default/106.jpg | Height: 224 Width: 224 |
| ./train/default/107.jpg | Height: 224 Width: 224 |
+-------------------------+------------------------+
[2028 rows x 2 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
+-------------------------+------------------------+---------+
| path | image | label |
+-------------------------+------------------------+---------+
| ./train/default/1.jpg | Height: 224 Width: 224 | default |
| ./train/default/10.jpg | Height: 224 Width: 224 | default |
| ./train/default/100.jpg | Height: 224 Width: 224 | default |
| ./train/default/101.jpg | Height: 224 Width: 224 | default |
| ./train/default/102.jpg | Height: 224 Width: 224 | default |
| ./train/default/103.jpg | Height: 224 Width: 224 | default |
| ./train/default/104.jpg | Height: 224 Width: 224 | default |
| ./train/default/105.jpg | Height: 224 Width: 224 | default |
| ./train/default/106.jpg | Height: 224 Width: 224 | default |
| ./train/default/107.jpg | Height: 224 Width: 224 | default |
+-------------------------+------------------------+---------+
[2028 rows x 3 columns]
Note: Only the head of the SFrame is printed.
You can use print_rows(num_rows=m, num_columns=n) to print more rows and columns.
```
### Making the Model
```python
import turicreate as tc
# Load the data
data = tc.SFrame('fire-smoke.sframe')
# Make a train-test split
train_data, test_data = data.random_split(0.8)
# Create the model
model = tc.image_classifier.create(train_data, target='label')
# Save predictions to an SArray
predictions = model.predict(test_data)
# Evaluate the model and print the results
metrics = model.evaluate(test_data)
print(metrics['accuracy'])
# Save the model for later use in Turi Create
model.save('fire-smoke.model')
# Export for use in Core ML
model.export_coreml('fire-smoke.mlmodel')
```
\
```Termcap
Performing feature extraction on resized images...
Completed 64/1633
Completed 128/1633
Completed 192/1633
Completed 256/1633
Completed 320/1633
Completed 384/1633
Completed 448/1633
Completed 512/1633
Completed 576/1633
Completed 640/1633
Completed 704/1633
Completed 768/1633
Completed 832/1633
Completed 896/1633
Completed 960/1633
Completed 1024/1633
Completed 1088/1633
Completed 1152/1633
Completed 1216/1633
Completed 1280/1633
Completed 1344/1633
Completed 1408/1633
Completed 1472/1633
Completed 1536/1633
Completed 1600/1633
Completed 1633/1633
PROGRESS: Creating a validation set from 5 percent of training data. This may take a while.
You can set ``validation_set=None`` to disable validation tracking.
Logistic regression:
--------------------------------------------------------
Number of examples : 1551
Number of classes : 3
Number of feature columns : 1
Number of unpacked features : 2048
Number of coefficients : 4098
Starting L-BFGS
--------------------------------------------------------
+-----------+----------+-----------+--------------+-------------------+---------------------+
| Iteration | Passes | Step size | Elapsed Time | Training Accuracy | Validation Accuracy |
+-----------+----------+-----------+--------------+-------------------+---------------------+
| 0 | 6 | 0.018611 | 0.891830 | 0.553836 | 0.560976 |
| 1 | 10 | 0.390832 | 1.622383 | 0.744681 | 0.792683 |
| 2 | 11 | 0.488541 | 1.943987 | 0.733075 | 0.804878 |
| 3 | 14 | 2.442703 | 2.512545 | 0.727917 | 0.841463 |
| 4 | 15 | 2.442703 | 2.826964 | 0.861380 | 0.853659 |
| 9 | 28 | 2.340435 | 5.492035 | 0.941328 | 0.975610 |
+-----------+----------+-----------+--------------+-------------------+---------------------+
Performing feature extraction on resized images...
Completed 64/395
Completed 128/395
Completed 192/395
Completed 256/395
Completed 320/395
Completed 384/395
Completed 395/395
0.9316455696202531
```
We just got an accuracy of 94% on Training Data and 97% on Validation Data!