KERAS CLASSIFY
Image classification for projects and researches
About The Project
Image classification is a commonly used problem in the experimental part of scientific papers and also frequently appears as part of the projects. With the desire to reduce time and effort, Keras Classify was created.
Getting Started
Installation
-
Clone the repo: https://github.com/nguyentruonglau/keras-classify.git
-
Install packages
> python -m venv> activate.bat (in scripts folder) > pip install -r requirements.txt
Todo List:
- Cosine learning rate scheduler
- Gradient-based Localization
- Sota models
- Synthetic data
- Smart Resize
- Support Python 3.X and Tf 2.X
- Use imagaug for augmentation data
- Use prefetching and multiprocessing to training.
- Analysis Of Input Shape
- Compiled using XLA, auto-clustering on GPU
- Receiver operating characteristic
Quick Start
Analysis Of Input Shape
If your data has random input_shape, you don't know which input_shape to choose, the analysis program is the right choice for you. The algorithm is applied to analyze: Kernel Density Estimation.
Convert Data
From tensorflow 2.3.x already support auto fit_generator, however moving the data to npy file will make it easier to manage. The algorithm is applied to shuffle data: Random Permutation. Read more here.
Run: python convert/convert_npy.py
Training Model.
Design your model at model/models.py, we have made EfficientNetB0 the default. Adjust the appropriate hyperparameters and run: python train.py
Evaluate Model.
-
Statistics number of images per class after suffle on test data.
-
Provide model evalution indicators such as: Accuracy, Precesion, Recall, F1-Score and AUC (Area Under the Curve).
-
Plot training history of Accuracy, Loss, Receiver Operating Characteristic curve and Confusion Matrix.
Explainable AI.
Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. "We propose a technique for producing 'visual explanations' for decisions from a large class of CNN-based models, making them more transparent" Ramprasaath R. Selvaraju ... Read more here.
Example Code
Use for projects
from keras.preprocessing.image import load_img, img_to_array
from keras.preprocessing.image import smart_resize
from tensorflow.keras.models import load_model
import tensorflow as tf
import numpy as np
#load pretrained model
model_path = 'data/output/model/val_accuracy_max.h5'
model = load_model(model_path)
#load data
img_path = 'images/images.jpg'
img = load_img(img_path)
img = img_to_array(img)
img = smart_resize(img, (72,72)) #resize to HxW
img = np.expand_dims(img, axis=0)
#prediction
y_pred = model.predict(img)
y_pred = np.argmax(y_pred, axis=1)
#see convert/output/label_decode.json
print(y_pred)
Smart resize (tf < 2.4.1)
from tensorflow.keras.preprocessing.image import img_to_array
from tensorflow.keras.preprocessing.image load_img
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import image_ops
import numpy as np
def smart_resize(img, new_size, interpolation='bilinear'):
"""Resize images to a target size without aspect ratio distortion.
Arguments:
img (3D array): image data
new_size (tuple): HxW
Returns:
[3D array]: image after resize
"""
# Get infor of the image
height, width, _ = img.shape
target_height, target_width = new_size
crop_height = (width * target_height) // target_width
crop_width = (height * target_width) // target_height
# Set back to input height / width if crop_height / crop_width is not smaller.
crop_height = np.min([height, crop_height])
crop_width = np.min([width, crop_width])
crop_box_hstart = (height - crop_height) // 2
crop_box_wstart = (width - crop_width) // 2
# Infor to resize image
crop_box_start = array_ops.stack([crop_box_hstart, crop_box_wstart, 0])
crop_box_size = array_ops.stack([crop_height, crop_width, -1])
img = array_ops.slice(img, crop_box_start, crop_box_size)
img = image_ops.resize_images_v2(
images=img,
size=new_size,
method=interpolation)
return img.numpy()
Contributor
- BS Nguyen Truong Lau ([email protected])
- PhD Thai Trung Hieu ([email protected])
License
Distributed under the MIT License. See LICENSE for more information.
1.5k Dec 31, 2022
82 Jan 1, 2023
22 Dec 2, 2022
0 Mar 30, 2022
4 Feb 22, 2022
2 Dec 9, 2021
14 Oct 31, 2022
19 Dec 12, 2022
129 Dec 15, 2022
2 Oct 31, 2022
87 Jan 01, 2023
313 Dec 22, 2022
96 Dec 22, 2022
87 Dec 28, 2022
76 Dec 14, 2022
239 Dec 26, 2022
3 May 09, 2022
33 Dec 23, 2022
2 Nov 07, 2022
3 Jan 16, 2022
2 Nov 13, 2021
1.1k Jan 06, 2023
1 Jan 06, 2022
11 Sep 30, 2022
2.8k Jan 08, 2023
58 Oct 08, 2022
299 Dec 17, 2022
0 Nov 17, 2022