chitra 0.1.0a0

A Deep Learning Computer Vision Utility library

chitra

What is chitra?

chitra (चित्र) is a Deep Learning Computer Vision library for easy data loading, model building and model interpretation with GradCAM/GradCAM++.

Highlights:

• Faster data loading without any boilerplate.
• Progressive resizing of images.
• Rapid experiments with different models using chitra.trainer module.
• Train models with cyclic learning rate.
• Model interpretation using GradCAM/GradCAM++ with no extra code.

📢 If you have more use case please raise an issue/PR with the feature you want.

📢 Join discord channel - https://discord.gg/TdnAfDw3kB

Installation

Using pip (recommended)

pip install -U chitra

From source

git clone https://github.com/aniketmaurya/chitra.git
cd chitra
pip install -e .

From GitHub

pip install git+https://github.com/aniketmaurya/chitra@master

Usage

Loading data for image classification

Chitra dataloader and datagenerator modules for loading data. dataloader is a minimal dataloader that returns tf.data.Dataset object. datagenerator provides flexibility to users on how they want to load and manipulate the data.

import numpy as np
import tensorflow as tf
import chitra
from chitra.dataloader import Clf, show_batch
import matplotlib.pyplot as plt

clf_dl = Clf()
data = clf_dl.from_folder(cat_dog_path, target_shape=(224, 224))

clf_dl.show_batch(8, figsize=(8,8))

for e in data.take(1):
    image = e[0].numpy().astype('uint8')
    label = e[1].numpy()
plt.imshow(image)
plt.show()

Image datagenerator

Dataset class provides the flexibility to load image dataset by updating components of the class.

Components of Dataset class are:

• image file generator
• resizer
• label generator
• image loader

These components can be updated with custom function by the user according to their dataset structure. For example the Tiny Imagenet dataset is organized as-

train_folder/
.....folder1/
    .....file.txt
    .....folder2/
           .....image1.jpg
           .....image2.jpg
                     .
                     .
                     .
           ......imageN.jpg
                    
                      

The inbuilt file generator search for images on the folder1, now we can just update the image file generator and rest of the functionality will remain same.

Dataset also support progressive resizing of images.

Updating component

from chitra.datagenerator import Dataset
from glob import glob

ds = Dataset(data_path)
# it will load the folders and NOT images
ds.filenames[:3]

No item present in the image size list





['/Users/aniket/Pictures/data/tiny-imagenet-200/train/n02795169/n02795169_boxes.txt',
 '/Users/aniket/Pictures/data/tiny-imagenet-200/train/n02795169/images',
 '/Users/aniket/Pictures/data/tiny-imagenet-200/train/n02769748/images']

def load_files(path):
    return glob(f'{path}/*/images/*')

def get_label(path):
    return path.split('/')[-3]
    
ds.update_component('get_filenames', load_files)
ds.filenames[:3]

get_filenames updated with <function load_files at 0x7fad6916d0e0>
No item present in the image size list





['/Users/aniket/Pictures/data/tiny-imagenet-200/train/n02795169/images/n02795169_369.JPEG',
 '/Users/aniket/Pictures/data/tiny-imagenet-200/train/n02795169/images/n02795169_386.JPEG',
 '/Users/aniket/Pictures/data/tiny-imagenet-200/train/n02795169/images/n02795169_105.JPEG']

Progressive resizing

It is the technique to sequentially resize all the images while training the CNNs on smaller to bigger image sizes. Progressive Resizing is described briefly in his terrific fastai course, “Practical Deep Learning for Coders”. A great way to use this technique is to train a model with smaller image size say 64x64, then use the weights of this model to train another model on images of size 128x128 and so on. Each larger-scale model incorporates the previous smaller-scale model layers and weights in its architecture. ~KDnuggets

image_sz_list = [(28, 28), (32, 32), (64, 64)]

ds = Dataset(data_path, image_size=image_sz_list)
ds.update_component('get_filenames', load_files)
ds.update_component('get_label', get_label)


print()
# first call to generator
for img, label in ds.generator():
    print('first call to generator:', img.shape)
    break

# seconds call to generator
for img, label in ds.generator():
    print('seconds call to generator:', img.shape)
    break

# third call to generator
for img, label in ds.generator():
    print('third call to generator:', img.shape)
    break

get_filenames updated with <function load_files at 0x7fad6916d0e0>
get_label updated with <function get_label at 0x7fad6916d8c0>

first call to generator: (28, 28, 3)
seconds call to generator: (32, 32, 3)
third call to generator: (64, 64, 3)

tf.data support

Creating a tf.data dataloader was never as easy as this one liner. It converts the Python generator into tf.data.Dataset for a faster data loading, prefetching, caching and everything provided by tf.data.

image_sz_list = [(28, 28), (32, 32), (64, 64)]

ds = Dataset(data_path, image_size=image_sz_list)
ds.update_component('get_filenames', load_files)
ds.update_component('get_label', get_label)

dl = ds.get_tf_dataset()

for e in dl.take(1):
    print(e[0].shape)

for e in dl.take(1):
    print(e[0].shape)

for e in dl.take(1):
    print(e[0].shape)

get_filenames updated with <function load_files at 0x7fad6916d0e0>
get_label updated with <function get_label at 0x7fad6916d8c0>
(28, 28, 3)
(32, 32, 3)
(64, 64, 3)

Trainer

The Trainer class inherits from tf.keras.Model, it contains everything that is required for training. It exposes trainer.cyclic_fit method which trains the model using Cyclic Learning rate discovered by Leslie Smith.

from chitra.trainer import Trainer, create_cnn
from chitra.datagenerator import Dataset
from PIL import Image

ds = Dataset(cat_dog_path, image_size=(224,224))
model = create_cnn('mobilenetv2', num_classes=2, name='Cat_Dog_Model')
trainer = Trainer(ds, model)
# trainer.summary()

WARNING:tensorflow:`input_shape` is undefined or non-square, or `rows` is not in [96, 128, 160, 192, 224]. Weights for input shape (224, 224) will be loaded as the default.

trainer.compile2(batch_size=8,
                 optimizer=tf.keras.optimizers.SGD(1e-3, momentum=0.9, nesterov=True),
                 lr_range=(1e-6, 1e-3),
                 loss='binary_crossentropy', 
                 metrics=['binary_accuracy'])

Model compiled!

trainer.cyclic_fit(epochs=5,
                   batch_size=8,
                   lr_range=(0.00001, 0.0001),                   
                  )

cyclic learning rate already set!
Epoch 1/5
1/1 [==============================] - 0s 14ms/step - loss: 6.4702 - binary_accuracy: 0.2500
Epoch 2/5
Returning the last set size which is: (224, 224)
1/1 [==============================] - 0s 965us/step - loss: 5.9033 - binary_accuracy: 0.5000
Epoch 3/5
Returning the last set size which is: (224, 224)
1/1 [==============================] - 0s 977us/step - loss: 5.9233 - binary_accuracy: 0.5000
Epoch 4/5
Returning the last set size which is: (224, 224)
1/1 [==============================] - 0s 979us/step - loss: 2.1408 - binary_accuracy: 0.7500
Epoch 5/5
Returning the last set size which is: (224, 224)
1/1 [==============================] - 0s 982us/step - loss: 1.9062 - binary_accuracy: 0.8750





<tensorflow.python.keras.callbacks.History at 0x7f8b1c3f2410>

Model Visualization

It is important to understand what is going inside the model. Techniques like GradCam and Saliency Maps can visualize what the Network is learning. trainer module has InterpretModel class which creates GradCam and GradCam++ visualization with almost no additional code.

from chitra.trainer import InterpretModel
trainer = Trainer(ds, create_cnn('mobilenetv2', num_classes=1000, keras_applications=False))
model_interpret = InterpretModel(True, trainer)

image = ds[1][0].numpy().astype('uint8')
image = Image.fromarray(image)
model_interpret(image)
print(IMAGENET_LABELS[285])

Returning the last set size which is: (224, 224)
index: 282

Egyptian Mau

Data Visualization

Image annotation

Thanks to fizyr keras-retinanet.

from chitra.visualization import draw_annotations

labels = np.array([label])
bbox = np.array([[30, 50, 170, 190]])
label_to_name = lambda x: 'Cat' if x==0 else 'Dog'

draw_annotations(image, ({'bboxes': bbox, 'labels':labels,}), label_to_name=label_to_name)
plt.imshow(image)
plt.show()

Utils

Limit GPU memory or enable dynamic GPU memory growth for Tensorflow

from chitra.utils import limit_gpu, gpu_dynamic_mem_growth

# limit the amount of GPU required for your training
limit_gpu(gpu_id=0, memory_limit=1024*2)

No GPU:0 found in your system!

gpu_dynamic_mem_growth()

No GPU found on the machine!

Contributing

Contributions of any kind are welcome. Please check the Contributing Guidelines before contributing.