Keras implementation of Deeplabv3+
This repo is not longer maintained. I won't respond to issues but will merge PR
DeepLab is a state-of-art deep learning model for semantic image segmentation.
Model is based on the original TF frozen graph. It is possible to load pretrained weights into this model. Weights are directly imported from original TF checkpoint.
Segmentation results of original TF model. Output Stride = 8
Segmentation results of this repo model with loaded weights and OS = 8
Results are identical to the TF model
Segmentation results of this repo model with loaded weights and OS = 16
Results are still good
How to get labels
Model will return tensor of shape (batch_size, height, width, num_classes). To obtain labels, you need to apply argmax to logits at exit layer. Example of predicting on image1.jpg:
import numpy as np
from PIL import Image
from matplotlib import pyplot as plt
from model import Deeplabv3
# Generates labels using most basic setup. Supports various image sizes. Returns image labels in same format
# as original image. Normalization matches MobileNetV2
trained_image_width=512
mean_subtraction_value=127.5
image = np.array(Image.open('imgs/image1.jpg'))
# resize to max dimension of images from training dataset
w, h, _ = image.shape
ratio = float(trained_image_width) / np.max([w, h])
resized_image = np.array(Image.fromarray(image.astype('uint8')).resize((int(ratio * h), int(ratio * w))))
# apply normalization for trained dataset images
resized_image = (resized_image / mean_subtraction_value) - 1.
# pad array to square image to match training images
pad_x = int(trained_image_width - resized_image.shape[0])
pad_y = int(trained_image_width - resized_image.shape[1])
resized_image = np.pad(resized_image, ((0, pad_x), (0, pad_y), (0, 0)), mode='constant')
# make prediction
deeplab_model = Deeplabv3()
res = deeplab_model.predict(np.expand_dims(resized_image, 0))
labels = np.argmax(res.squeeze(), -1)
# remove padding and resize back to original image
if pad_x > 0:
labels = labels[:-pad_x]
if pad_y > 0:
labels = labels[:, :-pad_y]
labels = np.array(Image.fromarray(labels.astype('uint8')).resize((h, w)))
plt.imshow(labels)
plt.waitforbuttonpress()
How to use this model with custom input shape and custom number of classes
from model import Deeplabv3
deeplab_model = Deeplabv3(input_shape=(384, 384, 3), classes=4)
#or you can use None as shape
deeplab_model = Deeplabv3(input_shape=(None, None, 3), classes=4)
After that you will get a usual Keras model which you can train using .fit and .fit_generator methods.
How to train this model
Useful parameters can be found in the original repository.
Important notes:
- This model doesn’t provide default weight decay, user needs to add it themselves.
- Due to huge memory use with
OS=8, Xception backbone should be trained withOS=16and only inferenced withOS=8. - User can freeze feature extractor for Xception backbone (first 356 layers) and only fine-tune decoder. Right now (March 2019), there is a problem with finetuning Keras models with BN. You can read more about it here.
Known issues
This model can be retrained check this notebook. Finetuning is tricky and difficult because of the confusion between training and trainable in Keras. See this issue for a discussion and possible alternatives.
How to load model
In order to load model after using model.save() use this code:
from model import relu6
deeplab_model = load_model('example.h5')
Xception vs MobileNetv2
There are 2 available backbones. Xception backbone is more accurate, but has 25 times more parameters than MobileNetv2.
For MobileNetv2 there are pretrained weights only for alpha=1. However, you can initiate model with different values of alpha.
Requirement
The latest vesrion of this repo uses TF Keras, so you only need TF 2.0+ installed
tensorflow-gpu==2.0.0a0
CUDA==9.0
If you want to use older version, use following commands:
git clone https://github.com/bonlime/keras-deeplab-v3-plus/
cd keras-deeplab-v3-plus/
git checkout 714a6b7d1a069a07547c5c08282f1a706db92e20
tensorflow-gpu==1.13
Keras==2.2.4
We use the first img as input, and resize it from (427, 640, 3) to (512, 512, 3), and normalize it from [0,255] to [0,1].
As below, the output of OS=16 is normal,
795 Jan 03, 2023
5 Jan 03, 2023
65 Oct 06, 2022
181 Dec 28, 2022
6 Jun 27, 2022
73 Dec 04, 2022
85 Oct 18, 2022
559 Dec 14, 2022
4 Mar 01, 2022
72 Dec 24, 2022
31 Oct 19, 2019
28 Dec 15, 2022
33 Dec 01, 2022
51 Aug 29, 2022
10 Nov 02, 2022
111 Dec 11, 2022
96 Oct 26, 2022
119 Jan 04, 2023
30 Dec 28, 2022
87 Nov 30, 2022