this repo is guide to segmentation using deeplabv3plus as sample code
this repo is want to guide to Semantic Segmentation with Deep Learning using gluon
this repo attemps to reproduce Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation use gluon reimplementation of DeepLabV3+ , but still different in current version
I use DenseNet as backbone instead of DCNN layer(Atrous Conv)
mxnet 1.1.0
1. clone( download)
2. execution deeplabv3+ ( training on VOC_less just 7 image)
3. show our training result ( you can change code in line 351)generate dataloader for segmentaion model
in the segmentaion task, we want to assign
each pixel in the image an object class.
2 case of segmentaion task
1. binary case (front / background)
2. multiple case (many object in one image)
label_indices: for mutiple class we want formulate segmentation task as classfication problem and use lookup table for any class
read_images: read dataset from our folder
load image first we should check which label format is our expect(binary or multiple class) and normalize our training data
normalize_image img = img / 255
getitem we can use any data augmentation method in this function like random crop or resize ,in this sample we just use resize
class SegDataset(gluon.data.Dataset):
def __init__(self,root,resize,colormap=None,classes=None):
self.root = root
self.resize = resize
self.colormap = colormap
self.classes = classes
self.colormap2label = None
self.load_images()
def label_indices(self,img):
if self.colormap2label is None:
self.colormap2label = nd.zeros(256**3)
for i, cm in enumerate(self.colormap):
self.colormap2label[(cm[0] * 256 + cm[1]) * 256 + cm[2]] = i
data = img.astype('int32')
idx = (data[:, :, 0] * 256 + data[:, :, 1]) * 256 + data[:, :, 2]
return self.colormap2label[idx]
def read_images(self,root):
dataroot = root + 'data/' # left_frames #data
labelroot = root + 'label/' # labels #label
DataNamelist = [f for f in listdir(dataroot)]
labelNamelist = [f for f in listdir(labelroot)]
if len(DataNamelist) != len(labelNamelist):
raise ValueError('number of your data is different from label')
else:
data, label = [None] * len(DataNamelist), [None] * len(labelNamelist)
for i, name in enumerate(DataNamelist):
data[i] = image.imread(dataroot + name)
for i, name in enumerate(labelNamelist):
label[i] = image.imread(labelroot + name)
return data, label
def load_images(self):
data, label = self.read_images(root=self.root)
self.data = [self.normalize_image(im) for im in data]
if self.colormap is None:
self.label = [self.normalize_image(im) for im in label]
if self.colormap != None:
self.label = label
print('read ' + str(len(self.data)) + ' examples')
def normalize_image(self,data):
return data.astype('float32') / 255
def __getitem__(self, item):
if self.colormap is None:
data = image.imresize(self.data[item], self.resize[0], self.resize[1])
label = image.imresize(self.label[item], self.resize[0], self.resize[1])
return data.transpose((2, 0, 1)), label.transpose((2,0,1))
if self.colormap != None:
data = image.imresize(self.data[item], self.resize[0], self.resize[1])
label = image.imresize(self.label[item], self.resize[0], self.resize[1])
return data.transpose((2, 0, 1)), self.label_indices(label)
def __len__(self):
return len(self.data)voc_colormap = [[0, 0, 0], [128, 0, 0], [0, 128, 0],
[128, 128, 0],
[0, 0, 128], [128, 0, 128], [0, 128, 128],
[128, 128, 128], [64, 0, 0], [192, 0, 0],
[64, 128, 0], [192, 128, 0], [64, 0, 128],
[192, 0, 128], [64, 128, 128], [192, 128, 128],
[0, 64, 0], [128, 64, 0], [0, 192, 0],
[128, 192, 0], [0, 64, 128]]
classes = ['background', 'aeroplane', 'bicycle', 'bird',
'boat', 'bottle', 'bus', 'car', 'cat', 'chair',
'cow', 'diningtable', 'dog', 'horse', 'motorbike',
'person', 'potted plant', 'sheep', 'sofa',
'train', 'tv/monitor']
def LoadDataset(dir, batchsize, output_shape, colormap = None, classes=None):
# select your dataset format (colormap, classes optional)
dataset = SegDataset(dir, output_shape,colormap,classes)
data_iter = gdata.DataLoader(dataset, batchsize, shuffle=False)
return data_iter
dir = 'VOC_less/'
ctx = mx.gpu()
batch_size = 3
resize = (480, 320)
train_iter = LoadDataset(dir, batch_size, resize, voc_colormap, classes) # default is for 2 class if you want to multiclass
follow this architecture but use densenet instead of DCNN layer
class Deeplabv3(nn.HybridBlock):
def __init__(self,growth_rate,numofcls):
super(Deeplabv3, self).__init__()
self.feature_extract = nn.HybridSequential()
with self.feature_extract.name_scope():
self.feature_extract.add(
stemblock(256),
DenseBlcok(6, growth_rate),
nn.BatchNorm(),
nn.Activation('relu')
)
self.conv1 = nn.HybridSequential()
with self.conv1.name_scope():
self.conv1.add(
nn.Conv2D(32, kernel_size=1, strides=2),
nn.BatchNorm(),
nn.Activation('relu')
)
self.conv3r6 = nn.HybridSequential()
with self.conv3r6.name_scope():
self.conv3r6.add(
nn.Conv2D(32, kernel_size=3, strides=2, padding=6, dilation=6),
nn.BatchNorm(),
nn.Activation('relu')
)
self.conv3r12 = nn.HybridSequential()
with self.conv3r12.name_scope():
self.conv3r12.add(
nn.Conv2D(32, kernel_size=3, strides=2, padding=12, dilation=12),
nn.BatchNorm(),
nn.Activation('relu')
)
self.conv3r18 = nn.HybridSequential()
with self.conv3r18.name_scope():
self.conv3r18.add(
nn.Conv2D(32,kernel_size=3,strides=2,padding=18,dilation=18),
nn.BatchNorm(),
nn.Activation('relu')
)
self.maxpool = nn.MaxPool2D(pool_size=2,strides=2)
self.concatconv1 = nn.HybridSequential()
with self.concatconv1.name_scope():
self.concatconv1.add(
nn.Conv2D(256,kernel_size=1),
nn.BatchNorm(),
nn.Activation('relu')
)
self.feconv1 = nn.HybridSequential()
with self.feconv1.name_scope():
self.feconv1.add(
nn.Conv2D(256,kernel_size=1),
nn.BatchNorm(),
nn.Activation('relu')
)
self.transUp = nn.HybridSequential()
with self.transUp.name_scope():
self.transUp.add(
nn.Conv2DTranspose(256,kernel_size=4,padding=1,strides=2),
nn.BatchNorm(),
nn.Activation('relu')
)
self.decodeConv3 = nn.HybridSequential()
with self.decodeConv3.name_scope():
self.decodeConv3.add(
nn.Conv2D(256, kernel_size=3, padding=1, strides=1),
nn.BatchNorm(),
nn.Activation('relu')
)
self.Up4 = nn.HybridSequential()
with self.Up4.name_scope():
self.Up4.add(
nn.Conv2DTranspose(256, kernel_size=4, padding=1, strides=2),
nn.BatchNorm(),
nn.Activation('relu'),
nn.Conv2DTranspose(numofcls,kernel_size=4,padding=1,strides=2)
)
def hybrid_forward(self, F, x):
out = self.feature_extract(x)
conv1out = self.conv1(out)
conv3r6out = self.conv3r6(out)
conv3r12out = self.conv3r12(out)
conv3r18out = self.conv3r18(out)
maxpoolout = self.maxpool(out)
second_out = ndarray.concat(conv1out,conv3r6out,conv3r12out,conv3r18out,maxpoolout, dim = 1)
encoder_out = self.concatconv1(second_out)
encoderUp = self.transUp(encoder_out)
feconv1out = self.feconv1(out)
combine_out = ndarray.concat(encoderUp, feconv1out, dim=1)
output = self.decodeConv3(combine_out)
output = self.Up4(output)
return output
training current model on VOC is ongoing i will keep going on Xception model and Atrous Conv, and fine tuning on benchmark dataset
this repo is want to help people learn more about gluon if you want to complete reproduce ,follow this repo https://github.com/duducheng/deeplabv3p_gluon