fusionlab 0.1.4

Useful packages for DL

FusionLab

FusionLab is an open-source frameworks built for Deep Learning research written in PyTorch and Tensorflow. The code is easy to read and modify especially for newbie. Feel free to send pull requests :D

• What's New
• Installation
• How to use
• Encoders
• Losses
• Segmentation
• 1D, 2D, 3D Model
• Acknowledgements

Installation

With pip

pip install fusionlab

For Mac M1 chip users

Install on Macbook M1 chip

How to use

import fusionlab as fl

# PyTorch
encoder = fl.encoders.VGG16()
# Tensorflow
encoder = fl.encoders.TFVGG16()

Encoders

encoder list

Losses

Loss func list

• Dice Loss
• Tversky Loss
• IoU Loss

# Dice Loss (Multiclass)
import fusionlab as fl

# PyTorch
pred = torch.randn(1, 3, 4, 4) # (N, C, *)
target = torch.randint(0, 3, (1, 4, 4)) # (N, *)
loss_fn = fl.losses.DiceLoss()
loss = loss_fn(pred, target)

# Tensorflow
pred = tf.random.normal((1, 4, 4, 3), 0., 1.) # (N, *, C)
target = tf.random.uniform((1, 4, 4), 0, 3) # (N, *)
loss_fn = fl.losses.TFDiceLoss("multiclass")
loss = loss_fn(target, pred)


# Dice Loss (Binary)

# PyTorch
pred = torch.randn(1, 1, 4, 4) # (N, 1, *)
target = torch.randint(0, 3, (1, 4, 4)) # (N, *)
loss_fn = fl.losses.DiceLoss("binary")
loss = loss_fn(pred, target)

# Tensorflow
pred = tf.random.normal((1, 4, 4, 1), 0., 1.) # (N, *, 1)
target = tf.random.uniform((1, 4, 4), 0, 3) # (N, *)
loss_fn = fl.losses.TFDiceLoss("binary")
loss = loss_fn(target, pred)

Segmentation

import fusionlab as fl
# PyTorch UNet
unet = fl.segmentation.UNet(cin=3, num_cls=10)

# Tensorflow UNet
# Multiclass Segmentation
unet = tf.keras.Sequential([
   fl.segmentation.TFUNet(num_cls=10, base_dim=64),
   tf.keras.layers.Activation(tf.nn.softmax),
])

# Binary Segmentation
unet = tf.keras.Sequential([
   fl.segmentation.TFUNet(num_cls=1, base_dim=64),
   tf.keras.layers.Activation(tf.nn.sigmoid),
])

Segmentation model list

• UNet
• ResUNet
• UNet2plus

N Dimensional Model

some models can be used in 1D, 2D, 3D

import fusionlab as fl

resnet1d = fl.encoders.ResNet50V1(cin=3, spatial_dims=1)
resnet2d = fl.encoders.ResNet50V1(cin=3, spatial_dims=2)
resnet3d = fl.encoders.ResNet50V1(cin=3, spatial_dims=3)

unet1d = fl.segmentation.UNet(cin=3, num_cls=10, spatial_dims=1)
unet2d = fl.segmentation.UNet(cin=3, num_cls=10, spatial_dims=2)
unet3d = fl.segmentation.UNet(cin=3, num_cls=10, spatial_dims=3)

News

Release logs

Acknowledgements

• BloodAxe/pytorch-toolbelt