Releases: jgrss/cultionet
Releases · jgrss/cultionet
v1.7.3
v1.7.2
v1.7.1
v1.7.0
What changed?
- Improved and fixed issues in the ResUNet 3+ Psi architecture, which was introduced in
v1.6.5
- More flexible user arguments. The user can now specify:
- the model architecture
- convolution blocks
- dilations
- attention weights
- Improvements in the train optimizer stability
- Deep supervision
- Cultionet uses the UNet 3+ style of deep supervision along three decoders
- These are optional during training
- Cultionet uses the UNet 3+ style of deep supervision along three decoders
- Improved training efficiency using PyTorch’s parallel data loader
- Improved inference efficiency using PyTorch’s batch loader
v1.6.5
v1.6.4
v1.6.3
v1.6.2
v1.6.1
v1.6.0
What's new?
- New architecture design based on UNet 3+ and residual convolutions
- The new design is a multi-head connection of the UNet 3+ architecture
- Added optional crop-type model for finer crop learning
- Modified total loss quantification with deep supervision of crop type in RNN layer
- The tanimoto loss is used on all layers
- Added
num_workers
option inDataLoader
for faster train/predict - Added .pt data compression by changing
torch.save|load
tojoblib.dump|load