Fields of The World - v3.1 "PRUE" CC-BY release

This release contains CC-BY versions of the previous PRUE model checkpoints. The models are trained using the same prue_efnet{3,5,7}.yaml configs located in configs/release/prue with the exception of removing noncommercial licensed countries from the train dataset.

Metrics below were computed on the Full FTW dataset using ftw model test -p3 -t3 --temporal_options stacked --model path/to/model.ckpt --countries full_data

Below are the original PRUE CC-BY-NC checkpoint metrics reference:

Consumption - Agriculture and Nutrition - Jupyter Notebook
Published by isaaccorley 6 months ago

Fields of The World - v3 "PRUE" release

This release contains a new series of models corresponding to the configs at configs/release/prue/:

• prue_efnet{3,5,7}_checkpoint.ckpt are 3-class U-Net models trained with channel shuffle, normalization augmentation, resize augmentation, a logcosh dice loss function, class weighting of background:0.05, field:0.2, field boundary:0.75 and efficientnet b{3,5,7} encoders on the full FTW dataset.
• prue_efnet{3,5,7}_standard_weight_checkpoint.ckpt are the same as above, but using the same class weights as the v1 models.
• prue_logcoshdice_only_checkpoint.ckpt is the same as the v1 model, but using a logcosh dice loss function.

We find these models outperform the previously released models in both performance and deployment metrics.

Consumption - Agriculture and Nutrition - Jupyter Notebook
Published by calebrob6 7 months ago

Fields of The World - v2 model release

This release contains two models:

• 3_Class_FULL_FTW_Pretrained_v2.ckpt was trained in the same way as the previous 3_Class_FULL_FTW_Pretrained.ckpt, but with a random shuffling of the order of window A and window B. This model has almost identical test set performance to the v1 model when the window ordering is the same, and much better performance when the window ordering is swapped.
• 3_Class_FULL_FTW_Pretrained_singleWindow_v2.ckpt was also trained in the same way as 3_Class_FULL_FTW_Pretrained.ckpt, but with a random selection of window A or window B for each sample. As a result, this model takes a single 4-channel input instead of the previous concatenated 8-channel input.

Consumption - Agriculture and Nutrition - Jupyter Notebook
Published by calebrob6 9 months ago

Fields of The World -

Consumption - Agriculture and Nutrition - Jupyter Notebook
Published by aninda-ghosh over 1 year ago

Fields of The World - Example Pre-trained Model

This release contains a src.ftw.trainers.CustomSemanticSegmentationTask checkpoint that has been pre-trained on the 3-class semantic segmentation labels using the training set from each country in FTW.

Consumption - Agriculture and Nutrition - Jupyter Notebook
Published by calebrob6 over 1 year ago