Data from: f4-statistics-based ancestry profiling and convolutional neural network phenotyping shed new light on the structure of genetic and spike shape diversity in Aegilops tauschii Coss.

Spike image dataset of Aegilops tauschii Coss.

The dataset contains "dry" and "green" spike image files.

dry-spike-images.zip

Used for traininig and validating convolutional neural network (CNN).

green-spike-images.zip

Used for training and validating CNN.

dagestan-spike-images.zip

Green spike images of Dagestani and Russian accessions used for blind test of the trained CNN.

The Accession numbrer of the plant individual from which the spike was sampled, the lineage (TauL1, TauL2, or TauL3) to which the plant individual belongs, and md5sum value are shown in the following CSV files. Lineage information is left blank in labels-dagestan.csv because we used Dagestani and Russian accessions for a blind test of the trained CNN models.

labels-dry.csv

labels-green.csv

labels-dagestan.csv

GradCAM, Guided Backpropagation and Guided-GradCAM visualization results

CNN prediction explanation image files are archived into gradcam-results.zip. The image files are named in the following format.

__

The is one of gradcam_cam, gradcam_gb, gradcam_cam_gb, and these names represent GradCAM, Guided Backpropagation, Guided-GradCAM, respectively.

Trained ResNet weights

The trained ResNet model weights are archived into the following files. Each weight files are named according to the fold number and F1 score on  the cross-validation dataset in the fold; e.g. resnet50_2_0.98392.pth represents ResNet weight of fold 2, and the F1 score is 0.98392. 

resnet-dry-weights.zip

resnet-green-weights.zip

The weight files are PyTorch state_dict objects, and therefore can be loaded using the example Python code below.

import timm
import torch

model = timm.create_model("timm/resnet50.a1_in1k", num_classes=3)
state_dict = torch.load("resnet50_1_1.00000.pth")
model.load_state_dict(state_dict)