Update README.md

paper/3.0/README.md

@@ -4,9 +4,17 @@
 
 To train the super-generalist cyto3 segmentation model, we did not perform any image processing on the external datasets other than normalizing the image per channel. 
 The livecell masks contain overlaps, so we removed overlaps with this [function](https://github.com/MouseLand/cellpose/blob/ae795e0f95cb2ecfc3e5b24185a2370bd2cd2225/paper/2.0/datasets.py#L271). 
-The training command for the model is as follows:
+The training command for the model is as follows (assuming the tiffs are normalized per channel already):
 
-python -m cellpose --file_list probs_generalist.npy --dir /media/carsen/ssd4/datasets_cellpose/ --verbose --train --min_train_masks 0 --chan 1 --chan2 2 --pretrained_model None --no_norm --nimg_per_epoch 800 --nimg_test_per_epoch 100 --SGD 0 --learning_rate 0.005 --n_epochs 2000 --use_gpu --train_size
+python -m cellpose --file_list probs_generalist.npy --dir /media/carsen/ssd4/datasets_cellpose/ --verbose --train --min_train_masks 0 --chan 1 --chan2 2 --pretrained_model None --no_norm --nimg_per_epoch 800 --nimg_test_per_epoch 100 --SGD 0 --learning_rate 0.005 --n_epochs 5000 --use_gpu --train_size
 
-See the example npy file in the folder for how to structure cellpose to use filelist as an input for training.
+See the example probs_generalist.npy file [here](https://www.cellpose.org/static/data/probs_generalist.npy) for how to structure cellpose to use filelist as an input for training. If you're on windows, you may need to set the pathlib path before loading the npy:
+```
+import pathlib
+pathlib.PosixPath = pathlib.WindowsPath
+import numpy as np
+
+dat = np.load("probs_generalist.npy", allow_pickle=True).item()
+print(dat.keys())
+```