cannot load new trained model for inference

[JunMa11]

Dear Cellpose developers,

Train

I trained a new model with my own data by

python -m cellpose --train --use_gpu --dir /media/cellpose/imagesTr/imagesTrTif/ --chan 2 --chan2 1 --n_epochs 600 --pretrained_model None --batch_size 8 --dir_above --save_each --verbose --img_filter '_img' --mask_filter '_masks'

The trained model was saved in the image path

Inference

from cellpose import models, io
import os
join = os.path.join
import numpy as np
from tqdm import tqdm
import tifffile as tif

# model_type='cyto' or model_type='nuclei'
model_path = '/media/cellpose/imagesTr/imagesTrTif/models/cellpose.028618_epoch_599'
model = models.CellposeModel(gpu=True, model_type='cyto', pretrained_model = model_path)
print('model info:', model.pretrained_model)
print('start predicting....')

img_path = '/media/cellpose/imagesTs/images'
names = sorted(os.listdir(img_path))
seg_path = '/media/seg-images/cellpose-seg'
os.makedirs(seg_path, exist_ok=True)

chan = [2,1]
for i, name in enumerate(tqdm(names)):
    save_name = name.split('_img')[0]+'_label.tiff'
    if not os.path.isfile(join(seg_path, save_name)):
        if name.endswith('.tif') or name.endswith('.tiff'):
            img = tif.imread(join(img_path, name))
        else:
            img = io.imread(join(img_path, name))
        masks, flows, styles = model.eval(img, diameter=None, channels=chan, net_avg=False, progress=True)

However, I find the script does not load my trained model. It still loads the default model [/home/.cellpose/models/cytotorch_0]

How can I load the new trained model for inference?

[JunMa11]

It works by replacing the default model file in .cellpose/models with the trained model.

[carsen-stringer]

I've added the function io.add_model(/full/path/to/model) to add the model file to the hidden folder to use with GUI and CLI. You can also access the function by running python -m cellpose --add_model /full/path/to/model.

[carsen-stringer]

In your case though, it wasn't working because you were using model_type='cyto' -- the model was by default using that input, instead in the future if you don't want to add your models to the hidden folder use command

model = models.CellposeModel(gpu=True, pretrained_model = model_path)