MouseLand
[BUG] Custom model not able to detect ROIs after training on cyto3,
Describe the bug I have been struggling with this for a few days. Please bear with me as I am not very good at coding yet and trying to fix this problem has tested the limits of my admittedly minimal knowledge of python, coding in general, etc. I am trying to generate ROIs of simple, GFP-tranfected CHO cells which I can then load into FIJI to extract fluorescence intensity data from another channel.
The default cyto3 model works very well in general, HOWEVER: I have to set the diameter to between 95-107 to capture most of the ROIs. It detects very few ROIs with the default diameter of 30. The problem is that sometimes it also detects miniscule ROIs with no area or that don't correspond to a cell at all. I thought that by annotating about 40 images where I remove these small ROIs, it would eventually learn not to detect them. However, instead every trained model detects 0 ROIs, despite cyto3 detecting between 40 and 60 per image I'm using for training.
I have so far tried:
Training from scratch (ie, no default model): 0 ROIs detected Training from models that aren't Cyto3 (namely, the bac_fluor model, which has worked as a training model w/ which to identify bacteria in a much more challenging set of images): 0 ROIs detected Forcing the use of RAdam instead of SGD, as was suggested previously, which required me to run the training through the command line instead of the GUI, but I did eventually figure this out: 0 ROIs detected Removing all images that have fewer than 20 ROIs from the training set: 0 ROIs detected Increasing Epochs from 100 to 300 to 500 to 1000: 0 ROIs detected in all conditions.
Reproducing Here is a google drive link containing the images I have been using for training + the _seg.npy outputs (https://drive.google.com/drive/folders/1884j88XyEyndEBN8BfIIyQk4b2wUc2lU?usp=sharing) See if you can train on these using Cyto3 as a baseline?
Run log Here is the terminal output from the last time I tried to train on cyto3 and make a model without those annoying little ROIs.
`2024-12-03 12:49:15,919 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f25d1.tif'> OME series is missing 1 frames. Missing data are zeroed
2024-12-03 12:49:15,976 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f26d1.tif'> OME series is missing 1 frames. Missing data are zeroed
2024-12-03 12:49:16,019 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f27d1.tif'> OME series is missing 1 frames. Missing data are zeroed
2024-12-03 12:49:16,063 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f28d1.tif'> OME series is missing 1 frames. Missing data are zeroed
2024-12-03 12:49:16,104 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f29d1.tif'> OME series is missing 1 frames. Missing data are zeroed
2024-12-03 12:49:16,145 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f30d1.tif'> OME series is missing 1 frames. Missing data are zeroed
2024-12-03 12:49:44,219 [INFO] training with ['scan_Plate_R_p00_0_C04f25d1.tif', 'scan_Plate_R_p00_0_C04f26d1.tif', 'scan_Plate_R_p00_0_C04f27d1.tif', 'scan_Plate_R_p00_0_C04f28d1.tif', 'scan_Plate_R_p00_0_C04f29d1.tif', 'scan_Plate_R_p00_0_C04f30d1.tif']
2024-12-03 12:49:44,219 [INFO] training new model starting at model cyto3
2024-12-03 12:49:44,219 [INFO] training with chan = 0: gray, chan2 = 0: none
2024-12-03 12:49:44,219 [INFO] >> cyto3 << model set to be used
2024-12-03 12:49:44,220 [INFO] ** TORCH MPS version installed and working. **
2024-12-03 12:49:44,221 [INFO] >>>> using GPU (MPS)
2024-12-03 12:49:44,307 [INFO] >>>> loading model /Users/Alexander.Morano/.cellpose/models/cyto3
2024-12-03 12:49:44,388 [INFO] >>>> model diam_mean = 30.000 (ROIs rescaled to this size during training) GUI_INFO: name of new model: CP_20241203_124811
2024-12-03 12:49:44,388 [INFO] computing flows for labels 100%|█████████████████████████████████████████████| 6/6 [00:14<00:00, 2.43s/it]
2024-12-03 12:49:58,999 [INFO] >>> computing diameters 100%|████████████████████████████████████████████| 6/6 [00:00<00:00, 100.68it/s]
2024-12-03 12:49:59,059 [INFO] >>> using channels [0, 0]
2024-12-03 12:49:59,059 [INFO] >>> normalizing {'lowhigh': None, 'percentile': [1.0, 99.0], 'normalize': True, 'norm3D': True, 'sharpen_radius': 0, 'smooth_radius': 0, 'tile_norm_blocksize': 0, 'tile_norm_smooth3D': 1, 'invert': False}
2024-12-03 12:49:59,819 [INFO] >>> n_epochs=1000, n_train=6, n_test=None
2024-12-03 12:49:59,819 [INFO] >>> SGD, learning_rate=0.10000, weight_decay=0.00010, momentum=0.900
2024-12-03 12:50:00,366 [INFO] >>> saving model to /Users/Alexander.Morano/Desktop/cellpose_training/old_plate/untitled_folder/models/CP_20241203_124811
2024-12-03 12:50:01,234 [INFO] 0, train_loss=nan, test_loss=0.0000, LR=0.0000, time 0.87s
2024-12-03 12:50:03,862 [INFO] 5, train_loss=nan, test_loss=0.0000, LR=0.0556, time 3.50s
2024-12-03 12:50:06,394 [INFO] 10, train_loss=nan, test_loss=0.0000, LR=0.1000, time 6.03s
2024-12-03 12:50:11,418 [INFO] 20, train_loss=nan, test_loss=0.0000, LR=0.1000, time 11.05s
2024-12-03 12:50:16,631 [INFO] 30, train_loss=nan, test_loss=0.0000, LR=0.1000, time 16.26s
2024-12-03 12:50:21,745 [INFO] 40, train_loss=nan, test_loss=0.0000, LR=0.1000, time 21.38s
2024-12-03 12:50:26,784 [INFO] 50, train_loss=nan, test_loss=0.0000, LR=0.1000, time 26.42s
2024-12-03 12:50:31,839 [INFO] 60, train_loss=nan, test_loss=0.0000, LR=0.1000, time 31.47s
2024-12-03 12:50:36,883 [INFO] 70, train_loss=nan, test_loss=0.0000, LR=0.1000, time 36.52s
2024-12-03 12:50:41,890 [INFO] 80, train_loss=nan, test_loss=0.0000, LR=0.1000, time 41.52s
2024-12-03 12:50:46,994 [INFO] 90, train_loss=nan, test_loss=0.0000, LR=0.1000, time 46.63s
2024-12-03 12:50:52,022 [INFO] 100, train_loss=nan, test_loss=0.0000, LR=0.1000, time 51.66s
2024-12-03 12:50:57,150 [INFO] 110, train_loss=nan, test_loss=0.0000, LR=0.1000, time 56.78s
2024-12-03 12:51:02,305 [INFO] 120, train_loss=nan, test_loss=0.0000, LR=0.1000, time 61.94s
2024-12-03 12:51:07,308 [INFO] 130, train_loss=nan, test_loss=0.0000, LR=0.1000, time 66.94s
2024-12-03 12:51:12,327 [INFO] 140, train_loss=nan, test_loss=0.0000, LR=0.1000, time 71.96s
2024-12-03 12:51:17,350 [INFO] 150, train_loss=nan, test_loss=0.0000, LR=0.1000, time 76.98s
2024-12-03 12:51:22,393 [INFO] 160, train_loss=nan, test_loss=0.0000, LR=0.1000, time 82.03s
2024-12-03 12:51:27,499 [INFO] 170, train_loss=nan, test_loss=0.0000, LR=0.1000, time 87.13s
2024-12-03 12:51:32,573 [INFO] 180, train_loss=nan, test_loss=0.0000, LR=0.1000, time 92.21s
2024-12-03 12:51:37,602 [INFO] 190, train_loss=nan, test_loss=0.0000, LR=0.1000, time 97.24s
2024-12-03 12:51:42,637 [INFO] 200, train_loss=nan, test_loss=0.0000, LR=0.1000, time 102.27s
2024-12-03 12:51:47,677 [INFO] 210, train_loss=nan, test_loss=0.0000, LR=0.1000, time 107.31s
2024-12-03 12:51:52,751 [INFO] 220, train_loss=nan, test_loss=0.0000, LR=0.1000, time 112.39s
2024-12-03 12:51:57,823 [INFO] 230, train_loss=nan, test_loss=0.0000, LR=0.1000, time 117.46s
2024-12-03 12:52:02,951 [INFO] 240, train_loss=nan, test_loss=0.0000, LR=0.1000, time 122.58s
2024-12-03 12:52:08,185 [INFO] 250, train_loss=nan, test_loss=0.0000, LR=0.1000, time 127.82s
2024-12-03 12:52:13,248 [INFO] 260, train_loss=nan, test_loss=0.0000, LR=0.1000, time 132.88s
2024-12-03 12:52:18,512 [INFO] 270, train_loss=nan, test_loss=0.0000, LR=0.1000, time 138.15s
2024-12-03 12:52:23,587 [INFO] 280, train_loss=nan, test_loss=0.0000, LR=0.1000, time 143.22s
2024-12-03 12:52:28,610 [INFO] 290, train_loss=nan, test_loss=0.0000, LR=0.1000, time 148.24s
2024-12-03 12:52:33,724 [INFO] 300, train_loss=nan, test_loss=0.0000, LR=0.1000, time 153.36s
2024-12-03 12:52:38,912 [INFO] 310, train_loss=nan, test_loss=0.0000, LR=0.1000, time 158.55s
2024-12-03 12:52:43,991 [INFO] 320, train_loss=nan, test_loss=0.0000, LR=0.1000, time 163.63s
2024-12-03 12:52:49,064 [INFO] 330, train_loss=nan, test_loss=0.0000, LR=0.1000, time 168.70s
2024-12-03 12:52:54,048 [INFO] 340, train_loss=nan, test_loss=0.0000, LR=0.1000, time 173.68s
2024-12-03 12:52:59,019 [INFO] 350, train_loss=nan, test_loss=0.0000, LR=0.1000, time 178.65s
2024-12-03 12:53:04,085 [INFO] 360, train_loss=nan, test_loss=0.0000, LR=0.1000, time 183.72s
2024-12-03 12:53:09,124 [INFO] 370, train_loss=nan, test_loss=0.0000, LR=0.1000, time 188.76s
2024-12-03 12:53:14,248 [INFO] 380, train_loss=nan, test_loss=0.0000, LR=0.1000, time 193.88s
2024-12-03 12:53:19,416 [INFO] 390, train_loss=nan, test_loss=0.0000, LR=0.1000, time 199.05s
2024-12-03 12:53:24,425 [INFO] 400, train_loss=nan, test_loss=0.0000, LR=0.1000, time 204.06s
2024-12-03 12:53:29,634 [INFO] 410, train_loss=nan, test_loss=0.0000, LR=0.1000, time 209.27s
2024-12-03 12:53:34,740 [INFO] 420, train_loss=nan, test_loss=0.0000, LR=0.1000, time 214.37s
2024-12-03 12:53:39,817 [INFO] 430, train_loss=nan, test_loss=0.0000, LR=0.1000, time 219.45s
2024-12-03 12:53:44,851 [INFO] 440, train_loss=nan, test_loss=0.0000, LR=0.1000, time 224.49s
2024-12-03 12:53:49,923 [INFO] 450, train_loss=nan, test_loss=0.0000, LR=0.1000, time 229.56s
2024-12-03 12:53:55,020 [INFO] 460, train_loss=nan, test_loss=0.0000, LR=0.1000, time 234.65s
2024-12-03 12:54:00,040 [INFO] 470, train_loss=nan, test_loss=0.0000, LR=0.1000, time 239.67s
2024-12-03 12:54:05,191 [INFO] 480, train_loss=nan, test_loss=0.0000, LR=0.1000, time 244.82s
2024-12-03 12:54:10,331 [INFO] 490, train_loss=nan, test_loss=0.0000, LR=0.1000, time 249.97s
2024-12-03 12:54:15,347 [INFO] 500, train_loss=nan, test_loss=0.0000, LR=0.1000, time 254.98s
2024-12-03 12:54:20,658 [INFO] 510, train_loss=nan, test_loss=0.0000, LR=0.1000, time 260.29s
2024-12-03 12:54:25,725 [INFO] 520, train_loss=nan, test_loss=0.0000, LR=0.1000, time 265.36s
2024-12-03 12:54:30,811 [INFO] 530, train_loss=nan, test_loss=0.0000, LR=0.1000, time 270.44s
2024-12-03 12:54:35,895 [INFO] 540, train_loss=nan, test_loss=0.0000, LR=0.1000, time 275.53s
2024-12-03 12:54:41,073 [INFO] 550, train_loss=nan, test_loss=0.0000, LR=0.1000, time 280.71s
2024-12-03 12:54:46,296 [INFO] 560, train_loss=nan, test_loss=0.0000, LR=0.1000, time 285.93s
2024-12-03 12:54:51,456 [INFO] 570, train_loss=nan, test_loss=0.0000, LR=0.1000, time 291.09s
2024-12-03 12:54:56,469 [INFO] 580, train_loss=nan, test_loss=0.0000, LR=0.1000, time 296.10s
2024-12-03 12:55:01,617 [INFO] 590, train_loss=nan, test_loss=0.0000, LR=0.1000, time 301.25s
2024-12-03 12:55:06,719 [INFO] 600, train_loss=nan, test_loss=0.0000, LR=0.1000, time 306.35s
2024-12-03 12:55:11,825 [INFO] 610, train_loss=nan, test_loss=0.0000, LR=0.1000, time 311.46s
2024-12-03 12:55:16,819 [INFO] 620, train_loss=nan, test_loss=0.0000, LR=0.1000, time 316.45s
2024-12-03 12:55:21,880 [INFO] 630, train_loss=nan, test_loss=0.0000, LR=0.1000, time 321.51s
2024-12-03 12:55:27,092 [INFO] 640, train_loss=nan, test_loss=0.0000, LR=0.1000, time 326.73s
2024-12-03 12:55:32,194 [INFO] 650, train_loss=nan, test_loss=0.0000, LR=0.1000, time 331.83s
2024-12-03 12:55:37,367 [INFO] 660, train_loss=nan, test_loss=0.0000, LR=0.1000, time 337.00s
2024-12-03 12:55:42,365 [INFO] 670, train_loss=nan, test_loss=0.0000, LR=0.1000, time 342.00s
2024-12-03 12:55:47,405 [INFO] 680, train_loss=nan, test_loss=0.0000, LR=0.1000, time 347.04s
2024-12-03 12:55:52,401 [INFO] 690, train_loss=nan, test_loss=0.0000, LR=0.1000, time 352.04s
2024-12-03 12:55:57,513 [INFO] 700, train_loss=nan, test_loss=0.0000, LR=0.1000, time 357.15s
2024-12-03 12:56:02,576 [INFO] 710, train_loss=nan, test_loss=0.0000, LR=0.1000, time 362.21s
2024-12-03 12:56:07,571 [INFO] 720, train_loss=nan, test_loss=0.0000, LR=0.1000, time 367.20s
2024-12-03 12:56:12,654 [INFO] 730, train_loss=nan, test_loss=0.0000, LR=0.1000, time 372.29s
2024-12-03 12:56:17,701 [INFO] 740, train_loss=nan, test_loss=0.0000, LR=0.1000, time 377.34s
2024-12-03 12:56:22,759 [INFO] 750, train_loss=nan, test_loss=0.0000, LR=0.1000, time 382.39s
2024-12-03 12:56:27,838 [INFO] 760, train_loss=nan, test_loss=0.0000, LR=0.1000, time 387.47s
2024-12-03 12:56:32,834 [INFO] 770, train_loss=nan, test_loss=0.0000, LR=0.1000, time 392.47s
2024-12-03 12:56:37,782 [INFO] 780, train_loss=nan, test_loss=0.0000, LR=0.1000, time 397.42s
2024-12-03 12:56:42,877 [INFO] 790, train_loss=nan, test_loss=0.0000, LR=0.1000, time 402.51s
2024-12-03 12:56:48,049 [INFO] 800, train_loss=nan, test_loss=0.0000, LR=0.1000, time 407.68s
2024-12-03 12:56:53,209 [INFO] 810, train_loss=nan, test_loss=0.0000, LR=0.1000, time 412.84s
2024-12-03 12:56:58,218 [INFO] 820, train_loss=nan, test_loss=0.0000, LR=0.1000, time 417.85s
2024-12-03 12:57:03,236 [INFO] 830, train_loss=nan, test_loss=0.0000, LR=0.1000, time 422.87s
2024-12-03 12:57:08,401 [INFO] 840, train_loss=nan, test_loss=0.0000, LR=0.1000, time 428.04s
2024-12-03 12:57:13,435 [INFO] 850, train_loss=nan, test_loss=0.0000, LR=0.1000, time 433.07s
2024-12-03 12:57:18,535 [INFO] 860, train_loss=nan, test_loss=0.0000, LR=0.1000, time 438.17s
2024-12-03 12:57:23,577 [INFO] 870, train_loss=nan, test_loss=0.0000, LR=0.1000, time 443.21s
2024-12-03 12:57:28,553 [INFO] 880, train_loss=nan, test_loss=0.0000, LR=0.1000, time 448.19s
2024-12-03 12:57:33,504 [INFO] 890, train_loss=nan, test_loss=0.0000, LR=0.1000, time 453.14s
2024-12-03 12:57:38,488 [INFO] 900, train_loss=nan, test_loss=0.0000, LR=0.0500, time 458.12s
2024-12-03 12:57:43,568 [INFO] 910, train_loss=nan, test_loss=0.0000, LR=0.0250, time 463.20s
2024-12-03 12:57:48,638 [INFO] 920, train_loss=nan, test_loss=0.0000, LR=0.0125, time 468.27s
2024-12-03 12:57:53,791 [INFO] 930, train_loss=nan, test_loss=0.0000, LR=0.0063, time 473.43s
2024-12-03 12:57:58,935 [INFO] 940, train_loss=nan, test_loss=0.0000, LR=0.0031, time 478.57s
2024-12-03 12:58:03,986 [INFO] 950, train_loss=nan, test_loss=0.0000, LR=0.0016, time 483.62s
2024-12-03 12:58:09,090 [INFO] 960, train_loss=nan, test_loss=0.0000, LR=0.0008, time 488.72s
2024-12-03 12:58:14,165 [INFO] 970, train_loss=nan, test_loss=0.0000, LR=0.0004, time 493.80s
2024-12-03 12:58:19,211 [INFO] 980, train_loss=nan, test_loss=0.0000, LR=0.0002, time 498.85s
2024-12-03 12:58:24,203 [INFO] 990, train_loss=nan, test_loss=0.0000, LR=0.0001, time 503.84s
/Users/Alexander.Morano/Desktop/cellpose_training/old_plate/untitled_folder/models/CP_20241203_124811 copied to models folder /Users/Alexander.Morano/.cellpose/models GUI_INFO: selected model CP_20241203_124811, loading now 2024-12-03 12:58:28,847 [INFO] >> CP_20241203_124811 << model set to be used
2024-12-03 12:58:28,848 [INFO] ** TORCH MPS version installed and working. **
2024-12-03 12:58:28,848 [INFO] >>>> using GPU (MPS)
2024-12-03 12:58:28,921 [INFO] >>>> loading model /Users/Alexander.Morano/.cellpose/models/CP_20241203_124811
2024-12-03 12:58:28,983 [INFO] >>>> model diam_mean = 30.000 (ROIs rescaled to this size during training)
2024-12-03 12:58:28,983 [INFO] >>>> model diam_labels = 97.534 (mean diameter of training ROIs) GUI_INFO: diameter set to 97.53 (but can be changed)
2024-12-03 12:58:29,014 [INFO] >>>> diameter set to diam_labels ( = 97.534 )
2024-12-03 12:58:29,018 [WARNING] <tifffile.TiffFile 'scan_Plate_R_p00_0_C04f26d1.tif'> OME series is missing 1 frames. Missing data are zeroed GUI_INFO: restore: None GUI_INFO: normalization checked: computing saturation levels (and optionally filtered image) {'lowhigh': None, 'percentile': [1.0, 99.0], 'normalize': True, 'norm3D': False, 'sharpen_radius': 0, 'smooth_radius': 0, 'tile_norm_blocksize': 0, 'tile_norm_smooth3D': 1, 'invert': False} [0, 255.0] (1, 1536, 2048, 3) GUI_INFO: 42 masks found GUI_INFO: creating cellcolors and drawing masks GUI_INFO: loaded in previous changes
2024-12-03 12:58:29,389 [INFO] ** TORCH MPS version installed and working. **
2024-12-03 12:58:29,389 [INFO] >>>> using GPU (MPS)
2024-12-03 12:58:29,460 [INFO] >>>> loading model /Users/Alexander.Morano/.cellpose/models/CP_20241203_124811
2024-12-03 12:58:29,522 [INFO] >>>> model diam_mean = 30.000 (ROIs rescaled to this size during training)
2024-12-03 12:58:29,522 [INFO] >>>> model diam_labels = 97.534 (mean diameter of training ROIs) {'lowhigh': None, 'percentile': [1.0, 99.0], 'normalize': True, 'norm3D': False, 'sharpen_radius': 0, 'smooth_radius': 0, 'tile_norm_blocksize': 0, 'tile_norm_smooth3D': 1, 'invert': False}
2024-12-03 12:58:29,957 [INFO] No cell pixels found.
2024-12-03 12:58:30,178 [INFO] 0 cells found with model in 0.814 sec
GUI_INFO: 0 masks found
GUI_INFO: creating cellcolors and drawing masks
2024-12-03 12:58:30,237 [INFO] !!! computed masks for scan_Plate_R_p00_0_C04f26d1.tif from new model !!! ` I suspect the "train_loss=nan, test_loss=0.0000" output(s) have something to do with the failure, but I have no clue how to resolve that problem, or what could be causing it.
Hi,
I have similar training issue. I started with running cyto3 (the best performance model), then manually corrected the masks, and then train new model based on that. I did several iterations but still got no cell detection. I am not sure this is an innate bug or because of image features. Here is an example image used for training.
Looking forward to suggestions
Hi,
I have similar training issue. I started with running cyto3 (the best performance model), then manually corrected the masks, and then train new model based on that. I did several iterations but still got no cell detection. I am not sure this is an innate bug or because of image features. Here is an example image used for training.
Looking forward to suggestions
Hi! I figured out how to fix this: apparently this is a known problem in earlier versions of Cellpose. I started utilizing the most recent version on Github and unclicked / deactivated the GPU. running on CPU takes longer, but if I use the most recent version of cellpose and do not use GPU, It all works fine!!
Hi, I have similar training issue. I started with running cyto3 (the best performance model), then manually corrected the masks, and then train new model based on that. I did several iterations but still got no cell detection. I am not sure this is an innate bug or because of image features. Here is an example image used for training.
Looking forward to suggestions
Hi! I figured out how to fix this: apparently this is a known problem in earlier versions of Cellpose. I started utilizing the most recent version on Github and unclicked / deactivated the GPU. running on CPU takes longer, but if I use the most recent version of cellpose and do not use GPU, It all works fine!!
hello, My training works fine, but after training the initial cyto model, do you know how to load the trained model for evaluation, I tried the following code lines, unfortunately they are not working
`model_path_trained = "Image Analysis/trained_models/models/cyto3_finetuned_new"
model_trained= models.CellposeModel(gpu=False, pretrained_model=model_path_trained)
pred_masks= model_trained.eval(test_images, channels=channels, flow_threshold=flow_threshold, cellprob_threshold=cellprob_threshold,diameter=30)[0]`
these code lines do not work, since the trained model can't capture any ROIs during the evaluation procedure
Hi, I have similar training issue. I started with running cyto3 (the best performance model), then manually corrected the masks, and then train new model based on that. I did several iterations but still got no cell detection. I am not sure this is an innate bug or because of image features. Here is an example image used for training.
Hi! I figured out how to fix this: apparently this is a known problem in earlier versions of Cellpose. I started utilizing the most recent version on Github and unclicked / deactivated the GPU. running on CPU takes longer, but if I use the most recent version of cellpose and do not use GPU, It all works fine!!
hello, My training works fine, but after training the initial cyto model, do you know how to load the trained model for evaluation, I tried the following code lines, unfortunately they are not working
`model_path_trained = "Image Analysis/trained_models/models/cyto3_finetuned_new"
model_trained= models.CellposeModel(gpu=False, pretrained_model=model_path_trained)
pred_masks= model_trained.eval(test_images, channels=channels, flow_threshold=flow_threshold, cellprob_threshold=cellprob_threshold,diameter=30)[0]`
these code lines do not work, since the trained model can't capture any ROIs during the evaluation procedure
Hi Jiadalee,
Here's the code I use to run cellpose in python.
import numpy as np import matplotlib.pyplot as plt from cellpose import models, io, utils from cellpose.io import imread import os
modelpath = "/Users/Alexander.Morano/Desktop/cellpose_training/10x training/models/10X_NewCyto" filepath = "/Users/Alexander.Morano/Desktop/data/2024_12_19_GPCs_final_test/remnantplate/d0/"
os.chdir(filepath) files = os.listdir() print(len(files))
model = models.CellposeModel(gpu = False, pretrained_model = modelpath) channels = [[0,0]]
for i in files: image = io.imread(i) masks, flows, styles = model.eval(image, diameter = 35.0, channels = channels) io.save_rois(masks, i)
I can't tell if this is a problem with your model or with your code. Try loading up the CellPose GUI, UNCLICK the "use GPU" button, and see if you can detect ROIs with the new model on one image using the GUI. If you cannot, it is a problem with generating the trained model. But if you can, then it is a problem with the code, which is easier to fix.
First, is your model path working? In your terminal, if you run cd and put in "Image Analysis/trained_models/models/ ", does that take you to the models directory? I usually have to start at least from the Users base folder when I am writing out the path. If you're running it in Python and not CLI, try laying it out like I have here if you're using it in python?
@duckyaisha I would appreciated it if you could post you training codes for reference. I tried your method and I even cannot detect the ROIs with the new model using the GUI for n one image. I believe this is the problem of generating the trained model. I'm using Jupiter notebooks from the cloud and model path is a relative path. The path should not be an issue as it works fine. I posted my training codes below. Feel free to share your thoughts.
`
model_path_pretrained = "Microscopy/Image Analysis/trained_models/models/cyto3" model_pretrained = models.CellposeModel(gpu=False, model_type="cyto3") logger = io.logger_setup() channels = [2,1] model_name = "cyto3_finetuned_new"
new_model_path, train_losses, val_losses = train.train_seg(model_pretrained.net, train_data=train_images, train_labels=train_masks, test_data=val_images, test_labels=val_masks, channels=channels, save_path=model_dir, n_epochs=50, learning_rate=0.01, weight_decay=0.0001, SGD=True,
nimg_per_epoch=16, model_name=model_name, min_train_masks=0)
`
The training processing info is as follows:
`
2025-02-03 10:48:15,856 [INFO] >> cyto3 << model set to be used 2025-02-03 10:48:15,858 [INFO] ** TORCH CUDA version installed and working. ** 2025-02-03 10:48:15,859 [INFO] >>>> using GPU (CUDA) 2025-02-03 10:48:15,940 [INFO] >>>> loading model /home/jiada_li/.cellpose/models/cyto3 2025-02-03 10:48:16,022 [INFO] >>>> model diam_mean = 30.000 (ROIs rescaled to this size during training)
2025-02-03 10:48:23,082 [INFO] WRITING LOG OUTPUT TO /home/jiada_li/.cellpose/run.log 2025-02-03 10:48:23,083 [INFO] cellpose version: 3.1.0 platform: linux python version: 3.12.2 torch version: 2.6.0+cu124 2025-02-03 10:48:23,084 [INFO] flows precomputed 2025-02-03 10:48:23,091 [INFO] flows precomputed 2025-02-03 10:48:23,092 [INFO] >>> computing diameters 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 99/99 [00:00<00:00, 5328.80it/s] 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 11/11 [00:00<00:00, 4337.44it/s] 2025-02-03 10:48:23,118 [INFO] >>> using channels [2, 1] 2025-02-03 10:48:23,118 [INFO] >>> normalizing {'lowhigh': None, 'percentile': None, 'normalize': True, 'norm3D': False, 'sharpen_radius': 0, 'smooth_radius': 0, 'tile_norm_blocksize': 0, 'tile_norm_smooth3D': 1, 'invert': False} 2025-02-03 10:48:23,229 [INFO] >>> n_epochs=50, n_train=99, n_test=11 2025-02-03 10:48:23,229 [INFO] >>> SGD, learning_rate=0.01000, weight_decay=0.00010, momentum=0.900 2025-02-03 10:48:23,231 [INFO] >>> saving model to Microscopy/Image Analysis/trained_models/models/cyto3_finetuned_new
2025-02-03 10:48:23,426 [INFO] 0, train_loss=0.7856, test_loss=0.5658, LR=0.000000, time 0.19s 2025-02-03 10:48:24,124 [INFO] 5, train_loss=0.7441, test_loss=0.5198, LR=0.005556, time 0.89s 2025-02-03 10:48:24,822 [INFO] 10, train_loss=0.6728, test_loss=0.4600, LR=0.010000, time 1.59s 2025-02-03 10:48:26,178 [INFO] 20, train_loss=0.5037, test_loss=0.4434, LR=0.010000, time 2.95s 2025-02-03 10:48:27,531 [INFO] 30, train_loss=0.4266, test_loss=0.4271, LR=0.010000, time 4.30s 2025-02-03 10:48:28,881 [INFO] 40, train_loss=0.4520, test_loss=0.4364, LR=0.010000, time 5.65s 2025-02-03 10:48:30,059 [INFO] saving network parameters to Microscopy/Image Analysis/trained_models/models/cyto3_finetuned_new Diameter of labels in training images: 12.647065162658691
`
sorry there have been issues with this @duckyaisha , I think the training may still not be working on MPS. I don't currently have a Mac to test.
@sun1000yao could you please post what was written on the terminal during training? (with the version info at the beginning of the output)
@Jiadalee we can keep discussing in the other issue
Don't worry, I figured it out by just using the newest version of cellpose from github and disabling the CPU!
On Fri, Feb 7, 2025 at 12:57 PM carsen-stringer @.***> wrote:
sorry there have been issues with this @duckyaisha https://github.com/duckyaisha , I think the training may still not be working on MPS. I don't currently have a Mac to test.
@sun1000yao https://github.com/sun1000yao could you please post what was written on the terminal during training? (with the version info at the beginning of the output)
@Jiadalee https://github.com/Jiadalee we can keep discussing in the other issue
— Reply to this email directly, view it on GitHub https://github.com/MouseLand/cellpose/issues/1071#issuecomment-2643617443, or unsubscribe https://github.com/notifications/unsubscribe-auth/BNNTVHK4VFJZAOPWXTZPR5L2OTXYLAVCNFSM6AAAAABS6PSMWCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNBTGYYTONBUGM . You are receiving this because you were mentioned.Message ID: @.***>
--
PLEASE NOTE: This message, including any attachments, may include privileged, confidential and/or inside information belonging to IPI. Any distribution or use of this communication by anyone other than the intended recipient(s) is strictly prohibited and may be unlawful. If you are not the intended recipient, please notify the sender by replying to this message and then delete it from your system. Thank you.
Don't worry, I figured it out by just using the newest version of cellpose from github and disabling the CPU! …
@duckyaisha disabling CPU? isn't GPU?
i meant GPU my apologies,
On Fri, Feb 7, 2025 at 2:19 PM Jiada Li @.***> wrote:
Don't worry, I figured it out by just using the newest version of cellpose from github and disabling the CPU! … <#m_5056221544094689457_>
@duckyaisha https://github.com/duckyaisha disabling CPU? isn't GPU?
— Reply to this email directly, view it on GitHub https://github.com/MouseLand/cellpose/issues/1071#issuecomment-2643822159, or unsubscribe https://github.com/notifications/unsubscribe-auth/BNNTVHM6AQUMDB7QIIS3NTL2OUBLHAVCNFSM6AAAAABS6PSMWCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDMNBTHAZDEMJVHE . You are receiving this because you were mentioned.Message ID: @.***>
--
PLEASE NOTE: This message, including any attachments, may include privileged, confidential and/or inside information belonging to IPI. Any distribution or use of this communication by anyone other than the intended recipient(s) is strictly prohibited and may be unlawful. If you are not the intended recipient, please notify the sender by replying to this message and then delete it from your system. Thank you.
okay thanks for letting us know I will put some documentation on this