Train on my own datasets but error occurred: *** Check failure stack trace: ***

[Cogito2012]

I had followed this help https://github.com/zeyuanxy/fast-rcnn/tree/master/help/train, but when I run the file train_net.py with argumemt: --solver models/CaffeNet/solver.prototxt --weights data/imagenet_models/CaffeNet.v2.caffemodel --imdb pku_bigmm_train * Check failure stack trace: * occurred in train.py at line: self.solver = caffe.SGDSolver(solver_prototxt)

///////////// output in the console: ///////////// Called with args: Namespace(cfg_file=None, gpu_id=None, imdb_name='pku_bigmm_train', max_iters=40000, pretrained_model='data/imagenet_models/CaffeNet.v2.caffemodel', randomize=False, set_cfgs=None, solver='models/CaffeNet/solver.prototxt') Using config: {'DEDUP_BOXES': 0.0625, 'EPS': 1e-14, 'EXP_DIR': 'pku_bigmm', 'PIXEL_MEANS': array([[[ 102.9801, 115.9465, 122.7717]]]), 'RNG_SEED': 3, 'ROOT_DIR': 'E:\IIP-VS\ObjectDetection\Fast-RCNN\fast-rcnn-bwt', 'TEST': {'BBOX_REG': True, 'MAX_SIZE': 1000, 'NMS': 0.3, 'SCALES': [600], 'SVM': False}, 'TRAIN': {'BATCH_SIZE': 128, 'BBOX_REG': True, 'BBOX_THRESH': 0.5, 'BG_THRESH_HI': 0.5, 'BG_THRESH_LO': 0.1, 'FG_FRACTION': 0.25, 'FG_THRESH': 0.5, 'IMS_PER_BATCH': 2, 'MAX_SIZE': 1000, 'SCALES': [600], 'SNAPSHOT_INFIX': '', 'SNAPSHOT_ITERS': 10000, 'USE_FLIPPED': True, 'USE_PREFETCH': False}} Loaded dataset pku_bigmm_train for training Appending horizontally-flipped training examples... train ss roidb loaded from E:\IIP-VS\ObjectDetection\Fast-RCNN\fast-rcnn-bwt\data\cache\train_selective_search_roidb.pkl done Preparing training data... done Output will be saved to E:\IIP-VS\ObjectDetection\Fast-RCNN\fast-rcnn-bwt\output\pku_bigmm\train Computing bounding-box regression targets... done * Check failure stack trace: *

[nwestlake]

Are there images without any annotations/GT boxes?

[Cogito2012]

@nw362 No such images, all my training images had at lest one annotations/GT box

[nwestlake]

A different issue then. It's a shame it doesn't give a file name and line number to assist.

[Cogito2012]

@The error ocurred at the line 36 in the file ./lib/fast_rcnn/train.py line 36: self.solver = caffe.SGDSolver(solver_prototxt)

[nwestlake]

I don't think it knows about the dataset until line 46 but I may be wrong. Does it work if you change dataset back to voc? line 46: self.solver.net.layers[0].set_roidb(roidb)

But just in case it's the same issue I had, apply 75b3fdd083f0490244441933b96581dcea719873 which gives a nicer error message if it was this.

[Cogito2012]

@nw362 The same error happened when I run the original version with voc2007 datasets. I doubt the problem is that I'm not using rbg's caffe. My caffe comes from @happyear. (https://github.com/happynear/caffe-windows), Mybe his caffe is not support for fast-rcnn training...or my caffe compiling is not right.

[nwestlake]

Probably, especially if it's missing a layer that the fast-rcnn-caffe has!