I am new to Computer Vision and i was trying to run YOLOP using a custom dataset that i built. i have images of size 400*300 and the detection annotation labels are structured as attached below
11383.json
I am only training the encoder and detection heads, so, i have put _C.TRAIN.ENC_DET_ONLY = True and everything else to false. But to keep the dataset folders in line with the recommendation i have added other png images of size 400*300 which are completely black (i just made them using python to fill the folders) to the da_seg_annotation and ll_seg_annotation folders. so to summarize, i have actual frames in my images folder and their labels in the json file structured as above, then i have black png images o the same size as the ones in the image folder so that i dont have to modify much in the original code. But now while training i am getting a runtime warning: invalid value encountered in true_divide
IoU=intersection / union
and then when it starts to train,Epoch: [1][0/3275] Time 1488.698s (1488.698s) Speed 0.0 samples/s Data 27.704s (27.704s) Loss nan (nan)
NaN or Inf found in input tensor - is happening
my frames are of the type (this is a random image from my dataset and not associated with the label given above): -
and the frames which i used to fill the ll_seg_annotations and da_seg_annotations folder are as follows : -
could this be an issue that the images are very different even if i am not training the driving area and lane line segmentation heads?
i am also getting a warning that says WARNING: Extremely small objects found. 193 of 111455 labels are < 3 pixels in width or height. could this be related to the runtime warning: invalid value encountered in true_divide
IoU=intersection / union warning i am getting and also the speed is 0.0 samples/s and the loss is shown as nan (nan) NaN or Inf found in input tensor issue while training
I am new to Computer Vision and i was trying to run YOLOP using a custom dataset that i built. i have images of size 400*300 and the detection annotation labels are structured as attached below
11383.json
I am only training the encoder and detection heads, so, i have put _C.TRAIN.ENC_DET_ONLY = True and everything else to false. But to keep the dataset folders in line with the recommendation i have added other png images of size 400*300 which are completely black (i just made them using python to fill the folders) to the da_seg_annotation and ll_seg_annotation folders. so to summarize, i have actual frames in my images folder and their labels in the json file structured as above, then i have black png images o the same size as the ones in the image folder so that i dont have to modify much in the original code. But now while training i am getting a runtime warning: invalid value encountered in true_divide
IoU=intersection / union
and then when it starts to train,Epoch: [1][0/3275] Time 1488.698s (1488.698s) Speed 0.0 samples/s Data 27.704s (27.704s) Loss nan (nan)
NaN or Inf found in input tensor - is happening
my frames are of the type (this is a random image from my dataset and not associated with the label given above): -
and the frames which i used to fill the ll_seg_annotations and da_seg_annotations folder are as follows : -
could this be an issue that the images are very different even if i am not training the driving area and lane line segmentation heads?
i am also getting a warning that says WARNING: Extremely small objects found. 193 of 111455 labels are < 3 pixels in width or height. could this be related to the runtime warning: invalid value encountered in true_divide
IoU=intersection / union warning i am getting and also the speed is 0.0 samples/s and the loss is shown as nan (nan) NaN or Inf found in input tensor issue while training