Karthik-Suresh93 commented 6 years ago

Hi,

I am training YOLOv3 on my custom dataset. I trained for about 3400 iterations and the error seems to be converging at around 0.3. However, when I run the map test with ./darknet detector map... I get the results as shown below:

for thresh = 0.25, precision = -nan, recall = 0.00, F1-score = -nan for thresh = 0.25, TP = 0, FP = 0, FN = 42846, average IoU = -nan %

mean average precision (mAP) = 0.000000, or 0.00 % Total Detection Time: 53.000000 Seconds

I don't understand why this is happening. My data file looks like this:

classes= 12 train = data/train.txt valid = data/test.txt names = data/obj.names backup = backup/

Please help me and let me know where the training is going wrong

AlexeyAB commented 6 years ago

Hi, Try to set valid = data/train.txt And provide more information, did you use Yolo_mark, what did you change in the cfg-file...

Karthik-Suresh93 commented 6 years ago

Hi,

I tried to set valid = data/train.txt and got the same results.

I suspect it is because of the annotations. I used a custom dataset where they had given: <x>, <y>, <w>, <h> where x and y are the x-y coordinates of the top left corner of the bounding box. [I know yolo expects the center coordinates of the bounding box]

I used the following code to convert it into yolo format:

`def convert(size, box):

dw = 1./size[0]
dh = 1./size[1]

##box[2] is the width of the bounding box, box[3] is the height of the bounding box. 
##box[0] and box[1] are the x and y coordinates of the top left corners of the bounding box respectively

##convert from top left coordiantes to center coordinates

box[0] = float(box[0]) + float(box[2])/2.0
box[1] = float(box[1]) + float(box[3])/2.0

x = box[0]*dw
y = box[1]*dh
w = float(box[2])*dw
h = float(box[3])*dh
return (x,y,w,h)`

After using this code, I was able to convert all numbers between 0 and 1 as mentioned in your github page. The link to the data format is here: Please let me know if there is any mistake in my approach

Thanks and Regards, Karthik

AlexeyAB commented 6 years ago

Open your dataset in the https://github.com/AlexeyAB/Yolo_mark

Karthik-Suresh93 commented 6 years ago

By open my dataset do you mean open a new issue with the link of my data in the yolo_mark link you've provided?

AlexeyAB commented 6 years ago

No, I mean install Yolo_mark and run command: ./yolo_mark ./img ./data/train.txt ./data/obj.names

And you will see whether you marked the objects correctly.

Karthik-Suresh93 commented 6 years ago

I tried the Yolo_mark and all the objects are approximately correctly marked.

AlexeyAB commented 6 years ago

@Karthik-Suresh93 Can you detect anything using your cfg/weights? What batch and subdivision did you use for training? What also did you change in your cfg file?

Karthik-Suresh93 commented 6 years ago

I figured out my mistake. I was training with the coordinates for the top left corner of the bounding box instead of the center. I get a map of 0.01% after 1000 iterations after I corrected this mistake. I will close this issue after checking the map again after a few thousand more iterations. Thank you very much for your help

AlexeyAB commented 6 years ago

Thats why I always recommend to use https://github.com/AlexeyAB/Yolo_mark

Karthik-Suresh93 commented 6 years ago

The error seems to be stuck at 7.7-8 from the last 1000 iterations or so ( I am currently in iteration 3134, dataset has 12 classes and is a very difficult dataset). I am suspicious about the anchor sizes calculated by calc_anchors. Here is my output: anchors = 13.0788,16.8780, 31.4323,356.1869, 13.3417,964.0239, 289.3022,81.5903, 16.4985,1743.8701, 15.8474,2934.3330, 957.3582,210.2993, 1745.9518,205.4940, 2944.1074,208.2814

Aren't some anchors too big? Also, my image dimensions are varying in the training set, although all of them are greater than 448x448. Could you please let me know if there is an issue here?

AlexeyAB commented 6 years ago

Can you show entire [net] section from your cfg-file?

Karthik-Suresh93 commented 6 years ago

[net]

Testing

batch=1

subdivisions=1

Training

batch=64 subdivisions=16 width=416 height=416 channels=3 momentum=0.9 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1

learning_rate=0.001 burn_in=1000 max_batches = 50200 policy=steps steps=40000,45000 scales=.1,.1

[convolutional] batch_normalize=1 filters=32 size=3 stride=1 pad=1 activation=leaky

Downsample

[convolutional] batch_normalize=1 filters=64 size=3 stride=2 pad=1 activation=leaky

[convolutional] batch_normalize=1 filters=32 size=1 stride=1 pad=1 activation=leaky

[convolutional] batch_normalize=1 filters=64 size=3 stride=1 pad=1 activation=leaky