search

experiencor / keras-yolo3

Training and Detecting Objects with YOLO3
MIT License
1.6k stars 862 forks source link

How to train from scratch with own data set? #194

Open ZooZoo-tc opened 5 years ago

ZooZoo-tc commented 5 years ago

Thank you for providing repo... @experiencor could you give some suggestion about training from scratch with own data set ?

zhiyilearn commented 5 years ago

I have same question.

kabrau commented 5 years ago

I can help you, are many steps, you can read this link https://github.com/experiencor/keras-yolo2, it is the same with little differences. If you can't understand anyway, tell me and I will create a video for you.

zhiyilearn commented 5 years ago

If I use my own dataset to train YOLO v3, the step to label images. From the tutorial: train annotations in VOC format How to make train annotations in VOC format? What's exact VOC format.

kabrau commented 5 years ago

There are many options to annotate this I made my tool: https://github.com/kabrau/PyImageRoi (CreateBoundingBoxes) But many people use: https://github.com/tzutalin/labelImg

ZooZoo-tc commented 5 years ago

I can help you, are many steps, you can read this link https://github.com/experiencor/keras-yolo2, it is the same with little differences. If you can't understand anyway, tell me and I will create a video for you.

Hi @kabrau, I have 32x32 size images in my data set. I want to train yolo with whole(32x32 image itself) image... in my case whole image is the bounding box coordinates.... is it possible ? will it work?

kabrau commented 5 years ago

Hi @satish4github I don't know, what is the length of an object inside an image? In your case, would not a classifier be better?

ZooZoo-tc commented 5 years ago

Object size is same as image size(32x32) like cifar dataset Classifier would say only object is there or not... But i need the object localization details also while predicting..

kabrau commented 5 years ago

@satish4github I not sure, but I think that implementation does not work for you. I'm curious, I don't understand If the object has the same size, why you need localization?

zhiyilearn commented 5 years ago

With raccoon data set I got errors, any idea?

python train.py -c config.json

Using TensorFlow backend. Traceback (most recent call last): File "train.py", line 280, in main(args) File "train.py", line 170, in main config = json.loads(config_buffer.read()) File "/opt/apps/Anaconda3/2019.03/envs/powerai16_ibm/lib/python3.6/json/init.py", line 354, in loads return _default_decoder.decode(s) File "/opt/apps/Anaconda3/2019.03/envs/powerai16_ibm/lib/python3.6/json/decoder.py", line 339, in decode obj, end = self.raw_decode(s, idx=_w(s, 0).end()) File "/opt/apps/Anaconda3/2019.03/envs/powerai16_ibm/lib/python3.6/json/decoder.py", line 355, in raw_decode obj, end = self.scan_once(s, idx) json.decoder.JSONDecodeError: Expecting property name enclosed in double quotes: line 13 column 49 (char 505)

kabrau commented 5 years ago

You need to debug this line to see the problem File "/opt/apps/Anaconda3/2019.03/envs/powerai16_ibm/lib/python3.6/json/decoder.py", line 355

ZooZoo-tc commented 5 years ago

I will check...

I not sure, but I think that implementation does not work for you.

Basically i want train model with CIFAR image data set . I want to use the same model for predicting bounding boxes/object localization.

  1. In this case while training i don't need to need to do create bounding boxes while training data set.
  2. Memory
zhiyilearn commented 5 years ago

Download link for backend weight not available:

Download pretrained weights for backend at: https://1drv.ms/u/s!ApLdDEW3ut5fgQXa7GzSlG-mdza6

Doesn't matter, fixed.

zhiyilearn commented 5 years ago

Annotation file with YOLO v3 format like this is OK?

15 0.927600 0.282065 0.021600 0.035870 15 0.894800 0.274457 0.021600 0.027174 15 0.857200 0.265217 0.024800 0.041304 15 0.822000 0.247826 0.021600 0.036957

zhiyilearn commented 5 years ago

When I run training process for raccoon dataset as: python train.py -c config.json I got training error like this:

Epoch 00027: loss did not improve from 5.13942 Epoch 00027: early stopping /opt/apps/Anaconda3/2019.03/envs/powerai16_ibm/lib/python3.6/site-packages/Keras-2.2.4-py3.6.egg/keras/engine/saving.py:310: UserWarning: No training configuration found in save file: the model was not compiled. Compile it manually. warnings.warn('No training configuration found in save file: ' raccoon: 0.9681 mAP: 0.9681

Someone have any idea? My configuration is wrong?

look4pritam commented 5 years ago

You are getting 0.9681 mAP, which is good.

zhiyilearn commented 5 years ago

Training errors when use raccoon dataset: It used to train process work. Any hint? python train.py -c config.json

Using TensorFlow backend. valid_annot_folder not exists. Spliting the trainining set. Seen labels: {}

Given labels: ['raccoon'] Some labels have no annotations! Please revise the list of labels in the config.json. Traceback (most recent call last): File "train.py", line 280, in main(args) File "train.py", line 182, in main config['model']['labels'] ValueError: not enough values to unpack (expected 4, got 3)

config.json file looks like: { "model" : { "min_input_size": 352, "max_input_size": 448, "anchors": [10,13, 16,30, 33,23, 30,61, 62,45, 59,119, 116,90, 156,198, 373,326], "labels": ["raccoon"] },

"train": {
    "train_image_folder":   "/home/zli04/2019SummerWork/raccoon_datast-master/images/",
    "train_annot_folder":   "/home/zli04/2019SummerWork/raccoon_dataset-master/annotations/",
    "cache_name":           "raccoon_train.pkl",

    "train_times":          3,
    "batch_size":           16,
    "learning_rate":        1e-4,
    "nb_epochs":            100,
    "warmup_epochs":        3,
    "ignore_thresh":        0.5,
    "gpus":                 "0,1",

    "grid_scales":          [1,1,1],
    "obj_scale":            5,
    "noobj_scale":          1,
    "xywh_scale":           1,
    "class_scale":          1,

    "tensorboard_dir":      "log_raccoon",
    "saved_weights_name":   "raccoon.h5",
    "debug":                true
},

"valid": {
    "valid_image_folder":   "",
    "valid_annot_folder":   "",
    "cache_name":           "",

    "valid_times":          1
}

}

zhiyilearn commented 5 years ago

I try to train with my own dataset with YOLO version like:

15 0.514000 0.254891 0.013600 0.029348

In config.json file, should I need update/change anything? I noticed label option, should I change my label to 15 instead?

zhiyilearn commented 5 years ago

When I use annotation tools labelImg to annotate box for my own dataset, should I first change the images size to standard such as 288 x488 for raccoon? My own dataset image size is: 1250 x 920. I noticed Pascal/VOC format has image size information.

look4pritam commented 5 years ago

Keep image size and bounding boxes same as original image size. There is no need to change image size to any standard size. Data augmentation tool will take care of converting image size to standard size and so on.

On Sat, Jul 27, 2019, 02:03 Zhiyi Li notifications@github.com wrote:

When I use annotation tools labelImg to annotate box for my own dataset, should I first change the images size to standard such as 288 x488 for raccoon? My own dataset image size is: 1250 x 920. I noticed Pascal/VOC format has image size information.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/experiencor/keras-yolo3/issues/194?email_source=notifications&email_token=ABJOVYQRX4VO5Z2GQZU75EDQBNNRFA5CNFSM4H6SLIXKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOD25UTHI#issuecomment-515590557, or mute the thread https://github.com/notifications/unsubscribe-auth/ABJOVYTTSHFEHHB6ZUNUURLQBNNRFANCNFSM4H6SLIXA .

zhiyilearn commented 5 years ago

How to set up min_input_size, max_input_size in config.json for your own dataset?

zhiyilearn commented 5 years ago

How to specify input_size in config.json file? Assume my own image size is 1250 x 720.

look4pritam commented 5 years ago

There is no need to specify image size in config file. but you need to specify model size in the form of min and max model size.

waelausy commented 5 years ago

hello, what are the modifications that must be done in the json file to support multiple classes custom training ?

tejasmagia commented 5 years ago

You are getting 0.9681 mAP, which is good.

If I am getting mAP: 0.0981 is it consider good? I have trained my data against this model.

look4pritam commented 5 years ago

Accuracy is less. How many input images are there?

On Thu, Oct 31, 2019, 15:33 tejasmagia notifications@github.com wrote:

You are getting 0.9681 mAP, which is good.

If I am getting mAP: 0.0981 is it consider good? I have trained my data against this model.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/experiencor/keras-yolo3/issues/194?email_source=notifications&email_token=ABJOVYQFHMGHXFRANFJPL7LQRKUN5A5CNFSM4H6SLIXKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOECXFIIY#issuecomment-548295715, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABJOVYVRWKTSKX4GY753NSTQRKUN5ANCNFSM4H6SLIXA .

tejasmagia commented 5 years ago

Around 125 images but still my objects are accurately detected.

On Thu, 31 Oct, 2019, 19:10 look4pritam, notifications@github.com wrote:

Accuracy is less. How many input images are there?

On Thu, Oct 31, 2019, 15:33 tejasmagia notifications@github.com wrote:

You are getting 0.9681 mAP, which is good.

If I am getting mAP: 0.0981 is it consider good? I have trained my data against this model.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub < https://github.com/experiencor/keras-yolo3/issues/194?email_source=notifications&email_token=ABJOVYQFHMGHXFRANFJPL7LQRKUN5A5CNFSM4H6SLIXKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOECXFIIY#issuecomment-548295715 , or unsubscribe < https://github.com/notifications/unsubscribe-auth/ABJOVYVRWKTSKX4GY753NSTQRKUN5ANCNFSM4H6SLIXA

.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/experiencor/keras-yolo3/issues/194?email_source=notifications&email_token=ADURW3OOMXJG3JVWE2EBRFDQRLN3LA5CNFSM4H6SLIXKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOECXZZRQ#issuecomment-548379846, or unsubscribe https://github.com/notifications/unsubscribe-auth/ADURW3NUMM5ZSIOY3P2ONGDQRLN3LANCNFSM4H6SLIXA .