0 box detect

[dr-askar]

thanx for this project i have tried raccoon train on tiny yolo and useed this config `{ "model" : { "backend": "Tiny Yolo", "input_size": 416, "anchors": [3.97,5.13, 5.55,9.46, 7.56,11.61, 9.65,8.51, 11.26,11.97], "max_box_per_image": 10,
"labels": ["raccoon"] },

"train": {
    "train_image_folder":   "/content/raccoon_dataset/train_image_folder/",
    "train_annot_folder":   "/content/raccoon_dataset/train_annot_folder/",     
      
    "train_times":          100,
    "pretrained_weights":   "tiny_yolo_r.h5",
    "batch_size":           32,
    "learning_rate":        1e-4,
    "nb_epochs":            1,
    "warmup_epochs":        3,

    "object_scale":         5.0 ,
    "no_object_scale":      1.0,
    "coord_scale":          1.0,
    "class_scale":          1.0,

    "saved_weights_name":   "tiny_yolo_r.h5",
    "debug":                true
},

"valid": {
    "valid_image_folder":   "/content/raccoon_dataset/valid_image_folder/",
    "valid_annot_folder":   "/content/raccoon_dataset/valid_annot_folder/",

    "valid_times":          5
}

} ` but always when predict got 0 boxes what is the problem??

[dr-askar]

i run frontend.py do warm up then do training with warm up 0 and used the wieght from warm up and still 0 box

[zenetio]

I am facing the same issue but using the blood cell detection. After 7-8 iterations I get an early stop and when I run the detection I get the same image with boxes=[]. Its weird because I am using the same notebook, the same dataset and just downloaded the yolo.weights. The explore check is working correctly but I am getting no prediction.

[werfxz]

I'm facing the same problem. map decreases zero and no boxes detected, recall also becomes zero.

[zenetio]

I saw in a thread that we need normalize both train and valid datasets. So I saw that train was not normalized and when I fixed it the model learned quick and is predicting correctly. So, it was fixed!

[dr-askar]

Thanx .. How can i normalize the datasets ?

[werfxz]

@zenetio Both valid_generator and train_generator takes norm=self.feature_extractor.normalize as a parameter. Why train data is not normalized but validation is? How can we fix this problem?

train_generator = BatchGenerator(train_imgs, generator_config, norm=self.feature_extractor.normalize) valid_generator = BatchGenerator(valid_imgs, generator_config, norm=self.feature_extractor.normalize, jitter=False)

[zenetio]

@werfxz I am guessing I am using a different implementation where, in the code, only validation was normalized. So, I just changed the code to accept normalize in train and the model learned and is predicting well.

[werfxz]

Can you share your code? I am running code in python 3.6 could it be the problem?

[zenetio]

I am using the Blood Cell Detection notebook as code base and you can find it in this current repo. Note the difference between train and valid calculation.

[MBoaretto25]

@zenetio I think all images are already normalized according to its respective feature extractor as can be seen in the frontend.py here.

@dr-askar have you tried raising the values of the scales?

[zenetio]

@MBoaretto25 I am not using frontend.py. I am using the notebook code and so I had to normalize my dataset.

[lenhatdong]

I have the same problem. Did you find the solution?

[zenetio]

As I said, I just normalized both train and valid datasets and the issue was fixed. My project is based on the Blood Cell project on this repo.

[lenhatdong]

Thanks for your answer, but I mean on frontend.py

[zenetio]

Sorry @mikado3119 but I am not using frontend.py. I am using the notebook code.

[eirini5th]

@zenetio how did you normalize the datasets? The notebook uses this function:

def normalize(image): return image / 255.

which is then used when creating the Batch Generators:

train_imgs, seen_train_labels = parse_annotation(train_annot_folder, train_image_folder, labels=LABELS) train_batch = BatchGenerator(train_imgs, generator_config, norm=normalize) valid_imgs, seen_valid_labels = parse_annotation(valid_annot_folder, valid_image_folder, labels=LABELS) valid_batch = BatchGenerator(valid_imgs, generator_config, norm=normalize, jitter=False)

Is this not enough for normalization?

[zenetio]

@eirini5th As you stated, I just passed the normalize function to both train and valid BatchGenerator and it was enough to make the model generalize properly. Before doing that, the model was not learning.

[eirini5th]

@zenetio thanks for your answer! Unfortunately in my case it still predicts 0 boxes even after using the normalize function. Only difference is I don't use a valid_batch, I wanted to first test it on a single image, which I am also normalizing before the prediction. Maybe I should try with an actual batch of validation images?

[zenetio]

eirini5th Note that the example works with the current dataset. So, I would suggest you run the notebook example and then start to replace parts of the example with your needs. Then when you see that it stops working, you will have a good idea of where the problem is located. For example, use train and validation. Is it working? Then remove validation and see what happens.

---

De: eirini5th @.> Enviado: terça-feira, 20 de julho de 2021 12:36 Para: experiencor/keras-yolo2 @.> Cc: carlos @.>; Mention @.> Assunto: Re: [experiencor/keras-yolo2] 0 box detect (#409)

@zenetiohttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fzenetio&data=04%7C01%7C%7C2b4964556de2404ae0bb08d94b7b07e1%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637623814054931235%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=nGgKWH7YsdsEmXS2j22D2ZhNzifUFDhOYWg7%2BTfKbtw%3D&reserved=0 thanks for your answer! Unfortunately in my case it still predicts 0 boxes even after using the normalize function. Only difference is I don't use a valid_batch, I wanted to first test it on a single image, which I am also normalizing before the prediction. Maybe I should try with an actual batch of validation images?

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHubhttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fexperiencor%2Fkeras-yolo2%2Fissues%2F409%23issuecomment-883357891&data=04%7C01%7C%7C2b4964556de2404ae0bb08d94b7b07e1%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637623814054941230%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=G47Q8NwEXPZZx5rAs2VfWg%2BN2TnE81r4nWU05UKZwO0%3D&reserved=0, or unsubscribehttps://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAAENGIKFFHAMNC3QSL7TZ6TTYVUVVANCNFSM4HB5TU2A&data=04%7C01%7C%7C2b4964556de2404ae0bb08d94b7b07e1%7C84df9e7fe9f640afb435aaaaaaaaaaaa%7C1%7C0%7C637623814054941230%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=uKgSYUe50FLLywLjfXGybPoHfDStOf1slOZoc0jM1hA%3D&reserved=0.