My training is failing despite my images looking fine

[sp01186]

My training isn't doing what I want it to do and I don't understand why, I was hoping I could get some explanation or support with this from this forum.

As you can see from the image above, the data seems to be okay in its identification of the bounding boxes.

The training code is shown here.

However after I train them, the predictions I am getting are all over the place.

Please let me know if there are any suggestions or solutions.

[ExtReMLapin]

Can you share your training json file ? and the images ?

[sp01186]

It turns out that the problem was that I didn't have enough training iterations. I thought that it would at least give something that wasn't completely random if I used 300 iterations,.

The training data can be obtained with the following:

!pip install "opencv-python-headless<4.3"
!pip install --upgrade pip setuptools wheel
!pip install fiftyone

import fiftyone as fo
import fiftyone.zoo as foz
import numpy as np
import tensorflow as tf

Download and load the validation split of Open Images V6

#dataset = foz.load_zoo_dataset("open-images-v6", split="validation") Full dataset.
dataset = foz.load_zoo_dataset(
"open-images-v6",
split="train",
label_types=["detections", "classifications", "segmentations"],
classes=["Airplane","Cat", "Dog"],
max_samples=20000
)

Converting the dataset into the detectron2 format

from detectron2.structures import BoxMode

def get_data_dicts(dataset):

dataset.compute_metadata()
dataset_dicts = []

for sample in dataset.iter_samples():
# for idx, v in enumerate(imgs_anns.values()):
    record = {}

    filename = sample.filepath
    idx = sample.id
    
    record["file_name"] = filename # found
    record["image_id"] = idx # found
    

    detections = sample.detections;

    classes = ['Airplane', 'Dog', 'Cat'];

    
    objs = []
    for i in detections.detections:
    # print(i)
    # print(i.label)
      if i.label in classes:
        coors = []
        category = 0
        if i.label == "Dog":
          category = 1
        if i.label == "Cat":
          category = 2
        #print(i.bounding_box) #will need to convert from relative to pixel coordinates
        for indice in range(4):
          if (indice == 0 or indice == 2):
            coors.append(i.bounding_box[indice]*sample.metadata.width)
          else:
            coors.append(i.bounding_box[indice]*sample.metadata.height)
        obj = {
            "bbox": coors, # found
            "bbox_mode": BoxMode.XYWH_ABS, #default
            # "segmentation": [poly],
            "category_id": category,
          }
        objs.append(obj)
    record["annotations"] = objs
    dataset_dicts.append(record)

return dataset_dicts