Are the dimensions on name model related to the accuracy of training?

[Kirito_Kirigaya]

I tried a simple object training with https://github.com/tensorflow/models/blob/master/research/object_detection/model_main_tf2.py using several pre-trained models and my own images, but after hours of training the learning_rate is zero and loop continues for ever.

So I wonder if the dimensions in the name of models and my images are the problem. Used ssd models:

• ssd_mobilenet_v2_320x320_coco17_tpu-8
• mask_rcnn_inception_resnet_v2_1024x1024_coco17_gpu-8
• centernet_hg104_512x512_coco17_tpu-8
• and other from here https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_detection_zoo.md

As you can see, the models have a dimension on its name. But my images don’t have that dimensions.

So my question is: If the model has 320x320 on its name, Should the images to be used in the training, have that dimension?

Thanks

[Laxma_Reddy_Patlolla]

Hi @Kirito_Kirigaya ,

While the dimension mentioned in the model’s name (e.g., ssd_mobilenet_v2_320x320_coco17_tpu-8) typically refers to the resolution at which the model was trained or designed to operate optimally, it doesn’t necessarily dictate the dimensions of the images you use for training.

For optimal training results, it’s crucial to resize your training images to match those dimensions.

Remember, using images with different dimensions than the expected input size can lead to inaccurate training and suboptimal performance. By ensuring consistency between your image dimensions and the model’s expectations, you can achieve better results and avoid issues like the learning rate freezing at zero

Thanks.