Object Detection Tensorflow 2 - Evaluate model, determine "best" model -- 2

An Object Detection was done with Tensorflow 2. Several models were calculated. The task is to load the models or checkpoints and evaluate images or tfrecords and calculate appropriate parameters to describe the goodness of the model to decide which model is better. -------------------------------- The data used for this example is the bloodcell data. Two models (A (Res_2021-06-04_17-24-57) and B (Res_2021-06-07_12-03-36)) were calculated using Tensorflow 2.3 Object detection API. The models were calculated with [login to view URL] --pipeline_config_path=Res_2021-06-04_17-24-57\\[login to view URL] --model_dir=Res_2021-06-04_17-24-57\\training --alsologtostderr --num_train_steps=1000000 --sample_1_of_n_eval_examples=1 --num_eval_steps=1000 [login to view URL] --pipeline_config_path=Res_2021-06-07_12-03-36\\[login to view URL] --model_dir=Res_2021-06-07_12-03-36\\training --alsologtostderr --num_train_steps=1000000 --sample_1_of_n_eval_examples=1 --num_eval_steps=1000 -------------------------------- The files can be downloaded zip archive from: [login to view URL] (contains the [login to view URL] and for each of the models Res_2021-06-04_17-24-57 and Res_2021-06-07_12-03-36: - [login to view URL] (which was used for training, see above) - training checkpoint models (result of/during training) ) The data can be found in the zip archive (from [login to view URL]): [login to view URL] (contains the tfrecords, the *.pbtxt) Paths to the data have to be changed in the files. The base model can be downloaded here: modelA: [login to view URL] modelB: [login to view URL] -------------------------------- The following tasks should be done: 1) Model a) Load the model from a checkpoint and save it as a "saved model" which can be used without the source b) Load the model "saved model" 2) Use model a) Calculate single images (*.jpg) from a directory with the models b) Calculate data from tfrecord with the models In both cases images shall be saved with rectangles of objects drawn into the image (different colors for the different objects, probability printed alon the rectangle9. 3) Calculate parameters Use tfrecord to calculate parameters to describe goodness of the model (with the aim to decide which model is better). Therefore [login to view URL] should be used (see [login to view URL]) . -------------------------------- All calculations have to be done in Python (3.7) and must run under Windows 10. Documented code has to be delivered.