Hello MAFAT team,
I would like to inquire about the computation of runtime for the model's inference process. We have noted that the runtime limit is 2 seconds per image (1280X1280 pixels) times the number of images. However, we are unclear about how the runtime is actually computed.
As mentioned in the FAQ section, the model's inference process follows the below order of methods:
1. model.load(dir_path)
2. If the make_initial_pass_over_dataset field is set to True:
model.collect_statistics_iter(img, meta) - This method enables an initial pass over the test data (looping over pairs of frames and their corresponding metadata).
model.update_model() - After the initial pass, this function updates the model parameters.
3. model.predict(img, meta) - This method receives one image and its corresponding metadata at a time and returns predictions.
Could you please clarify which steps will be included in the runtime calculation? Will it be step 1, 2, and 3 altogether or step 2 and 3 only or just step 3?
Thank you for your kind assistance.
Best regards,
wangzhiyu918
Hi,
The runtime limit of 2 seconds per image (1280X1280 pixels) times the number of images refers to steps: 2 and 3.
We also took into account a proper additional time for model loading (up to 1 GB) - step 1.
MAFAT Challenge Team