heatmap – how to create test_model when using LSTM to predict

I trained a model which predicted heatmaps of the future three-days(6, 7, 8) based on previous six-days(0, 1, 2, 3, 4, 5).
Part1: the six-days’ heatmaps were sended to cnn [input channel is 6] then to LSTM;
Part2: the output of part1 will be part inputs of part2. the rest inputs of part2 are heatmaps of day5, day6, day7(which are sended to cnn[input channel is 3]). Noted, in training model, heatmaps of day 5, 6, 7 are truth value. finally, the outputs of part2 are heatmaps of the future three-days(6, 7, 8)

my problem is : when I tested the model, the part1’s inputs could be six-days’ heatmaps, but the part2’s input cannot be truth value; it’s a sequence model, which means day5 should print pre_day6 that is input of next cell, then using pre_day6 to predict pre_day7 . But I have trouble in organizing the inputs.

I got trained model already, should I save the parameters in trained model? How can I get a cell’s output meanwhiel using those parameters trained.enter image description here