different tensorRT vesrion int8 engine is not the same

my caffemodel is 42m ,in tensorRT2 int8 engine is about 10m ,but in tensorRT3 and tensorRT4,the int8 engine is about 32m,why? the mnist.caffemodel has the same problem, someone can help me?


Are you saying your model file size increased from 10mb to 32mb when migrating trt2 to trt3 and trt4? It’s hard to quantify if this is expected. Engine serialization sizes are affected by many variables (due to variability in layer timings, etc).

thanks for your reply