do_pytorch.py calls loss = model(data, target) (not out = model(data); loss = criterion(out, target) ). This is made possible using the stuff in helpers_pytorch.py which wrap the "raw" pytorch model, into something that does specific to the problem (i.e. applies the right loss).
do_pytorch.pycallsloss = model(data, target)(notout = model(data); loss = criterion(out, target)). This is made possible using the stuff inhelpers_pytorch.pywhich wrap the "raw" pytorch model, into something that does specific to the problem (i.e. applies the right loss).This is not done for SSD300...