def tomtom_refit(zipkmap):
tm.K = zipkmap[0]
kmap = zipkmap[1]
print('Currently refitting {} model with K={}'.format(tm.mtype, tm.K),
flush=True)
# each element in the second layer is a tensor nrep*nsample*datadim
stor1_seeds = []
stor1_maps = []
stor1_logprobs = []
for tn in np.arange(len(kmap)):
print('Where K = {} and sample size is {}'.format(
tm.K, n_sample_array[tn]),
flush=True)
# iterate through the first dimension of the tensor, fitting model for each layer
stor2_seeds = []
stor2_maps = []
stor2_logprobs = []
tens = kmap[tn]
for i in np.arange(tens.shape[0]):
if 'gr' in tm.mtype:
seed, mmap, mem, lp = tm.tomtom_svi(tens[i], print_fit=False)
elif 'di' in tm.mtype:
seed, mmap, lp = tm.tomtom_svi(tens[i], print_fit=False)
stor2_seeds.append(seed)
stor2_maps.append(detach_mmap(mmap))
stor2_logprobs.append(lp.detach())
stor1_seeds.append(stor2_seeds)
stor1_maps.append(stor2_maps)
stor1_logprobs.append(stor2_logprobs)
return stor1_seeds, stor1_maps, stor1_logprobs
def tomtom_refit(tens):
print('Where K = {} and sample size is {}'.format(tm.K, tens.shape[1]),flush=True)
# iterate through the first dimension of the tensor, fitting model for each layer
stor2_seeds = []
stor2_maps = []
stor2_logprobs = []
# tens = kmap[tn]
for i in np.arange(tens.shape[0]):
if 'gr' in tm.mtype:
seed, mmap, mem, lp = tm.tomtom_svi(tens[i], print_fit = False)
elif 'di' in tm.mtype:
seed, mmap, lp = tm.tomtom_svi(tens[i], print_fit = False)
stor2_seeds.append(seed)
stor2_maps.append(detach_mmap(mmap))
stor2_logprobs.append(lp.detach())
return stor2_seeds,stor2_maps,stor2_logprobs
def tomtom_refit(gendat, print_fit = False):
modrec_seeds = []
modrec_maps = []
modrec_logprobs = []
tm.K = 1
for kmap in gendat:
print('Currently refitting {} model with K={}'.format(tm.mtype, tm.K))
# each element in the second layer is a tensor nrep*nsample*datadim
stor1_seeds = []
stor1_maps = []
stor1_logprobs = []
for tn in np.arange(len(kmap)):
print('Where sample size is {}'.format(n_sample_array[tn]))
# iterate through the first dimension of the tensor, fitting model for each layer
stor2_seeds = []
stor2_maps = []
stor2_logprobs = []
tens = kmap[tn]
for i in np.arange(tens.shape[0]):
seed, mmap, mem, lp = tm.tomtom_svi(tens[i], print_fit = print_fit)
stor2_seeds.append(seed)
stor2_maps.append(mmap)
stor2_logprobs.append(lp)
stor1_seeds.append(stor2_seeds)
stor1_maps.append(stor2_maps)
stor1_logprobs.append(stor2_logprobs)
modrec_seeds.append(stor1_seeds)
modrec_maps.append(stor1_maps)
modrec_logprobs.append(stor1_logprobs)
tm.K += 1
return modrec_seeds, modrec_maps, modrec_logprobs
ttarg_raw_all_3d, ttarg_raw_noauto_3d, tavg_norm_all_3d,
tavg_norm_noauto_3d, tavg_raw_all_3d, tavg_raw_noauto_3d
] = pickle.load(f)
# set model fitting params
tm.K = 3
tm.mtype = 'group'
tm.target = 'self' # 'self','targ','avg'
tm.dtype = 'raw' # 'norm','raw'
tm.auto = 'noauto' # 'noauto','all'
tm.stickbreak = False
tm.optim = pyro.optim.Adam({'lr': 0.0005, 'betas': [0.8, 0.99]})
tm.elbo = TraceEnum_ELBO(max_plate_nesting=1)
dtname = 't{}_{}_{}_3d'.format(tm.target, tm.dtype, tm.auto)
data = globals()[dtname]
seed_grp, mapl_grp, mem_grp, lp_grp, guide_grp = tm.tomtom_svi(
data, return_guide=True)
# defining sparse-input model
@config_enumerate
def model_multi_obs_grp(obsmat):
# some parameters can be directly derived from the data passed
# K = 2
nparticipants = data.shape[0]
nfeatures = data.shape[1] # number of rows in each person's matrix
ncol = data.shape[2]
# Background probability of different groups
if tm.stickbreak:
# stick breaking process for assigning weights to groups
with pyro.plate("beta_plate", K - 1):
modrec_logprobs_self_norm_all_grp = []
# each element in the outermost list is all gendat for a singel k-MAP
tm.K = 1
for kmap in gendat_self_norm_all_grp:
print(tm.K)
# each element in the second layer is a tensor nrep*nsample*datadim
stor1_seeds_self_norm_all_grp = []
stor1_maps_self_norm_all_grp = []
stor1_logprobs_self_norm_all_grp = []
for tens in kmap:
# iterate through the first dimension of the tensor, fitting model for each layer
stor2_seeds_self_norm_all_grp = []
stor2_maps_self_norm_all_grp = []
stor2_logprobs_self_norm_all_grp = []
for i in np.arange(tens.shape[0]):
pyro.clear_param_store()
seed, mmap, mem, lp = tm.tomtom_svi(tens[i])
stor2_seeds_self_norm_all_grp.append(seed)
stor2_maps_self_norm_all_grp.append(mmap)
stor2_logprobs_self_norm_all_grp.append(lp)
stor1_seeds_self_norm_all_grp.append(stor2_seeds_self_norm_all_grp)
stor1_maps_self_norm_all_grp.append(stor2_maps_self_norm_all_grp)
stor1_logprobs_self_norm_all_grp.append(stor2_logprobs_self_norm_all_grp)
modrec_seeds_self_norm_all_grp.append(stor1_seeds_self_norm_all_grp)
modrec_maps_self_norm_all_grp.append(stor1_maps_self_norm_all_grp)
modrec_logprobs_self_norm_all_grp.append(stor1_logprobs_self_norm_all_grp)
tm.K += 1
with open('refit_mod_self_norm_all_grp','wb') as f:
pickle.dump([modrec_maps_self_norm_all_grp,modrec_seeds_self_norm_all_grp,modrec_logprobs_self_norm_all_grp],f)