options_ref['--freq'] = 4
options_ref = conf.setoptions(argv=None,kw=options_ref,configfile=configfile_ref)
if torch.cuda.is_available():
options_ref['--device'] = 'cuda'
else:
options_ref['--device'] = 'cpu'
globalnames_ref, callback_ref, model_ref, data_model_ref, sampling_ref, addnoise_ref = setenv.setenv(options_ref)
globals().update(globalnames_ref)
torch.cuda.manual_seed_all(globalnames_ref['torchseed'])
torch.manual_seed(globalnames_ref['torchseed'])
np.random.seed(globalnames_ref['npseed'])
# initialization of parameters
initparameters.initkernels(model_ref,'upwind')
initparameters.initexpr(model_ref, viscosity=viscosity, pattern=dataname)
# model_ref.polys[k].coeffs(iprint=1)
for poly in model_ref.polys:
poly.coeffs(iprint=1)
#%%
options_1 = {}
options_1['--name'] = 'cdr-noise0.001'
configfile_1 = 'checkpoint/'+options_1['--name']+'/options.yaml'
options_1 = conf.setoptions(argv=None,kw=None,configfile=configfile_1,isload=True)
if torch.cuda.is_available():
options_1['--device'] = 'cuda'
else:
options_1['--device'] = 'cpu'
options_ref['--batch_size'] = 2
options_ref['--data_timescheme'] = 'rk2'
options_ref['--channel_names'] = 'u'
options_ref = conf.setoptions(argv=None, kw=options_ref, configfile=configfile)
if torch.cuda.is_available():
options_ref['--device'] = 'cuda'
else:
options_ref['--device'] = 'cpu'
globalnames_ref, callback_ref, model_ref, data_model_ref, sampling_ref, addnoise_ref = setenv.setenv(
options_ref)
globals().update(globalnames_ref)
# initialization of parameters
initparameters.initkernels(model_ref)
initparameters.initexpr(model_ref, viscosity=viscosity, pattern=dataname)
# model_ref.polys[k].coeffs(iprint=1)
for poly in model_ref.polys:
poly.coeffs(iprint=1)
#%%
options_1 = {}
options_1['--name'] = 'heat-frozen-upwind-sparse0.005-noise0.001'
configfile_1 = 'checkpoint/' + options_1['--name'] + '/options.yaml'
options_1 = conf.setoptions(argv=None,
kw=None,
configfile=configfile_1,
isload=True)
if torch.cuda.is_available():
# }
options = conf.setoptions(argv=sys.argv[1:], kw=kw, configfile=None)
print(options)
globalnames, callback, model, data_model, sampling, addnoise = setenv.setenv(
options)
globals().update(globalnames)
torch.cuda.manual_seed_all(torchseed)
torch.manual_seed(torchseed)
np.random.seed(npseed)
# initialization of parameters
if start_from < 0:
initparameters.initkernels(model, scheme=scheme)
# initparameters.renormalize(model, u0)
initparameters.initexpr(model, viscosity=viscosity, pattern='random')
else: # load checkpoint of layer-$start_from
callback.load(start_from, iternum='final')
#%% train
for block in blocks:
if block <= start_from:
continue
print('block: ', block)
print('name: ', name)
r = np.random.randn() + torch.randn(1, dtype=torch.float64,
device=device).item()
with callback.open() as output:
print('device: ', device, file=output)