# path_to_save_variables=home+'/Documents/tmp/pytorch_bvae'+str(i)+'.pt'
# path_to_load_variables=home+'/Documents/tmp/pytorch_bvae'+str(i)+'.pt'
# path_to_save_variables=''
# model.load_state_dict(torch.load(path_to_save_variables, lambda storage, loc: storage))
# print 'loaded variables ' + path_to_save_variables
# path_to_save_variables=this_dir+'/params_'+args.model +'_'
if torch.cuda.is_available():
model.cuda()
this_ckt_file = path_to_save_variables + str(3000) + '.pt'
# this_ckt_file = path_to_save_variables + str(args.epoch) + '.pt'
model.load_params(path_to_load_variables=this_ckt_file)
# ffg_samps = [0,1,2,3]
# ffg_samps = [5,6,7,8]
# # used for standard
ffg_samps = [6, 3, 2, 5]
# ffg_samps = [5]
# ffg_samps = [6]
rows = 4
cols = len(ffg_samps) + 1 #for annotation
legend = False
fig = plt.figure(figsize=(3 + cols, 1 + rows), facecolor='white')
# 'rv_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'flow_hidden_size': 100
# }
# # model = VAE(hyper_config)
# else:
# print ('What')
# fadas
trained_model = VAE(hyper_config)
# model.load_params(home+'/Documents/tmp/first_try/'+args.model+'/params_'+args.model+'_2800.pt')
trained_model.load_params(home+'/Documents/tmp/new_training/standard/params_standard_1000.pt')
# now to reset encoder, or even load a new one with different architecture.
# could maybe init a new one. and get model to ignore the old one.
# antoher option is init the right model then just load the params you want
# dont init q
# how does it know which params are the decoder ones, besides shape?
#ah they have keys names! mine are mostly numbers
# # print (model.state_dict())
# print (model.state_dict().keys())
# fsdaas
# 'n_flows': 2,
# 'qv_arch': [[x_size,200],[200,200],[200,z_size*2]],
# 'qz_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'rv_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'flow_hidden_size': 100
# }
# # model = VAE(hyper_config)
else:
print('What')
fadas
model = VAE(hyper_config)
print('done init')
model.load_params(home + '/Documents/tmp/large_N_time/' + args.model +
'/params_' + args.model + '_1.pt')
#Train params
learning_rate = .001
batch_size = 100
k = 1
epochs = 3000
#save params and compute IW and AIS
start_at = 100
save_freq = 300
display_epoch = 10
# Test params
k_IW = 2000
batch_size_IW = 20
# path_to_save_variables=this_dir+'/params_'+args.model +'_'
if torch.cuda.is_available():
model.cuda()
this_ckt_file = path_to_save_variables + str(3000) + '.pt'
# this_ckt_file = path_to_save_variables + str(args.epoch) + '.pt'
model.load_params(path_to_load_variables=this_ckt_file)
# ffg_samps = [0,1,2,3]
# ffg_samps = [5,6,7,8]
# # used for standard
ffg_samps = [6,3,2,5]
# ffg_samps = [5]
# ffg_samps = [6]
# 'n_flows': 2,
# 'qv_arch': [[x_size,200],[200,200],[200,z_size*2]],
# 'qz_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'rv_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'flow_hidden_size': 100
# }
# # model = VAE(hyper_config)
# else:
# print ('What')
# fadas
trained_model = VAE(hyper_config)
# model.load_params(home+'/Documents/tmp/first_try/'+args.model+'/params_'+args.model+'_2800.pt')
trained_model.load_params(
home + '/Documents/tmp/new_training/standard/params_standard_1000.pt')
# now to reset encoder, or even load a new one with different architecture.
# could maybe init a new one. and get model to ignore the old one.
# antoher option is init the right model then just load the params you want
# dont init q
# how does it know which params are the decoder ones, besides shape?
#ah they have keys names! mine are mostly numbers
# # print (model.state_dict())
# print (model.state_dict().keys())
# fsdaas
# # for k,v in model.state_dict.items():
# # print (k)
# 'qv_arch': [[x_size,200],[200,200],[200,z_size*2]],
# 'qz_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'rv_arch': [[x_size+z_size,200],[200,200],[200,z_size*2]],
# 'flow_hidden_size': 100
# }
# # model = VAE(hyper_config)
else:
print ('What')
fadas
model = VAE(hyper_config)
print ('done init')
model.load_params(home+'/Documents/tmp/large_N_time/'+args.model+'/params_'+args.model+'_1.pt')
#Train params
learning_rate = .001
batch_size = 100
k = 1
epochs = 3000
#save params and compute IW and AIS
start_at = 100
save_freq = 300
display_epoch = 10