def get_run_parameters(SB):
#GET MISCELANEOUS RUN PARAMETERS
SB['src_path'] = path.dirname(path.realpath(__file__))
SB['run_path'] = writer.run_path
SB['start_time'] = writer.start_time;
writer.log("RUN PARAMETERS:");
writer.log_dict(SB)
def read_pot_file(SB):
if (SB['pot_type'] == "NN"):
file_path = SB['pot_file']
writer.log("READING NEURAL NETWORK FILE:")
if (path.exists(file_path)):
input_file = open(file_path, "r")
lines = []
for line in input_file:
parts = line.strip().split() #rm /n
if (len(parts) == 1):
parts = parts[0] #dont save single numbers as arrays
if (parts != []): lines.append(parts)
pot = neural.NN(lines,
SB) #send lines to NN class to create NN object
else:
raise ValueError("NN_FILE=" + str(file_path) + " DOESNT EXIST")
writer.log_dict(pot.info)
SB.update(pot.info)
SB['nn'] = pot
def read_input(SB):
#READ DEFAULT FILE
file_path = SB['src_path'] + '/defaults.json'
if (path.exists(file_path)):
writer.log(["READING DEFAULT PARAMETERS USING FILE:", file_path])
with open(file_path, "r") as read_file: #READ USER INPUT FILE
input_data = load(read_file)
writer.log_dict(input_data)
SB.update(input_data)
else:
raise ValueError("DEFAULT_FILE=" + str(file_path) + " DOESNT EXIST")
#READ INPUT FILE
file_path = SB['input_file']
if (path.exists(file_path)):
writer.log([
"OVERWRITING SELECT DEFAULTS USING INPUT FILE:", SB['input_file']
])
with open(file_path, "r") as read_file: #READ USER INPUT FILE
input_data = load(read_file)
writer.log_dict(input_data)
SB.update(input_data)
else:
raise ValueError("INPUT_FILE=" + str(file_path) + " DOESNT EXIST")
if (SB['use_cuda'] and cuda.is_available() == False):
writer.log("NOTE: CUDA IS NOT AVAILABLE (RE-SETTING)")
writer.log(" use_cuda : False")
SB['use_cuda'] = False
if ('pot_type' not in SB.keys() or 'pot_file' not in SB.keys()
or 'dataset_path' not in SB.keys()):
raise ValueError(
"INPUT FILE MUST CONTAIN KEYS FOLLOWING KEYS: pot_type, pot_file, dataset_path"
)
if (SB['pot_type'] != "NN"):
raise ValueError("REQUESTED POT_TYPE=" + str(SB['pot_type']) +
" IS NOT AVAILABLE")
def read_database(SB):
file_path = SB['dataset_path']
writer.log("READING DATASET FILE:")
full_set = data.Dataset("full", SB) #INITIALIZE FULL DATASET OBJECT
SID = 0
new_structure = True
if (path.exists(file_path)):
input_file = open(file_path, "r")
for line in input_file:
if (new_structure):
lines = []
new_structure = False
counter = 1
full_set.Ns += 1
else:
counter += 1
parts = line.strip().split()
if (len(parts) == 1):
parts = parts[0] #dont save single numbers as arrays
lines.append(parts)
#TODO: THIS HARDCODE FOR CURRENT POSCAR FORMAT (NEED TO GENERALIZE)
if (counter == 6): Natom = int(parts)
if (counter > 6):
if (counter == 6 + 1 + Natom +
1): #see dataset examples for format
full_set.Na += Natom
full_set.structures[SID] = data.Structure(lines, SID, SB)
#create structure object
GID = str(full_set.structures[SID].gid)
if (GID not in full_set.group_sids.keys()):
full_set.group_sids[GID] = []
full_set.group_sids[GID].append(
[full_set.structures[SID].v, SID])
new_structure = True
SID += 1
#print(lines)
else:
raise ValueError("DATABASE_FILE=" + str(file_path) + " DOESNT EXIST")
full_set.sort_group_sids()
writer.log([" TOTAL NUMBER OF STRUCTURES:", full_set.Ns])
writer.log([" TOTAL NUMBER OF ATOMS: ", full_set.Na])
SB['full_set'] = full_set
def add_neurons(self):
#add N neurons to each hidden layer
if (self.info['activation'] != 1):
raise ValueError(
"ERROR: CAN ONLY ADD NEURONS TO SHIFTD SIGMOID FUNCTION")
#START FRESH EVERY TIME
start_fresh = self.info['start_fresh']
if (start_fresh):
self.randomize()
max_rand_wb = self.info['max_rand_wb']
else:
max_rand_wb = 1.0
self.unset_grad()
new_nfit = 0
N_neuron_2_add = 2
writer.log("ADDING " + str(N_neuron_2_add) + " NEURONS TO EACH LAYER")
writer.log([" original num_fit_param =", self.info['num_fit_param']])
for layer_add in range(1, len(self.info['nn_layers']) - 1):
for neurons in range(0, N_neuron_2_add):
for i in range(0, len(self.submatrices)):
layer = 2 * (i - 1)
if (layer_add == (i + 2.0) / 2):
#ADD ROW (WEIGHT MATRIX)
shp2 = self.submatrices[i].shape[1]
TMP = max_rand_wb * torch.empty(1, shp2).uniform_(
-1.0, 1.0)
self.submatrices[i] = torch.cat(
(self.submatrices[i], TMP))
#ADD BIAS
shp2 = self.submatrices[i + 1].shape[1]
TMP = max_rand_wb * torch.empty(1, shp2).uniform_(
-1.0, 1.0)
self.submatrices[i + 1] = torch.cat(
(self.submatrices[i + 1], TMP))
# #ADD COL (WEIGHT MATRIX)
shp1 = self.submatrices[i + 2].shape[0]
TMP = max_rand_wb * torch.empty(shp1, 1).uniform_(
-1.0, 1.0)
self.submatrices[i + 2] = torch.cat(
(self.submatrices[i + 2], TMP), 1)
self.info['nn_layers'][
layer_add] = self.info['nn_layers'][layer_add] + 1
#COUNT NFIT
for i in range(0, len(self.submatrices)):
new_nfit += self.submatrices[i].shape[0] * self.submatrices[
i].shape[1]
self.info['num_fit_param'] = new_nfit
writer.log([" new num_fit_param =", new_nfit])
self.set_grad()
def __init__(self, lines, SB):
info = {}
info['lsp_type'] = int(lines[0][0])
info['pot_type'] = str(SB['pot_type'])
info['lsp_shift'] = float(lines[0][1])
info['activation'] = int(lines[0][2])
info['num_species'] = int(lines[1][0])
info['species'] = str(lines[2][0])
info['atomic_weight'] = float(lines[2][1])
info['randomize_nn'] = bool(int(lines[3][0]))
info['max_rand_wb'] = float(lines[3][1])
info['cutoff_dist'] = float(lines[3][2])
info['cutoff_range'] = float(lines[3][3])
info['lsp_sigma'] = float(lines[3][4])
info['N_lg_poly'] = int(lines[4][0])
#map converts str list to int list
info['lsp_lg_poly'] = list(map(int, lines[4][1:]))
info['N_ro_val'] = int(lines[5][0])
info['lsp_ro_val'] = list(map(float, lines[5][1:]))
#map converts str list to float list
info['ibaseline'] = bool(int(lines[6][0]))
info['bop_param'] = list(map(float, lines[6][1:]))
info['nn_layers'] = list(map(int, lines[7][1:]))
info['cnst_final_bias'] = SB['cnst_final_bias']
info['final_bias'] = SB['final_bias']
info['start_fresh'] = SB['start_fresh']
info['constrain_WB'] = SB['constrain_WB']
#DETERMINE NUMBER OF FITITNG PARAM AND RANDOMIZE IF NEEDED
nfit = 0
layers = info['nn_layers']
for i in range(1, len(layers)):
nfit = nfit + layers[i - 1] * layers[i] + layers[i]
info['num_fit_param'] = nfit
self.info = info
self.normalize_by_ro = SB["normalize_by_ro"]
self.dtype = torch.FloatTensor
if (SB['use_cuda']): self.dtype = torch.cuda.FloatTensor
if (info['randomize_nn'] or SB['re_randomize']):
writer.log([" RANDOMIZING NN MIN/MAX =", info['max_rand_wb']])
self.randomize()
else:
#always do LR ramp up when re-starting
SB['ramp_LR'] = True
WB = np.array(lines[8:]).astype(np.float)[:, 0]
self.submatrices = self.extract_submatrices(WB)
if (len(WB) != info['num_fit_param']):
raise ValueError("INCORRECT NUMBER OF FITTING PARAMETERS")
#SOME ERROR CHECKS
if (info['num_species'] != 1):
raise ValueError("NUM_SPECIES != 1 IN EURAL NETWORK FILE")
if (len(info['nn_layers']) != int(lines[7][0])):
raise ValueError(
"NUMBER OF LAYERS IN NEURAL NETWORK FILE IS INCORRECT")
if (int(lines[0][0]) not in [5, 6, 7, 20]):
raise ValueError("REQUESTED POT_TYPE=" + str(int(lines[0][0])) +
" NOT AVAILABLE")
if (info['pot_type'] == 'PINN_BOP' and info['nn_layers'][-1] != 8):
raise ValueError("ERROR: NN OUTPUT DIMENSION INCORRECT")
if (info['pot_type'] == 'NN' and info['nn_layers'][-1] != 1):
raise ValueError("ERROR: NN OUTPUT DIMENSION INCORRECT")
if (info['N_ro_val'] != len(info['lsp_ro_val'])):
raise ValueError("ERROR: N_ro_val != len(ro)")
if (info['N_lg_poly'] != len(info['lsp_lg_poly'])):
raise ValueError("ERROR: N_lg_poly != len(lsp_lg_poly)")
if (info['nn_layers'][0] !=
len(info['lsp_ro_val']) * len(info['lsp_lg_poly'])):
raise ValueError(
"ERROR: NN INPUT DIMENSION INCORRECT FOR Gi CHOICE")
optimizer=optim.LBFGS(SB['nn'].submatrices, max_iter=SB['lbfgs_max_iter'], lr=SB['LR_f'])
set_optim()
def closure():
global loss,OBE1,OBL1,OBLP,OB_DU,rmse,OBT
optimizer.zero_grad(); loss=0.0
[rmse,OBE1,OB_DU,OBL1,OBLP]=training_set.compute_objective(SB)
loss=OBE1+OB_DU+OBL1+OBLP
loss.backward();
OBE1=OBE1.item(); OB_DU=OB_DU.item(); OBLP=OBLP.item()
OBL1=OBL1.item(); OBT=loss.item();
return loss
#OPTIMIZATION LOOP
start=time();
writer.log('STARTING FITTING LOOP:')
writer.log([" INITIAL LR:",'%10.7s'%str(optimizer.param_groups[0]['lr'])])
N_TRY=1;
while(t<max_iter):
optimizer.step(closure)
if(SB['ramp_LR']): scheduler.step() #ADJUST LR
#CHECK CONVERGENCE
if(str(OBE1)=='nan' or rmse>1000000): #START OVER
writer.log("NOTE: THE OBJ FUNCTION BLEW UP (IM STARTING OVER)(MAYBE TRY SMALLER LR)")
SB['nn'].unset_grad(); SB['nn'].randomize(); set_optim(); N_TRY=N_TRY+1
delta1=((rmse_m1-rmse)**2.0)**0.5
delta2=((rmse_m2-rmse)**2.0)**0.5
def partition_data(SB):
test_set=data.Dataset("test",SB) #INITIALIZE DATASET OBJECT
training_set=data.Dataset("train",SB) #INITIALIZE DATASET OBJECT
validation_set=data.Dataset("validate",SB) #INITIALIZE DATASET OBJECT
no_dft_set=data.Dataset("no_dft",SB) #INITIALIZE DATASET OBJECT
writer.log("PARTITIONING DATA:")
writer.log([" TOTAL NUMBER OF GROUPS=",len(SB['full_set'].group_sids.keys())])
fraction_train=SB['fraction_train']
train_edges=SB['train_edges']
#ERROR CHECKS
if(fraction_train==0):
raise ValueError("FRACTION_TRAIN=0 (CANT TRAIN WITHOUT TRAINING DATA)");
if(fraction_train<0 or fraction_train>1):
ERR="BAD VALUE FOR FRACTION_TRAIN: (I.E. FRACTION_TRAIN<0 OR FRACTION_TRAIN>1)"
raise ValueError(ERR);
if(SB['n_rand_GIDS']>=len(SB['full_set'].group_sids.keys())):
ERR="N_RAND_GIDS IS LARGER THAN TOTAL NUMBER OF GIDS: USE N_RAND_GIDS<" \
+str(len(SB['full_set'].group_sids.keys()))
raise ValueError(ERR);
#-------------------------------------
#TEST-SET (EXTRAPOLATION)
#-------------------------------------
if(SB['fix_rand_seed']): random.seed(a=412122, version=2) #SAME RAND TEST SET EVERY TIME
SB['test_set_gids']=[]
#COLLECT GID FOR TEST SET
if(SB['n_rand_GIDS']!=0):
k=1
while(k<=SB['n_rand_GIDS']):
rand_GID=random.choice(list(SB['full_set'].group_sids.keys()))
#if(rand_GID not in SB['test_set_gids'] ):
for i1 in SB['exclude_from_test']:
if(i1 not in rand_GID):
keep=True
else:
keep=False; break
if(rand_GID not in SB['test_set_gids'] and keep and rand_GID != "NO_DFT"): #REMOVE
#writer.log(" "+rand_GID)
SB['test_set_gids'].append(rand_GID)
k=k+1
for key in SB['test_set_tags']:
for GID in SB['full_set'].group_sids.keys():
for i1 in SB['exclude_from_test']:
if(i1 not in GID):
keep=True
else:
keep=False; break
if(key in SB['test_set_gids']): keep=False
if(key in GID and keep and GID != "NO_DFT"):
SB['test_set_gids'].append(GID)
writer.log(" TEST SET (UNTRAINED):")
#COLLECT STRUCTURES FOR TEST SET
for GID in SB['full_set'].group_sids.keys():
if(GID in SB['test_set_gids']):
writer.log([" GID : ",GID])
#test_set.group_sids[GID]= SB['full_set'].group_sids[GID]
for SID in SB['full_set'].group_sids[GID]:
test_set.structures[SID] = SB['full_set'].structures[SID]
test_set.Ns+=1;
test_set.Na+=SB['full_set'].structures[SID].N
#EXTAPOLATION
if("NO_DFT" in SB['full_set'].group_sids.keys()):
for SID in SB['full_set'].group_sids["NO_DFT"]:
no_dft_set.structures[SID] = SB['full_set'].structures[SID]
no_dft_set.Ns+=1;
no_dft_set.Na+=SB['full_set'].structures[SID].N
#COLLECT WHATS LEFT (use for train+val)
remainder=[]
for SID in SB['full_set'].structures.keys():
if(SID not in test_set.structures.keys() and SID not in no_dft_set.structures.keys()):
remainder.append(SID)
#-------------------------------------
#TRAIN-VALIDATION-SET (TRAIN+INTERPOLATION SET)
#-------------------------------------
#TRAINING SIDS (LIST OF DICTIONARY KEYS)a
train_indices=np.random.choice(len(remainder),int(fraction_train*len(remainder)), replace=False).tolist() #keys for training structures
for i in train_indices:
training_set.structures[remainder[i]]= SB['full_set'].structures[remainder[i]]
training_set.Ns+=1;
training_set.Na+=SB['full_set'].structures[remainder[i]].N
# #ADD MIN/MAX VOLUME STRUCTURES IN EACH GROUP TO TRAINING SET
if(train_edges):
sid_2_add=[]
for i in SB['full_set'].group_sids.values():
if(len(i)>4): #already sorted by volume
sid_2_add.append(i[0]); sid_2_add.append(i[1])
sid_2_add.append(i[-2]); sid_2_add.append(i[-1])
#print(sid_2_add)
for SID in sid_2_add:
if(SID not in training_set.structures.keys() and SID not in test_set.structures.keys() \
and SID not in no_dft_set.structures.keys()):
training_set.structures[SID]= SB['full_set'].structures[SID]
training_set.Ns+=1;
training_set.Na+=SB['full_set'].structures[SID].N
# #VALIDATION SID
for SID in remainder:
if(SID not in training_set.structures.keys()):
validation_set.structures[SID]= SB['full_set'].structures[SID]
validation_set.Ns+=1;
validation_set.Na+=SB['full_set'].structures[SID].N
#if(SB['full_set'].Ns != training_set.Ns+test_set.Ns+validation_set.Ns):
# raise ValueError("LOST A STUCTURE IN DATA PARTITIONING");
# if(test_SIDS==[]): test_SIDS=validation_SIDS #not ideal but test_SIDS cant be empty
writer.log([" N_train_structures : ",training_set.Ns])
writer.log([" N_val_structures : ",validation_set.Ns])
writer.log([" N_test_structures : ",test_set.Ns])
writer.log([" N_combined : ",training_set.Ns+test_set.Ns+validation_set.Ns])
test_set.build_arrays(SB)
training_set.build_arrays(SB)
validation_set.build_arrays(SB)
SB['training_set']=training_set; SB['datasets']=['training_set']
if(validation_set.Ns>0):
SB['validation_set']=validation_set
SB['datasets'].append('validation_set')
if(test_set.Ns>0):
SB['test_set']=test_set
SB['datasets'].append('test_set')
if("NO_DFT" in SB['full_set'].group_sids.keys()):
no_dft_set.build_arrays(SB)
SB['no_dft_set']=no_dft_set
SB['datasets'].append('no_dft_set')
def compute_all_lsps(SB): writer.log(["COMPUTING LOCAL STRUCTURE PARAMETERS (LSP):"]) start = time.time(); for structure in SB['full_set'].structures.values(): structure.compute_lsp(SB); writer.log([" LSP CONSTRUCTION TIME (SEC) =",time.time()-start])
def compute_all_nbls(SB): writer.log(["COMPUTING NEIGHBOR LIST (NBL):"]) start = time.time(); for structure in SB['full_set'].structures.values(): structure.compute_nbl(SB); writer.log([" NBL CONSTRUCTION TIME (SEC) =",time.time()-start])