def fit(final_params_file):
config = ConfigLoader(
"config.yml"
) # We use ConfigLoader to read the information in the configuration file
# config.set_params("gen_params.json") # If not set, we will use random initial parameters
fit_result = config.fit(method="BFGS")
errors = config.get_params_error(
fit_result) # calculate Hesse errors of the parameters
print("\n########## fit parameters:")
for key, value in config.get_params().items():
print(key, error_print(value, errors.get(key, None)))
fit_result.save_as(final_params_file) # save fit_result to a json file
config.plot_partial_wave(
fit_result
) # Plot distributions of variables indicated in the configuration file
fit_frac, err_frac = config.cal_fitfractions()
print("\n########## fit fractions:")
for i in fit_frac:
if not isinstance(i, tuple): # fit fraction
name = i
else:
name = "{}x{}".format(*i) # interference term
print(name + ": " + error_print(fit_frac[i], err_frac.get(i, None)))
def fit(config="config.yml", init_params="init_params.json", method="BFGS"):
"""
simple fit script
"""
# load config.yml
config = ConfigLoader(config)
# set initial parameters if have
try:
config.set_params(init_params)
print("using {}".format(init_params))
except Exception as e:
if str(e) != "[Errno 2] No such file or directory: 'init_params.json'":
print(e)
print("\nusing RANDOM parameters", flush=True)
# print("\n########### initial parameters")
# json_print(config.get_params())
# fit
data, phsp, bg, inmc = config.get_all_data()
try:
fit_result = config.fit(batch=65000, method=method)
except KeyboardInterrupt:
config.save_params("break_params.json")
raise
except Exception as e:
print(e)
config.save_params("break_params.json")
raise
json_print(fit_result.params)
fit_result.save_as("final_params.json")
# calculate parameters error
fit_error = config.get_params_error(fit_result, batch=13000)
fit_result.set_error(fit_error)
fit_result.save_as("final_params.json")
pprint(fit_error)
print("\n########## fit results:")
for k, v in config.get_params().items():
print(k, error_print(v, fit_error.get(k, None)))
# plot partial wave distribution
config.plot_partial_wave(fit_result, plot_pull=True)
# calculate fit fractions
phsp_noeff = config.get_phsp_noeff()
fit_frac, err_frac = config.cal_fitfractions({}, phsp_noeff)
print("########## fit fractions")
fit_frac_string = ""
for i in fit_frac:
if isinstance(i, tuple):
name = "{}x{}".format(*i)
else:
name = i
fit_frac_string += "{} {}\n".format(
name, error_print(fit_frac[i], err_frac.get(i, None)))
print(fit_frac_string)
def write_some_results_combine(config, fit_result, save_root=False):
from tf_pwa.applications import fit_fractions
for i, c in enumerate(config.configs):
c.plot_partial_wave(fit_result,
prefix="figure/s{}_".format(i),
save_root=save_root)
for it, config_i in enumerate(config.configs):
print("########## fit fractions {}:".format(it))
print(f"nll{it}", config_i.get_fcn()({}).numpy())
mcdata = config_i.get_phsp_noeff()
fit_frac, err_frac = fit_fractions(
config_i.get_amplitude(),
mcdata,
config.inv_he,
fit_result.params,
)
fit_frac_string = ""
for i in fit_frac:
if isinstance(i, tuple):
name = "{}x{}".format(*i) # interference term
else:
name = i # fit fraction
fit_frac_string += "{} {}\n".format(
name, error_print(fit_frac[i], err_frac.get(i, None)))
print(fit_frac_string)
save_frac_csv(f"fit_frac{it}.csv", fit_frac)
save_frac_csv(f"fit_frac{it}_err.csv", err_frac)
def cal_fitfractions(params_file):
config = ConfigLoader("config.yml")
config.set_params(params_file)
params = config.get_params()
config.get_params_error(params)
mcdata = (
config.get_phsp_noeff()
) # use the file of PhaseSpace MC without efficiency indicated in config.yml
fit_frac, err_frac = fit_fractions(
config.get_amplitude(), mcdata, config.inv_he, params
)
print("########## fit fractions:")
fit_frac_string = ""
for i in fit_frac:
if isinstance(i, tuple):
name = "{}x{}".format(*i) # interference term
else:
name = i # fit fraction
fit_frac_string += "{} {}\n".format(
name, error_print(fit_frac[i], err_frac.get(i, None))
)
print(fit_frac_string)
print("########## fit fractions table:")
print_frac_table(
fit_frac_string
) # print the fit-fractions as a 2-D table. The codes below are just to implement the print function.
def fit(config, init_params="", method="BFGS", loop=1):
"""
simple fit script
"""
# load config.yml
# config = ConfigLoader(config_file)
# load data
all_data = config.get_all_data()
fit_results = []
for i in range(loop):
# set initial parameters if have
if config.set_params(init_params):
print("using {}".format(init_params))
else:
print("\nusing RANDOM parameters", flush=True)
# try to fit
try:
fit_result = config.fit(batch=65000, method=method)
except KeyboardInterrupt:
config.save_params("break_params.json")
raise
except Exception as e:
print(e)
config.save_params("break_params.json")
raise
fit_results.append(fit_result)
# reset parameters
try:
config.reinit_params()
except Exception as e:
print(e)
fit_result = fit_results.pop()
for i in fit_results:
if i.success:
if not fit_result.success or fit_result.min_nll > i.min_nll:
fit_result = i
config.set_params(fit_result.params)
json_print(fit_result.params)
fit_result.save_as("final_params.json")
# calculate parameters error
fit_error = config.get_params_error(fit_result, batch=13000)
fit_result.set_error(fit_error)
fit_result.save_as("final_params.json")
pprint(fit_error)
print("\n########## fit results:")
print("Fit status: ", fit_result.success)
print("Minimal -lnL = ", fit_result.min_nll)
for k, v in config.get_params().items():
print(k, error_print(v, fit_error.get(k, None)))
return fit_result
def write_some_results(config, fit_result):
# plot partial wave distribution
config.plot_partial_wave(fit_result, plot_pull=True)
# calculate fit fractions
phsp_noeff = config.get_phsp_noeff()
fit_frac, err_frac = config.cal_fitfractions({}, phsp_noeff)
print("########## fit fractions")
fit_frac_string = ""
for i in fit_frac:
if isinstance(i, tuple):
name = "{}x{}".format(*i)
else:
name = i
fit_frac_string += "{} {}\n".format(
name, error_print(fit_frac[i], err_frac.get(i, None)))
print(fit_frac_string)
def cal_errors(params_file):
config = ConfigLoader("config.yml")
config.set_params(params_file)
fcn = config.get_fcn()
fcn.model.Amp.vm.rp2xy_all(
) # we can use this to transform all complex parameters to xy-coordinates, since the Hesse errors of xy are more statistically reliable
params = config.get_params()
errors, config.inv_he = cal_hesse_error(
fcn, params, check_posi_def=True, save_npy=True
) # obtain the Hesse errors and the error matrix (inverse Hessian)
print("\n########## fit parameters in XY-coordinates:")
errors = dict(zip(fcn.model.Amp.vm.trainable_vars, errors))
for key, value in config.get_params().items():
print(key, error_print(value, errors.get(key, None)))
print("\n########## correlation matrix:")
print("Matrix index:\n", fcn.model.Amp.vm.trainable_vars)
print("Correlation Coefficients:\n", corr_coef_matrix(config.inv_he)
) # obtain the correlation matrix using the inverse Hessian
def fit():
dtype = "float64"
config_list = load_config_file("Resonances")
decs, final_particles, decay = get_decay_chains(config_list)
amp = get_amplitude(decs, config_list, decay)
# data/pure1w.dat data/flat_mc30w.dat
data = prepare_data(
"data/data4600_new.dat", decs, particles=final_particles, dtype=dtype
)
mcdata = prepare_data(
"data/PHSP4600_new.dat", decs, particles=final_particles, dtype=dtype
)
bg = prepare_data(
"data/bg4600_new.dat", decs, particles=final_particles, dtype=dtype
)
w_bkg = 0.768331
model = Model(amp, w_bkg=w_bkg)
load_params(model.Amp, "glb_params_rp.json")
model.Amp.vm.trans_params(True)
args_name = model.Amp.vm.trainable_vars
fcn = FCN(model, data, mcdata, bg, batch=65000)
SCIPY = False
hesse = True
if SCIPY:
def callback(x):
print(fcn.cached_nll)
f_g = fcn.model.Amp.vm.trans_fcn_grad(fcn.nll_grad)
now = time.time()
s = minimize(
f_g,
np.array(fcn.model.Amp.vm.get_all_val(True)),
method="BFGS",
jac=True,
callback=callback,
options={"disp": 1, "gtol": 1e-4, "maxiter": 1000},
)
model.Amp.vm.trans_params(True)
xn = fcn.model.Amp.vm.get_all_val()
print("########## fit state:")
print(s)
print("\nTime for fitting:", time.time() - now)
val = {k: v.numpy() for k, v in model.Amp.variables.items()}
err = {}
if hesse:
inv_he = cal_hesse_error(
model.Amp, val, w_bkg, data, mcdata, bg, args_name, batch=20000
)
diag_he = inv_he.diagonal()
hesse_error = np.sqrt(np.fabs(diag_he)).tolist()
print(hesse_error)
err = dict(zip(model.Amp.vm.trainable_vars, hesse_error))
else:
from tf_pwa.fit import fit_minuit
m = fit_minuit(fcn, hesse=False)
model.Amp.vm.trans_params(True)
err_mtrx = m.np_covariance()
np.save("error_matrix.npy", err_mtrx)
val = dict(m.values)
err = dict(m.errors)
print("\n########## fit results:")
for i in val:
print(" ", i, ":", error_print(val[i], err.get(i, None)))
frac = True
err_frac = {}
if frac:
err_frac = {}
if hesse:
frac, grad = cal_fitfractions(
model.Amp, list(split_generator(mcdata, 25000))
)
else:
frac = cal_fitfractions_no_grad(
model.Amp, list(split_generator(mcdata, 45000))
)
for i in frac:
if hesse:
err_frac[i] = np.sqrt(np.dot(np.dot(inv_he, grad[i]), grad[i]))
print("########## fit fractions")
for i in frac:
print(i, ":", error_print(frac[i], err_frac.get(i, None)))
outdic = {
"value": val,
"error": err,
"config": config_list,
"frac": frac,
"err_frac": err_frac,
"NLL": fcn.cached_nll.numpy(),
}
with open("glb_params_rp.json", "w") as f:
json.dump(outdic, f, indent=2)
def fit(method="BFGS", init_params="init_params.json", hesse=True, frac=True):
POLAR = True # fit in polar coordinates. should be consistent with init_params.json if any
GEN_TOY = False # use toy data (mcdata and bg stay the same). REMEMBER to update gen_params.json
dtype = "float64"
w_bkg = 0.768331
# set_gpu_mem_growth()
tf.keras.backend.set_floatx(dtype)
# open Resonances list as dict
config_list = load_config_file("Resonances")
config_list2 = load_config_file("Resonances")
vm = VarsManager()
decs, final_particles, decay = get_decay_chains(config_list)
data, bg, mcdata = prepare_data(
data_file_name, decs, particles=final_particles, dtype=dtype
)
decs2, final_particles2, decay2 = get_decay_chains(config_list)
data2, bg2, mcdata2 = prepare_data(
data_file_name2, decs2, particles=final_particles2, dtype=dtype
)
amp = get_amplitude(decs, config_list, decay, polar=POLAR, vm=vm)
amp2 = get_amplitude(decs2, config_list2, decay2, polar=POLAR, vm=vm)
load_params(amp, init_params)
load_params(amp2, init_params)
model = Model(amp, w_bkg=w_bkg)
model2 = Model(amp2, w_bkg=w_bkg)
pprint(model.get_params())
fcn = CombineFCN(
[model, model2],
[data, data2],
[mcdata, mcdata2],
bg=[bg, bg2],
batch=65000,
)
# fit configure
args = {}
args_name = vm.trainable_vars
x0 = []
bnds = []
bounds_dict = {
# "Zc_4160_m:0":(4.1,4.22),
# "Zc_4160_g:0":(0,None)
}
for i in model.Amp.trainable_variables:
args[i.name] = i.numpy()
x0.append(i.numpy())
if i.name in bounds_dict:
bnds.append(bounds_dict[i.name])
else:
bnds.append((None, None))
args["error_" + i.name] = 0.1
points = []
nlls = []
now = time.time()
maxiter = 1000
method = "BFGS"
def callback(x):
if np.fabs(x).sum() > 1e7:
x_p = dict(zip(args_name, x))
raise Exception("x too large: {}".format(x_p))
points.append(vm.get_all_val())
nlls.append(float(fcn.cached_nll))
print(fcn.cached_nll)
vm.set_bound(bounds_dict)
f_g = vm.trans_fcn_grad(fcn.nll_grad)
s = minimize(
f_g,
np.array(vm.get_all_val(True)),
method=method,
jac=True,
callback=callback,
options={"disp": 1, "gtol": 1e-4, "maxiter": maxiter},
)
xn = vm.get_all_val()
print("########## fit state:")
print(s)
print("\nTime for fitting:", time.time() - now)
model.Amp.vm.trans_params(POLAR)
val = {k: v.numpy() for k, v in model.Amp.variables.items()}
with open("final_params.json", "w") as f:
json.dump({"value": val}, f, indent=2)
err = {}
inv_he = cal_hesse_error(
model.Amp, val, w_bkg, data, mcdata, bg, args_name, batch=20000
)
diag_he = inv_he.diagonal()
hesse_error = np.sqrt(np.fabs(diag_he)).tolist()
print(hesse_error)
err = dict(zip(model.Amp.vm.trainable_vars, hesse_error))
print("\n########## fit results:")
for i in val:
print(" ", i, ":", error_print(val[i], err.get(i, None)))
outdic = {"value": val, "error": err, "config": config_list}
with open("final_params.json", "w") as f:
json.dump(outdic, f, indent=2)
err_frac = {}
frac, grad = cal_fitfractions(
model.Amp, list(split_generator(mcdata, 25000))
)
for i in frac:
err_frac[i] = np.sqrt(np.dot(np.dot(inv_he, grad[i]), grad[i]))
print("########## fit fractions")
for i in frac:
print(i, ":", error_print(frac[i], err_frac.get(i, None)))
print("\nEND\n")
"""outdic = {"value": val, "error": err, "frac": frac, "err_frac": err_frac}
