def main():
df = make_df(constants.PATH_JSON)
train_ds, valid_ds = data_generator(constants.PATH_IMG, df)
input_main, output_main = multimodal_multistream_model()
model = Model(inputs=input_main, outputs=output_main)
model.compile(optimizer=tf.keras.optimizers.Adam(
learning_rate=constants.LEARNING_RATE),
loss=tf.losses.BinaryCrossentropy(),
metrics=['accuracy'])
model.summary()
checkpoint_path = constants.PATH_MODEL_SAVE
cb_checkpoint = tf.keras.callbacks.ModelCheckpoint(
filepath=checkpoint_path,
monitor='val_loss',
save_weights_only=False,
save_best_only=True,
verbose=2)
cb_earlystopper = tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=10)
with tf.device('/device:GPU:0'):
fit_history = model.fit(train_ds,
epochs=100,
validation_data=valid_ds,
callbacks=[cb_earlystopper, cb_checkpoint])
plot_model(fit_history)
def test_plot_model(self):
"""
Test if plotted model is saved to disk
"""
mock_model = self.mock_model
plot_model(mock_model, self.test_plots_dir)
files = os.listdir(self.test_plots_dir)
if files:
test_ext = os.path.splitext(files[0])[1]
# Saves a models
self.assertTrue(files, "no model plot saved")
self.assertEqual(".png", test_ext, "model plot not saved as '.png'")
def main(args):
path = os.path.abspath(__file__).rsplit('/', 1)[0]
defaults = parse_config(path + '/default.cfg')
opt = parse_config('model.cfg', defaults)
print pprint.pformat(opt)
print '### Read parameters from model.cfg ###'
filename, = args
samples = loadobj(filename)
mean_accept = samples['accept'].mean()
print 'Mean acceptance fraction', mean_accept
nwalkers, nsamples, npar = samples['chain'].shape
if not os.path.lexists('fig/'):
os.mkdir('fig')
if filename.startswith('samples_burn'):
# estimate maximum likelihood as the point in the chain with
# the highest likelihood.
i = samples['lnprob'].ravel().argmax()
P['ml'] = samples['chain'].reshape(-1, npar)[i]
print 'Plotting burn-in sample posteriors'
# bins for plotting posterior histograms
P['bins'] = [np.linspace(lo, hi, opt.Nhistbins) for
lo,hi in zip(P['min'], P['max'])]
fig,axes = plot_posteriors_burn(samples['chain'], P, npar=opt.npar)
fig.suptitle('%i samples of %i walkers' % (
nsamples, nwalkers), fontsize=14)
fig.savefig('fig/posterior_burnin.' + opt.plotformat)
print 'Plotting traces'
fig, nwplot = plot_trace(samples['chain'])
fig.suptitle('Chain traces for %i of %i walkers' % (nwplot,nwalkers))
fig.savefig('fig/traces.' + opt.plotformat)
if opt.autocorr:
print 'Plotting autocorrelation'
fig, axes = plot_autocorr(samples['chain'])
fig.suptitle('Autocorrelation for %i walkers with %i samples. '
'(Mean acceptance fraction %.2f)' %
(nwalkers, nsamples, mean_accept), fontsize=14)
fig.savefig('fig/autocorr.' + opt.plotformat)
else:
# make a chain of independent samples
Ns, Nt = opt.Nsamp, opt.Nthin
assert Ns * Nt <= nsamples
chain = samples['chain'][:,0:Ns*Nt:Nt,:].reshape(-1, npar)
# bins for plotting posterior histograms
P['bins'] = []
for i in xrange(len(P['names'])):
x0, x1 = chain[:,i].min(), chain[:,i].max()
dx = x1 - x0
lo = x0 - 0.1*dx
hi = x1 + 0.1*dx
P['bins'].append( np.linspace(lo, hi, opt.Nhistbins) )
levels = 0.6827, 0.9545
P['p1sig'] = [find_min_interval(chain[:, i], levels[0]) for i
in range(npar)]
P['p2sig'] = [find_min_interval(chain[:, i], levels[1]) for i
in range(npar)]
# if hasattr(P, 'nuisance') and any(P.nuisance):
# print 'marginalising over nuisance parameters'
# marginalised_chain = chain[:, [i for i in range(npar)
# if not P.nuisance[i]]]
# print chain.shape, marginalised_chain.shape
# ijoint_sig = get_levels(marginalised_chain, levels)
lnprob = samples['lnprob'][:,0:Ns*Nt:Nt].ravel()
isort = lnprob.argsort()
P['ijoint_sig'] = [isort[int((1-l)*len(lnprob)):] for l in levels]
# the joint 1 and 2 sigma regions, simulatenously estimating
# all parameters.
P['p1sig_joint'] = []
P['p2sig_joint'] = []
for i in range(npar):
lo = chain[P['ijoint_sig'][0], i].min()
hi = chain[P['ijoint_sig'][0], i].max()
P['p1sig_joint'].append((lo, hi))
lo = chain[P['ijoint_sig'][1], i].min()
hi = chain[P['ijoint_sig'][1], i].max()
P['p2sig_joint'].append((lo, hi))
P['median'] = np.median(chain, axis=0)
# estimate maximum likelihood as the point in the chain with
# the highest likelihood.
i = samples['lnprob'].ravel().argmax()
P['ml'] = samples['chain'].reshape(-1, npar)[i]
if opt.find_maximum_likelihood:
if not scipy:
raise ImportError('Scipy minimize not available')
print 'Finding maximum likelihood parameter values'
P['ml'] = minimize(lambda *x: -ln_likelihood(*x), P['ml'])
print 'done'
if opt.plotposteriors:
print 'Plotting sample posteriors'
fig, axes = plot_posteriors(chain, P, npar=opt.npar)
fig.suptitle('%i of %i samples, %i walkers, thinning %i' % (
Ns, nsamples, nwalkers, Nt), fontsize=14)
fig.savefig('fig/posterior_mcmc.' + opt.plotformat)
if opt.plotdata:
print 'Plotting the maximum likelihood model and data'
from model import plot_model
fig = plot_model(P['ml'])
fig.savefig('fig/model.' + opt.plotformat)
if opt.printpar and not filename.startswith('samples_burn'):
from model import print_par
print_par(P)
if opt.display:
print 'Displaying...'
pl.show()
print 'Done!'
# if you get an error when running emcee in parallel, you often get an
# unhelpful error messgae, making debugging very difficult. To check
# everything is working before you start running emcee, use thsi test
# file.
from model import \
ln_likelihood, P, x, ydata, ysigma, get_initial_positions, plot_model
p0 = get_initial_positions(1)
plot_model(p0)
def main(args):
path = os.path.abspath(__file__).rsplit("/", 1)[0]
defaults = parse_config(path + "/default.cfg")
opt = parse_config("model.cfg", defaults)
print pprint.pformat(opt)
print "### Read parameters from model.cfg ###"
filename, = args
samples = loadobj(filename)
mean_accept = samples["accept"].mean()
print "Mean acceptance fraction", mean_accept
nwalkers, nsamples, npar = samples["chain"].shape
if not os.path.lexists("fig/"):
os.mkdir("fig")
if filename.startswith("samples_burn"):
# estimate maximum likelihood as the point in the chain with
# the highest likelihood.
i = samples["lnprob"].ravel().argmax()
P["ml"] = samples["chain"].reshape(-1, npar)[i]
print "Plotting burn-in sample posteriors"
# bins for plotting posterior histograms
P["bins"] = [np.linspace(lo, hi, opt.Nhistbins) for lo, hi in zip(P["min"], P["max"])]
fig, axes = plot_posteriors_burn(samples["chain"], P, npar=opt.npar)
fig.suptitle("%i samples of %i walkers" % (nsamples, nwalkers), fontsize=14)
fig.savefig("fig/posterior_burnin." + opt.plotformat)
print "Plotting traces"
fig, nwplot = plot_trace(samples["chain"])
fig.suptitle("Chain traces for %i of %i walkers" % (nwplot, nwalkers))
fig.savefig("fig/traces." + opt.plotformat)
if opt.autocorr:
print "Plotting autocorrelation"
fig, axes = plot_autocorr(samples["chain"])
fig.suptitle(
"Autocorrelation for %i walkers with %i samples. "
"(Mean acceptance fraction %.2f)" % (nwalkers, nsamples, mean_accept),
fontsize=14,
)
fig.savefig("fig/autocorr." + opt.plotformat)
else:
# make a chain of independent samples
Ns, Nt = opt.Nsamp, opt.Nthin
assert Ns * Nt <= nsamples
chain = samples["chain"][:, 0 : Ns * Nt : Nt, :].reshape(-1, npar)
# bins for plotting posterior histograms
P["bins"] = []
for i in xrange(len(P["names"])):
x0, x1 = chain[:, i].min(), chain[:, i].max()
dx = x1 - x0
lo = x0 - 0.1 * dx
hi = x1 + 0.1 * dx
P["bins"].append(np.linspace(lo, hi, opt.Nhistbins))
levels = 0.6827, 0.9545
P["p1sig"] = [find_min_interval(chain[:, i], levels[0]) for i in range(npar)]
P["p2sig"] = [find_min_interval(chain[:, i], levels[1]) for i in range(npar)]
# if hasattr(P, 'nuisance') and any(P.nuisance):
# print 'marginalising over nuisance parameters'
# marginalised_chain = chain[:, [i for i in range(npar)
# if not P.nuisance[i]]]
# print chain.shape, marginalised_chain.shape
# ijoint_sig = get_levels(marginalised_chain, levels)
lnprob = samples["lnprob"][:, 0 : Ns * Nt : Nt].ravel()
isort = lnprob.argsort()
P["ijoint_sig"] = [isort[int((1 - l) * len(lnprob)) :] for l in levels]
# the joint 1 and 2 sigma regions, simulatenously estimating
# all parameters.
P["p1sig_joint"] = []
P["p2sig_joint"] = []
for i in range(npar):
lo = chain[P["ijoint_sig"][0], i].min()
hi = chain[P["ijoint_sig"][0], i].max()
P["p1sig_joint"].append((lo, hi))
lo = chain[P["ijoint_sig"][1], i].min()
hi = chain[P["ijoint_sig"][1], i].max()
P["p2sig_joint"].append((lo, hi))
P["median"] = np.median(chain, axis=0)
# estimate maximum likelihood as the point in the chain with
# the highest likelihood.
i = samples["lnprob"].ravel().argmax()
P["ml"] = samples["chain"].reshape(-1, npar)[i]
if opt.find_maximum_likelihood:
if not scipy:
raise ImportError("Scipy minimize not available")
print "Finding maximum likelihood parameter values"
P["ml"] = minimize(lambda *x: -ln_likelihood(*x), P["ml"])
print "done"
if opt.plotposteriors:
print "Plotting sample posteriors"
fig, axes = plot_posteriors(chain, P, nplot=opt.npar)
fig.suptitle("%i of %i samples, %i walkers, thinning %i" % (Ns, nsamples, nwalkers, Nt), fontsize=14)
fig.savefig("fig/posterior_mcmc." + opt.plotformat, dpi=200)
if opt.plotdata:
print "Plotting the maximum likelihood model and data"
from model import plot_model
if opt.nsamp_plot > 1:
chain = samples["chain"].reshape(-1, npar)
step = int(len(chain) / opt.nsamp_plot)
samp = chain[::step]
fig = plot_model(samp)
else:
fig = plot_model([P["median"]])
if opt.printpar and not filename.startswith("samples_burn"):
from model import print_par
print_par(P)
if opt.display:
print "Displaying..."
pl.show()
print "Done!"
def main(args):
path = os.path.abspath(__file__).rsplit('/', 1)[0]
defaults = parse_config(path + '/default.cfg')
opt = parse_config('model.cfg', defaults)
print pprint.pformat(opt)
print '### Read parameters from model.cfg ###'
filename, = args
samples = loadobj(filename)
mean_accept = samples['accept'].mean()
print 'Mean acceptance fraction', mean_accept
nwalkers, nsamples, npar = samples['chain'].shape
if not os.path.lexists('fig/'):
os.mkdir('fig')
if filename.startswith('samples_burn'):
# estimate maximum likelihood as the point in the chain with
# the highest likelihood.
i = samples['lnprob'].ravel().argmax()
P['ml'] = samples['chain'].reshape(-1, npar)[i]
print 'Plotting burn-in sample posteriors'
# bins for plotting posterior histograms
P['bins'] = [
np.linspace(lo, hi, opt.Nhistbins)
for lo, hi in zip(P['min'], P['max'])
]
fig, axes = plot_posteriors_burn(samples['chain'], P, npar=opt.npar)
fig.suptitle('%i samples of %i walkers' % (nsamples, nwalkers),
fontsize=14)
fig.savefig('fig/posterior_burnin.' + opt.plotformat)
print 'Plotting traces'
fig, nwplot = plot_trace(samples['chain'])
fig.suptitle('Chain traces for %i of %i walkers' % (nwplot, nwalkers))
fig.savefig('fig/traces.' + opt.plotformat)
if opt.autocorr:
print 'Plotting autocorrelation'
fig, axes = plot_autocorr(samples['chain'])
fig.suptitle('Autocorrelation for %i walkers with %i samples. '
'(Mean acceptance fraction %.2f)' %
(nwalkers, nsamples, mean_accept),
fontsize=14)
fig.savefig('fig/autocorr.' + opt.plotformat)
else:
# make a chain of independent samples
Ns, Nt = opt.Nsamp, opt.Nthin
assert Ns * Nt <= nsamples
chain = samples['chain'][:, 0:Ns * Nt:Nt, :].reshape(-1, npar)
# bins for plotting posterior histograms
P['bins'] = []
for i in xrange(len(P['names'])):
x0, x1 = chain[:, i].min(), chain[:, i].max()
dx = x1 - x0
lo = x0 - 0.1 * dx
hi = x1 + 0.1 * dx
P['bins'].append(np.linspace(lo, hi, opt.Nhistbins))
levels = 0.6827, 0.9545
P['p1sig'] = [
find_min_interval(chain[:, i], levels[0]) for i in range(npar)
]
P['p2sig'] = [
find_min_interval(chain[:, i], levels[1]) for i in range(npar)
]
# if hasattr(P, 'nuisance') and any(P.nuisance):
# print 'marginalising over nuisance parameters'
# marginalised_chain = chain[:, [i for i in range(npar)
# if not P.nuisance[i]]]
# print chain.shape, marginalised_chain.shape
# ijoint_sig = get_levels(marginalised_chain, levels)
lnprob = samples['lnprob'][:, 0:Ns * Nt:Nt].ravel()
isort = lnprob.argsort()
P['ijoint_sig'] = [isort[int((1 - l) * len(lnprob)):] for l in levels]
# the joint 1 and 2 sigma regions, simulatenously estimating
# all parameters.
P['p1sig_joint'] = []
P['p2sig_joint'] = []
for i in range(npar):
lo = chain[P['ijoint_sig'][0], i].min()
hi = chain[P['ijoint_sig'][0], i].max()
P['p1sig_joint'].append((lo, hi))
lo = chain[P['ijoint_sig'][1], i].min()
hi = chain[P['ijoint_sig'][1], i].max()
P['p2sig_joint'].append((lo, hi))
P['median'] = np.median(chain, axis=0)
# estimate maximum likelihood as the point in the chain with
# the highest likelihood.
i = samples['lnprob'].ravel().argmax()
P['ml'] = samples['chain'].reshape(-1, npar)[i]
if opt.find_maximum_likelihood:
if not scipy:
raise ImportError('Scipy minimize not available')
print 'Finding maximum likelihood parameter values'
P['ml'] = minimize(lambda *x: -ln_likelihood(*x), P['ml'])
print 'done'
if opt.plotposteriors:
print 'Plotting sample posteriors'
fig, axes = plot_posteriors(chain, P, nplot=opt.npar)
fig.suptitle('%i of %i samples, %i walkers, thinning %i' %
(Ns, nsamples, nwalkers, Nt),
fontsize=14)
fig.savefig('fig/posterior_mcmc.' + opt.plotformat, dpi=200)
if opt.plotdata:
print 'Plotting the maximum likelihood model and data'
from model import plot_model
if opt.nsamp_plot > 1:
chain = samples['chain'].reshape(-1, npar)
step = int(len(chain) / opt.nsamp_plot)
samp = chain[::step]
fig = plot_model(samp)
else:
fig = plot_model([P['median']])
if opt.printpar and not filename.startswith('samples_burn'):
from model import print_par
print_par(P)
if opt.display:
print 'Displaying...'
pl.show()
print 'Done!'
