avivajpeyi
commented
3 years ago Do we know of an example TOI that I can try out?
Closed avivajpeyi closed 2 years ago
avivajpeyi
commented
3 years ago Do we know of an example TOI that I can try out?
dfm
commented
3 years ago A good place to start might be to find examples with period < 0 in the catalog. I did start implementing single transits at one point if you want to take a look there as a starting point: https://github.com/exoplanet-dev/tess.world/blob/main/src/tess_world/templates/template.py#L305
avivajpeyi
commented
3 years ago From https://exofop.ipac.caltech.edu/tess/view_toi.php, I found 66/2394 candidates have period < 0. Going to test PE with TOI 2180.
avivajpeyi
commented
3 years ago single transit --> most likely a short period single transit
Hence incorporate that into the prior --> steep power-law prior in period
avivajpeyi
commented
3 years ago Re un with above, period, duration, b, probably very correlated. We can look at corner and see whats going on (Plot log_period)
avivajpeyi
commented
3 years ago Plot divergences on top of the corner to see where and which parameters might be causing the issue. Maybe this always happens at particular parameters?
avivajpeyi
commented
3 years ago Check if this works now that I have merged in #76
avivajpeyi
commented
2 years ago Reran 2180:
failed when testing the model https://github.com/dfm/tess-atlas/blob/81db5436b8f86f9da758e6272eaffa02e7890301/src/tess_atlas/notebook_templates/toi_template.py#L330
There is an issue with the model. The model(testval) has a nan:
dur_interval__ NaN
r_log__ -0.92
b_impact__ -1.39
t0_interval__ -1.19
p_1_lowerbound__ -inf
f0 -3.22
u_quadlimbdark__ -2.77
jitter_log__ 0.11
sigma_log__ 0.11
rho_log__ -0.55
obs -362020.84The error is arising from the transit duration dur and the orbital period p.
{'TOI': 2180.01,
'Period (days)': 0.0,
'Epoch (TBJD)': 1830.7900000000373,
'Depth (ppt)': 2.4,
'Duration (days)': 1.3089166666666667,
'Single Transit': True} def build_planet_transit_model(tic_entry):
t = tic_entry.lightcurve.time
y = tic_entry.lightcurve.flux
yerr = tic_entry.lightcurve.flux_err
n = tic_entry.planet_count
t0s = np.array([planet.t0 for planet in tic_entry.candidates])
depths = np.array([planet.depth for planet in tic_entry.candidates])
periods = np.array([planet.period for planet in tic_entry.candidates])
tmaxs = np.array([planet.tmax for planet in tic_entry.candidates])
durations = np.array([planet.duration for planet in tic_entry.candidates])
max_durations = np.array(
[planet.duration_max for planet in tic_entry.candidates]
)
min_durations = np.array(
[planet.duration_min for planet in tic_entry.candidates]
)
with pm.Model() as my_planet_transit_model:
## define planet parameters
# 1) d: transit duration (duration of eclipse)
d_priors = pm.Bound(
pm.Lognormal, lower=min_durations, upper=max_durations
)(
name=DURATION,
mu=np.log(durations),
sigma=np.log(1.2),
shape=n,
testval=durations,
)
...
# 1) t0: the time of the first transit in data (a reference time)
t0_norm = pm.Bound(
pm.Normal, lower=t0s - max_durations, upper=t0s + max_durations
)
t0_priors = t0_norm(
TIME_START, mu=t0s, sigma=0.5 * durations, shape=n, testval=t0s
)
# 2) period: the planets' orbital period
p_params, p_priors_list, tmax_priors_list = [], [], []
for n, planet in enumerate(tic_entry.candidates):
# if only one transit in data we use the period
if planet.has_data_only_for_single_transit:
p_prior = pm.Pareto(
name=f"{ORBITAL_PERIOD}_{planet.index}",
m=planet.period_min,
alpha=2.0 / 3.0,
testval=planet.period_min,
)
p_param = p_prior
tmax_prior = planet.t0
...
p_priors = pm.Deterministic(ORBITAL_PERIOD, tt.stack(p_priors_list))
tmax_priors = pm.Deterministic(TIME_END, tt.stack(tmax_priors_list))
...LOL
tic_entry.candidates[0].num_periods
>>> -9223372036854775808With the above changes, I was able to get the model to compile and run on the test val (no nans! woohoo!)
However, got the darn celerite2.backprop.LinAlgError: failed to factorize or solve matrix error again!
avivajpeyi
commented
2 years ago Hmm looking at the logs, I see that there are two 0s in the point being optimized:
{'dur_interval__': array([4.59231013]),
'r_log__': array([-3.01614327]),
'b_impact__': array([-0.09347175]),
't0_interval__': array([0.]),
'p_1_lowerbound__': array(4.99338746),
'f0': array(0.),
'u_quadlimbdark__': array([4.4408921e-16, 0.0000000e+00]),
'jitter_log__': array(-0.70142785),
'sigma_log__': array(19.9134664),
'rho_log__': array(30.25482971)}The error is being thrown at
theta = pmx.optimize(**kwargs, vars=noise_params)and
noise_params = [jitter_prior, sigma_prior, rho_prior]run_toi(1847) results with an error at the start of the sampling step, because p_1_lowerbound -inf
An error occurred while executing the following cell:
------------------
if tic_entry.inference_data is None:
inference_data = run_inference(planet_transit_model)
else:
logger.info("Using cached run")
inference_data = tic_entry.inference_data
inference_data
------------------
---------------------------------------------------------------------------
SamplingError Traceback (most recent call last)
/tmp/ipykernel_176089/141900171.py in <module>
1 if tic_entry.inference_data is None:
----> 2 inference_data = run_inference(planet_transit_model)
3 else:
4 logger.info("Using cached run")
5 inference_data = tic_entry.inference_data
/tmp/ipykernel_176089/3349958539.py in run_inference(model)
4 sampling_kwargs = dict(tune=2000, draws=2000, chains=2, cores=2)
5 logger.info(f"Run sampler with kwargs: {sampling_kwargs}")
----> 6 inference_data = pmx.sample(
7 **sampling_kwargs, start=init_params, return_inferencedata=True
8 )
/fred/oz200/avajpeyi/envs/tess/lib/python3.8/site-packages/pymc3_ext/sampling/sampling.py in sample(draws, tune, model, step_kwargs, warmup_window, adapt_window, cooldown_window, initial_accept, target_accept, gamma, k, t0, parameter_groups, adapt_type, recompute_interval, regularization_steps, regularization_variance, **kwargs)
95 )
96
---> 97 return pm.sample(draws=draws, tune=tune, model=model, step=step, **kwargs)
/fred/oz200/avajpeyi/envs/tess/lib/python3.8/site-packages/pymc3/sampling.py in sample(draws, step, init, n_init, start, trace, chain_idx, chains, cores, tune, progressbar, model, random_seed, discard_tuned_samples, compute_convergence_checks, callback, jitter_max_retries, return_inferencedata, idata_kwargs, mp_ctx, pickle_backend, **kwargs)
433 for chain_start_vals in start:
434 update_start_vals(chain_start_vals, model.test_point, model)
--> 435 check_start_vals(start, model)
436
437 if cores is None:
/fred/oz200/avajpeyi/envs/tess/lib/python3.8/site-packages/pymc3/util.py in check_start_vals(start, model)
235
236 if not np.all(np.isfinite(initial_eval)):
--> 237 raise SamplingError(
238 "Initial evaluation of model at starting point failed!\n"
239 "Starting values:\n{}\n\n"
SamplingError: Initial evaluation of model at starting point failed!
Starting values:
{'dur_interval__': array([-1.75828217, 0.66137893]), 'r_log__': array([-2.3121765, -2.9045715]), 'b_impact__': array([-0.18072936, -0.10394958]), 't0_interval__': array([0., 0.]), 'p_1_lowerbound__': array(-inf), 'tmax_2_interval__': array(0.), 'f0': array(0.), 'u_quadlimbdark__': array([4.4408921e-16, 0.0000000e+00]), 'jitter_log__': array(0.5556663), 'sigma_log__': array(0.5556663), 'rho_log__': array(0.05250777), 'dur': array([0.149375 , 0.08883333]), 'r': array([0.09904544, 0.05477226]), 'b': array([0.5, 0.5]), 't0': array([1390.70655 , 2130.150082]), 'p_1': array(750.87597302), 'tmax_2': array(2130.150082), 'p': array([750.87597302, nan]), 'tmax': array([1390.70655 , 2130.150082]), 'u': array([0.70710678, 0. ]), 'jitter': array(1.74310202), 'sigma': array(1.74310202), 'rho': array(1.05391075), 'rho_circ': array([128.85573981, nan])}
Initial evaluation results:
dur_interval__ 0.28
r_log__ -1.84
b_impact__ -2.78
t0_interval__ 0.47
p_1_lowerbound__ -inf
tmax_2_interval__ -0.51
f0 -3.22
u_quadlimbdark__ -2.77
jitter_log__ 0.11
sigma_log__ 0.11
rho_log__ -0.55
obs -37873.84
Name: Log-probability of test_point, dtype: float64
SamplingError: Initial evaluation of model at starting point failed!So the issue might be in the optimise_init_param step?
From the flux vs time plot, looks like we've just missed the transit?
notebook: toi_1847.ipynb.zip
^ the test_model function should also check for infs, not just isnull
avivajpeyi
commented
2 years ago Looks like im missing the transits for several single-transit TOIs!
Here are some examples:
avivajpeyi
commented
2 years ago Another set of funky errors -- the transit duration seems wrong?
avivajpeyi
commented
2 years ago The 77 analyses for single-transit systems have p_n_lowerbound = -inf
Eg from 1812:
avivajpeyi
commented
2 years ago The test-val is wrong -- currently set to prior's lower bound!
avivajpeyi
commented
2 years ago So although the periods are super long with even wider uncertainties -- the phase plots look ok?
I wonder if there is any more physics that we can use to constrain the period a bit more for these single-transit systems
avivajpeyi
commented
2 years ago Some more diagnostics:
Hhmm why are my ranges going so far out I wonder... i probably samples really far away from the 1,2,3 sigma contours
avivajpeyi
commented
2 years ago Finally, the trace plot:
looks like there are a few divergences
avivajpeyi
commented
2 years ago For single transits:
avivajpeyi
commented
2 years ago period -- are we just getting back the prior?
avivajpeyi
commented
2 years ago This could be a cool system to talk about + followup + maybe highlight this in the paper
There might be a way to deal with single transits now!
assert np.all(periods > 0), "We haven't implemented single transits yet"