diff --git a/example/GW150914.py b/example/GW150914.py deleted file mode 100644 index 559b5b7c..00000000 --- a/example/GW150914.py +++ /dev/null @@ -1,138 +0,0 @@ -import time - -import jax -import jax.numpy as jnp - -from jimgw.jim import Jim -from jimgw.prior import Composite, Unconstrained_Uniform -from jimgw.single_event.detector import H1, L1 -from jimgw.single_event.likelihood import TransientLikelihoodFD -from jimgw.single_event.waveform import RippleIMRPhenomD -from flowMC.strategy.optimization import optimization_Adam - -jax.config.update("jax_enable_x64", True) - -########################################### -########## First we grab data ############# -########################################### - -total_time_start = time.time() - -# first, fetch a 4s segment centered on GW150914 -gps = 1126259462.4 -duration = 4 -post_trigger_duration = 2 -start_pad = duration - post_trigger_duration -end_pad = post_trigger_duration -fmin = 20.0 -fmax = 1024.0 - -ifos = ["H1", "L1"] - -H1.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) -L1.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) - -Mc_prior = Unconstrained_Uniform(10.0, 80.0, naming=["M_c"]) -q_prior = Unconstrained_Uniform( - 0.125, - 1.0, - naming=["q"], - transforms={"q": ("eta", lambda params: params["q"] / (1 + params["q"]) ** 2)}, -) -s1z_prior = Unconstrained_Uniform(-1.0, 1.0, naming=["s1_z"]) -s2z_prior = Unconstrained_Uniform(-1.0, 1.0, naming=["s2_z"]) -dL_prior = Unconstrained_Uniform(0.0, 2000.0, naming=["d_L"]) -t_c_prior = Unconstrained_Uniform(-0.05, 0.05, naming=["t_c"]) -phase_c_prior = Unconstrained_Uniform(0.0, 2 * jnp.pi, naming=["phase_c"]) -cos_iota_prior = Unconstrained_Uniform( - -1.0, - 1.0, - naming=["cos_iota"], - transforms={ - "cos_iota": ( - "iota", - lambda params: jnp.arccos( - jnp.arcsin(jnp.sin(params["cos_iota"] / 2 * jnp.pi)) * 2 / jnp.pi - ), - ) - }, -) -psi_prior = Unconstrained_Uniform(0.0, jnp.pi, naming=["psi"]) -ra_prior = Unconstrained_Uniform(0.0, 2 * jnp.pi, naming=["ra"]) -sin_dec_prior = Unconstrained_Uniform( - -1.0, - 1.0, - naming=["sin_dec"], - transforms={ - "sin_dec": ( - "dec", - lambda params: jnp.arcsin( - jnp.arcsin(jnp.sin(params["sin_dec"] / 2 * jnp.pi)) * 2 / jnp.pi - ), - ) - }, -) - -prior = Composite( - [ - Mc_prior, - q_prior, - s1z_prior, - s2z_prior, - dL_prior, - t_c_prior, - phase_c_prior, - cos_iota_prior, - psi_prior, - ra_prior, - sin_dec_prior, - ] -) -likelihood = TransientLikelihoodFD( - [H1, L1], - waveform=RippleIMRPhenomD(), - trigger_time=gps, - duration=4, - post_trigger_duration=2, -) - - -mass_matrix = jnp.eye(11) -mass_matrix = mass_matrix.at[1, 1].set(1e-3) -mass_matrix = mass_matrix.at[5, 5].set(1e-3) -local_sampler_arg = {"step_size": mass_matrix * 3e-3} - -Adam_optimizer = optimization_Adam(n_steps=3000, learning_rate=0.01, noise_level=1) - -import optax -n_epochs = 20 -n_loop_training = 100 -total_epochs = n_epochs * n_loop_training -start = total_epochs//10 -learning_rate = optax.polynomial_schedule( - 1e-3, 1e-4, 4.0, total_epochs - start, transition_begin=start -) - - -jim = Jim( - likelihood, - prior, - n_loop_training=n_loop_training, - n_loop_production=20, - n_local_steps=10, - n_global_steps=1000, - n_chains=500, - n_epochs=n_epochs, - learning_rate=learning_rate, - n_max_examples=30000, - n_flow_samples=100000, - momentum=0.9, - batch_size=30000, - use_global=True, - train_thinning=1, - output_thinning=10, - local_sampler_arg=local_sampler_arg, - strategies=[Adam_optimizer,"default"], -) - -jim.sample(jax.random.PRNGKey(42)) diff --git a/example/GW150914_D.py b/example/GW150914_D.py new file mode 100644 index 00000000..efa58b04 --- /dev/null +++ b/example/GW150914_D.py @@ -0,0 +1,152 @@ +import jax +import jax.numpy as jnp + +from jimgw.jim import Jim +from jimgw.prior import CombinePrior, UniformPrior, CosinePrior, SinePrior, PowerLawPrior +from jimgw.single_event.detector import H1, L1 +from jimgw.single_event.likelihood import TransientLikelihoodFD +from jimgw.single_event.waveform import RippleIMRPhenomD +from jimgw.transforms import BoundToUnbound +from jimgw.single_event.transforms import ComponentMassesToChirpMassSymmetricMassRatioTransform, SkyFrameToDetectorFrameSkyPositionTransform, ComponentMassesToChirpMassMassRatioTransform +from jimgw.single_event.utils import Mc_q_to_m1_m2 +from flowMC.strategy.optimization import optimization_Adam + +jax.config.update("jax_enable_x64", True) + +########################################### +########## First we grab data ############# +########################################### + +# first, fetch a 4s segment centered on GW150914 +gps = 1126259462.4 +duration = 4 +post_trigger_duration = 2 +start_pad = duration - post_trigger_duration +end_pad = post_trigger_duration +fmin = 20.0 +fmax = 1024.0 + +ifos = [H1, L1] + +for ifo in ifos: + ifo.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) + +M_c_min, M_c_max = 10.0, 80.0 +q_min, q_max = 0.125, 1.0 +m_1_prior = UniformPrior(Mc_q_to_m1_m2(M_c_min, q_max)[0], Mc_q_to_m1_m2(M_c_max, q_min)[0], parameter_names=["m_1"]) +m_2_prior = UniformPrior(Mc_q_to_m1_m2(M_c_min, q_min)[1], Mc_q_to_m1_m2(M_c_max, q_max)[1], parameter_names=["m_2"]) +s1z_prior = UniformPrior(-1.0, 1.0, parameter_names=["s1_z"]) +s2z_prior = UniformPrior(-1.0, 1.0, parameter_names=["s2_z"]) +dL_prior = PowerLawPrior(1.0, 2000.0, 2.0, parameter_names=["d_L"]) +t_c_prior = UniformPrior(-0.05, 0.05, parameter_names=["t_c"]) +phase_c_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["phase_c"]) +iota_prior = SinePrior(parameter_names=["iota"]) +psi_prior = UniformPrior(0.0, jnp.pi, parameter_names=["psi"]) +ra_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["ra"]) +dec_prior = CosinePrior(parameter_names=["dec"]) + +prior = CombinePrior( + [ + m_1_prior, + m_2_prior, + s1z_prior, + s2z_prior, + dL_prior, + t_c_prior, + phase_c_prior, + iota_prior, + psi_prior, + ra_prior, + dec_prior, + ] +) + +sample_transforms = [ + ComponentMassesToChirpMassMassRatioTransform(name_mapping=[["m_1", "m_2"], ["M_c", "q"]]), + BoundToUnbound(name_mapping = [["M_c"], ["M_c_unbounded"]], original_lower_bound=M_c_min, original_upper_bound=M_c_max), + BoundToUnbound(name_mapping = [["q"], ["q_unbounded"]], original_lower_bound=q_min, original_upper_bound=q_max), + BoundToUnbound(name_mapping = [["s1_z"], ["s1_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), + BoundToUnbound(name_mapping = [["s2_z"], ["s2_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), + BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=1.0, original_upper_bound=2000.0), + BoundToUnbound(name_mapping = [["t_c"], ["t_c_unbounded"]] , original_lower_bound=-0.05, original_upper_bound=0.05), + BoundToUnbound(name_mapping = [["phase_c"], ["phase_c_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), + BoundToUnbound(name_mapping = [["iota"], ["iota_unbounded"]], original_lower_bound=0., original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["psi"], ["psi_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), + SkyFrameToDetectorFrameSkyPositionTransform(name_mapping = [["ra", "dec"], ["zenith", "azimuth"]], gps_time=gps, ifos=ifos), + BoundToUnbound(name_mapping = [["zenith"], ["zenith_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["azimuth"], ["azimuth_unbounded"]], original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), +] + +likelihood_transforms = [ + ComponentMassesToChirpMassSymmetricMassRatioTransform(name_mapping=[["m_1", "m_2"], ["M_c", "eta"]]), +] + +likelihood = TransientLikelihoodFD( + ifos, + waveform=RippleIMRPhenomD(), + trigger_time=gps, + duration=4, + post_trigger_duration=2, +) + + +mass_matrix = jnp.eye(11) +mass_matrix = mass_matrix.at[1, 1].set(1e-3) +mass_matrix = mass_matrix.at[5, 5].set(1e-3) +local_sampler_arg = {"step_size": mass_matrix * 3e-3} + +Adam_optimizer = optimization_Adam(n_steps=3000, learning_rate=0.01, noise_level=1) + +n_epochs = 30 +n_loop_training = 100 +learning_rate = 1e-4 + + +jim = Jim( + likelihood, + prior, + sample_transforms=sample_transforms, + likelihood_transforms=likelihood_transforms, + n_loop_training=n_loop_training, + n_loop_production=20, + n_local_steps=10, + n_global_steps=1000, + n_chains=500, + n_epochs=n_epochs, + learning_rate=learning_rate, + n_max_examples=30000, + n_flow_samples=100000, + momentum=0.9, + batch_size=30000, + use_global=True, + train_thinning=1, + output_thinning=10, + local_sampler_arg=local_sampler_arg, + strategies=[Adam_optimizer, "default"], +) + +jim.sample(jax.random.PRNGKey(42)) +jim.get_samples() +jim.print_summary() + + +########################################### +########## Visualize the Data ############# +########################################### +import corner +import matplotlib.pyplot as plt +import numpy as np + +production_summary = jim.sampler.get_sampler_state(training=False) +production_chain = production_summary["chains"].reshape(-1, len(jim.parameter_names)).T +if jim.sample_transforms: + transformed_chain = jim.add_name(production_chain) + for transform in reversed(jim.sample_transforms): + transformed_chain = transform.backward(transformed_chain) +result = transformed_chain +labels = list(transformed_chain.keys()) + +samples = np.array(list(result.values())).reshape(int(len(labels)), -1) # flatten the array +transposed_array = samples.T # transpose the array +figure = corner.corner(transposed_array, labels=labels, plot_datapoints=False, title_quantiles=[0.16, 0.5, 0.84], show_titles=True, title_fmt='g', use_math_text=True) +plt.savefig("GW1500914_D.jpeg") diff --git a/example/GW150914_D_heterodyne.py b/example/GW150914_D_heterodyne.py new file mode 100644 index 00000000..91e04ba2 --- /dev/null +++ b/example/GW150914_D_heterodyne.py @@ -0,0 +1,157 @@ +import jax +import jax.numpy as jnp + +from jimgw.jim import Jim +from jimgw.prior import CombinePrior, UniformPrior, CosinePrior, SinePrior, PowerLawPrior +from jimgw.single_event.detector import H1, L1 +from jimgw.single_event.likelihood import HeterodynedTransientLikelihoodFD +from jimgw.single_event.waveform import RippleIMRPhenomD +from jimgw.transforms import BoundToUnbound +from jimgw.single_event.transforms import ComponentMassesToChirpMassSymmetricMassRatioTransform, SkyFrameToDetectorFrameSkyPositionTransform, ComponentMassesToChirpMassMassRatioTransform +from jimgw.single_event.utils import Mc_q_to_m1_m2 +from flowMC.strategy.optimization import optimization_Adam + +jax.config.update("jax_enable_x64", True) + +########################################### +########## First we grab data ############# +########################################### + +# first, fetch a 4s segment centered on GW150914 +gps = 1126259462.4 +duration = 4 +post_trigger_duration = 2 +start_pad = duration - post_trigger_duration +end_pad = post_trigger_duration +fmin = 20.0 +fmax = 1024.0 + +ifos = [H1, L1] + +for ifo in ifos: + ifo.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) + +M_c_min, M_c_max = 10.0, 80.0 +q_min, q_max = 0.125, 1.0 +m_1_prior = UniformPrior(Mc_q_to_m1_m2(M_c_min, q_max)[0], Mc_q_to_m1_m2(M_c_max, q_min)[0], parameter_names=["m_1"]) +m_2_prior = UniformPrior(Mc_q_to_m1_m2(M_c_min, q_min)[1], Mc_q_to_m1_m2(M_c_max, q_max)[1], parameter_names=["m_2"]) +s1z_prior = UniformPrior(-1.0, 1.0, parameter_names=["s1_z"]) +s2z_prior = UniformPrior(-1.0, 1.0, parameter_names=["s2_z"]) +dL_prior = PowerLawPrior(1.0, 2000.0, 2.0, parameter_names=["d_L"]) +t_c_prior = UniformPrior(-0.05, 0.05, parameter_names=["t_c"]) +phase_c_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["phase_c"]) +iota_prior = SinePrior(parameter_names=["iota"]) +psi_prior = UniformPrior(0.0, jnp.pi, parameter_names=["psi"]) +ra_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["ra"]) +dec_prior = CosinePrior(parameter_names=["dec"]) + +prior = CombinePrior( + [ + m_1_prior, + m_2_prior, + s1z_prior, + s2z_prior, + dL_prior, + t_c_prior, + phase_c_prior, + iota_prior, + psi_prior, + ra_prior, + dec_prior, + ] +) + +sample_transforms = [ + ComponentMassesToChirpMassMassRatioTransform(name_mapping=[["m_1", "m_2"], ["M_c", "q"]]), + BoundToUnbound(name_mapping = [["M_c"], ["M_c_unbounded"]], original_lower_bound=M_c_min, original_upper_bound=M_c_max), + BoundToUnbound(name_mapping = [["q"], ["q_unbounded"]], original_lower_bound=q_min, original_upper_bound=q_max), + BoundToUnbound(name_mapping = [["s1_z"], ["s1_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), + BoundToUnbound(name_mapping = [["s2_z"], ["s2_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), + BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=1.0, original_upper_bound=2000.0), + BoundToUnbound(name_mapping = [["t_c"], ["t_c_unbounded"]] , original_lower_bound=-0.05, original_upper_bound=0.05), + BoundToUnbound(name_mapping = [["phase_c"], ["phase_c_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), + BoundToUnbound(name_mapping = [["iota"], ["iota_unbounded"]], original_lower_bound=0., original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["psi"], ["psi_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), + SkyFrameToDetectorFrameSkyPositionTransform(name_mapping = [["ra", "dec"], ["zenith", "azimuth"]], gps_time=gps, ifos=ifos), + BoundToUnbound(name_mapping = [["zenith"], ["zenith_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["azimuth"], ["azimuth_unbounded"]], original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), +] + +likelihood_transforms = [ + ComponentMassesToChirpMassSymmetricMassRatioTransform(name_mapping=[["m_1", "m_2"], ["M_c", "eta"]]), +] + +likelihood = HeterodynedTransientLikelihoodFD( + ifos, + prior=prior, + waveform=RippleIMRPhenomD(), + trigger_time=gps, + duration=4, + post_trigger_duration=2, + sample_transforms=sample_transforms, + likelihood_transforms=likelihood_transforms, + n_steps=5, + popsize=10, +) + + +mass_matrix = jnp.eye(11) +mass_matrix = mass_matrix.at[1, 1].set(1e-3) +mass_matrix = mass_matrix.at[5, 5].set(1e-3) +local_sampler_arg = {"step_size": mass_matrix * 3e-3} + +Adam_optimizer = optimization_Adam(n_steps=3000, learning_rate=0.01, noise_level=1) + +n_epochs = 30 +n_loop_training = 100 +learning_rate = 1e-4 + + +jim = Jim( + likelihood, + prior, + sample_transforms=sample_transforms, + likelihood_transforms=likelihood_transforms, + n_loop_training=n_loop_training, + n_loop_production=20, + n_local_steps=10, + n_global_steps=1000, + n_chains=500, + n_epochs=n_epochs, + learning_rate=learning_rate, + n_max_examples=30000, + n_flow_samples=100000, + momentum=0.9, + batch_size=30000, + use_global=True, + train_thinning=1, + output_thinning=10, + local_sampler_arg=local_sampler_arg, + strategies=[Adam_optimizer, "default"], +) + +jim.sample(jax.random.PRNGKey(42)) +jim.get_samples() +jim.print_summary() + + +########################################### +########## Visualize the Data ############# +########################################### +import corner +import matplotlib.pyplot as plt +import numpy as np + +production_summary = jim.sampler.get_sampler_state(training=False) +production_chain = production_summary["chains"].reshape(-1, len(jim.parameter_names)).T +if jim.sample_transforms: + transformed_chain = jim.add_name(production_chain) + for transform in reversed(jim.sample_transforms): + transformed_chain = transform.backward(transformed_chain) +result = transformed_chain +labels = list(transformed_chain.keys()) + +samples = np.array(list(result.values())).reshape(int(len(labels)), -1) # flatten the array +transposed_array = samples.T # transpose the array +figure = corner.corner(transposed_array, labels=labels, plot_datapoints=False, title_quantiles=[0.16, 0.5, 0.84], show_titles=True, title_fmt='g', use_math_text=True) +plt.savefig("GW1500914_D_heterodyne.jpeg") diff --git a/example/GW150914_PV2.py b/example/GW150914_PV2.py deleted file mode 100644 index 06209ba6..00000000 --- a/example/GW150914_PV2.py +++ /dev/null @@ -1,165 +0,0 @@ -import time - -import jax -import jax.numpy as jnp - -from jimgw.jim import Jim -from jimgw.prior import Composite, Sphere, Unconstrained_Uniform -from jimgw.single_event.detector import H1, L1 -from jimgw.single_event.likelihood import TransientLikelihoodFD -from jimgw.single_event.waveform import RippleIMRPhenomPv2 -from flowMC.strategy.optimization import optimization_Adam - - -jax.config.update("jax_enable_x64", True) - -########################################### -########## First we grab data ############# -########################################### - -total_time_start = time.time() - -# first, fetch a 4s segment centered on GW150914 -gps = 1126259462.4 -start = gps - 2 -end = gps + 2 -fmin = 20.0 -fmax = 1024.0 - -ifos = ["H1", "L1"] - -H1.load_data(gps, 2, 2, fmin, fmax, psd_pad=16, tukey_alpha=0.2) -L1.load_data(gps, 2, 2, fmin, fmax, psd_pad=16, tukey_alpha=0.2) - -waveform = RippleIMRPhenomPv2(f_ref=20) - -########################################### -########## Set up priors ################## -########################################### - -Mc_prior = Unconstrained_Uniform(10.0, 80.0, naming=["M_c"]) -q_prior = Unconstrained_Uniform( - 0.125, - 1.0, - naming=["q"], - transforms={"q": ("eta", lambda params: params["q"] / (1 + params["q"]) ** 2)}, -) -s1_prior = Sphere(naming="s1") -s2_prior = Sphere(naming="s2") -dL_prior = Unconstrained_Uniform(0.0, 2000.0, naming=["d_L"]) -t_c_prior = Unconstrained_Uniform(-0.05, 0.05, naming=["t_c"]) -phase_c_prior = Unconstrained_Uniform(0.0, 2 * jnp.pi, naming=["phase_c"]) -cos_iota_prior = Unconstrained_Uniform( - -1.0, - 1.0, - naming=["cos_iota"], - transforms={ - "cos_iota": ( - "iota", - lambda params: jnp.arccos( - jnp.arcsin(jnp.sin(params["cos_iota"] / 2 * jnp.pi)) * 2 / jnp.pi - ), - ) - }, -) -psi_prior = Unconstrained_Uniform(0.0, jnp.pi, naming=["psi"]) -ra_prior = Unconstrained_Uniform(0.0, 2 * jnp.pi, naming=["ra"]) -sin_dec_prior = Unconstrained_Uniform( - -1.0, - 1.0, - naming=["sin_dec"], - transforms={ - "sin_dec": ( - "dec", - lambda params: jnp.arcsin( - jnp.arcsin(jnp.sin(params["sin_dec"] / 2 * jnp.pi)) * 2 / jnp.pi - ), - ) - }, -) - -prior = Composite( - [ - Mc_prior, - q_prior, - s1_prior, - s2_prior, - dL_prior, - t_c_prior, - phase_c_prior, - cos_iota_prior, - psi_prior, - ra_prior, - sin_dec_prior, - ], -) - -epsilon = 1e-3 -bounds = jnp.array( - [ - [10.0, 80.0], - [0.125, 1.0], - [0, jnp.pi], - [0, 2 * jnp.pi], - [0.0, 1.0], - [0, jnp.pi], - [0, 2 * jnp.pi], - [0.0, 1.0], - [0.0, 2000], - [-0.05, 0.05], - [0.0, 2 * jnp.pi], - [-1.0, 1.0], - [0.0, jnp.pi], - [0.0, 2 * jnp.pi], - [-1.0, 1.0], - ] -) + jnp.array([[epsilon, -epsilon]]) - -likelihood = TransientLikelihoodFD( - [H1, L1], waveform=waveform, trigger_time=gps, duration=4, post_trigger_duration=2 -) -# likelihood = HeterodynedTransientLikelihoodFD([H1, L1], prior=prior, bounds=bounds, waveform=waveform, trigger_time=gps, duration=4, post_trigger_duration=2) - - -mass_matrix = jnp.eye(prior.n_dim) -mass_matrix = mass_matrix.at[1, 1].set(1e-3) -mass_matrix = mass_matrix.at[9, 9].set(1e-3) -local_sampler_arg = {"step_size": mass_matrix * 1e-3} - -Adam_optimizer = optimization_Adam(n_steps=3000, learning_rate=0.01, noise_level=1, bounds=bounds) - -import optax -n_epochs = 20 -n_loop_training = 100 -total_epochs = n_epochs * n_loop_training -start = total_epochs//10 -learning_rate = optax.polynomial_schedule( - 1e-3, 1e-4, 4.0, total_epochs - start, transition_begin=start -) - -jim = Jim( - likelihood, - prior, - n_loop_training=n_loop_training, - n_loop_production=20, - n_local_steps=10, - n_global_steps=1000, - n_chains=500, - n_epochs=n_epochs, - learning_rate=learning_rate, - n_max_examples=30000, - n_flow_sample=100000, - momentum=0.9, - batch_size=30000, - use_global=True, - keep_quantile=0.0, - train_thinning=1, - output_thinning=10, - local_sampler_arg=local_sampler_arg, - # strategies=[Adam_optimizer,"default"], -) - -import numpy as np -# chains = np.load('./GW150914_init.npz')['chain'] - -jim.sample(jax.random.PRNGKey(42))#,initial_guess=chains) diff --git a/example/GW150914_Pv2.py b/example/GW150914_Pv2.py new file mode 100644 index 00000000..c922c822 --- /dev/null +++ b/example/GW150914_Pv2.py @@ -0,0 +1,177 @@ +import time + +import jax +import jax.numpy as jnp + +from jimgw.jim import Jim +from jimgw.prior import CombinePrior, UniformPrior, CosinePrior, SinePrior, PowerLawPrior +from jimgw.single_event.detector import H1, L1 +from jimgw.single_event.likelihood import TransientLikelihoodFD +from jimgw.single_event.waveform import RippleIMRPhenomPv2 +from jimgw.transforms import BoundToUnbound +from jimgw.single_event.transforms import MassRatioToSymmetricMassRatioTransform, SpinToCartesianSpinTransform, ComponentMassesToChirpMassMassRatioTransform, SkyFrameToDetectorFrameSkyPositionTransform +from jimgw.single_event.utils import Mc_q_to_m1_m2 +from flowMC.strategy.optimization import optimization_Adam + +jax.config.update("jax_enable_x64", True) + +########################################### +########## First we grab data ############# +########################################### + +total_time_start = time.time() + +# first, fetch a 4s segment centered on GW150914 +gps = 1126259462.4 +duration = 4 +post_trigger_duration = 2 +start_pad = duration - post_trigger_duration +end_pad = post_trigger_duration +fmin = 20.0 +fmax = 1024.0 + +ifos = [H1, L1] + +f_ref = 20.0 + +for ifo in ifos: + ifo.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) + +M_c_min, M_c_max = 10.0, 80.0 +q_min, q_max = 0.125, 1.0 +m_1_prior = UniformPrior(Mc_q_to_m1_m2(M_c_min, q_max)[0], Mc_q_to_m1_m2(M_c_max, q_min)[0], parameter_names=["m_1"]) +m_2_prior = UniformPrior(Mc_q_to_m1_m2(M_c_min, q_min)[1], Mc_q_to_m1_m2(M_c_max, q_max)[1], parameter_names=["m_2"]) +theta_jn_prior = SinePrior(parameter_names=["theta_jn"]) +phi_jl_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["phi_jl"]) +theta_1_prior = SinePrior(parameter_names=["theta_1"]) +theta_2_prior = SinePrior(parameter_names=["theta_2"]) +phi_12_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["phi_12"]) +a_1_prior = UniformPrior(0.0, 1.0, parameter_names=["a_1"]) +a_2_prior = UniformPrior(0.0, 1.0, parameter_names=["a_2"]) +dL_prior = PowerLawPrior(10.0, 2000.0, 2.0, parameter_names=["d_L"]) +t_c_prior = UniformPrior(-0.05, 0.05, parameter_names=["t_c"]) +phase_c_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["phase_c"]) +psi_prior = UniformPrior(0.0, jnp.pi, parameter_names=["psi"]) +ra_prior = UniformPrior(0.0, 2 * jnp.pi, parameter_names=["ra"]) +dec_prior = CosinePrior(parameter_names=["dec"]) + +prior = CombinePrior( + [ + m_1_prior, + m_2_prior, + theta_jn_prior, + phi_jl_prior, + theta_1_prior, + theta_2_prior, + phi_12_prior, + a_1_prior, + a_2_prior, + dL_prior, + t_c_prior, + phase_c_prior, + psi_prior, + ra_prior, + dec_prior, + ] +) + +sample_transforms = [ + ComponentMassesToChirpMassMassRatioTransform(name_mapping=[["m_1", "m_2"], ["M_c", "q"]]), + BoundToUnbound(name_mapping = [["M_c"], ["M_c_unbounded"]], original_lower_bound=10.0, original_upper_bound=80.0), + BoundToUnbound(name_mapping = [["q"], ["q_unbounded"]], original_lower_bound=0.125, original_upper_bound=1.), + BoundToUnbound(name_mapping = [["theta_jn"], ["theta_jn_unbounded"]] , original_lower_bound=0.0, original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["phi_jl"], ["phi_jl_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), + BoundToUnbound(name_mapping = [["theta_1"], ["theta_1_unbounded"]] , original_lower_bound=0.0, original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["theta_2"], ["theta_2_unbounded"]] , original_lower_bound=0.0, original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["phi_12"], ["phi_12_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), + BoundToUnbound(name_mapping = [["a_1"], ["a_1_unbounded"]] , original_lower_bound=0.0, original_upper_bound=1.0), + BoundToUnbound(name_mapping = [["a_2"], ["a_2_unbounded"]] , original_lower_bound=0.0, original_upper_bound=1.0), + BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=10.0, original_upper_bound=2000.0), + BoundToUnbound(name_mapping = [["t_c"], ["t_c_unbounded"]] , original_lower_bound=-0.05, original_upper_bound=0.05), + BoundToUnbound(name_mapping = [["phase_c"], ["phase_c_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), + BoundToUnbound(name_mapping = [["psi"], ["psi_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), + SkyFrameToDetectorFrameSkyPositionTransform(name_mapping = [["ra", "dec"], ["zenith", "azimuth"]], gps_time=gps, ifos=ifos), + BoundToUnbound(name_mapping = [["zenith"], ["zenith_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), + BoundToUnbound(name_mapping = [["azimuth"], ["azimuth_unbounded"]], original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), +] + +likelihood_transforms = [ + ComponentMassesToChirpMassMassRatioTransform(name_mapping=[["m_1", "m_2"], ["M_c", "q"]]), + SpinToCartesianSpinTransform(name_mapping=[["theta_jn", "phi_jl", "theta_1", "theta_2", "phi_12", "a_1", "a_2"], ["iota", "s1_x", "s1_y", "s1_z", "s2_x", "s2_y", "s2_z"]], freq_ref=f_ref), + MassRatioToSymmetricMassRatioTransform(name_mapping=[["q"], ["eta"]]), +] + +likelihood = TransientLikelihoodFD( + ifos, + waveform=RippleIMRPhenomPv2(f_ref=f_ref), + trigger_time=gps, + duration=4, + post_trigger_duration=2, +) + + +mass_matrix = jnp.eye(15) +mass_matrix = mass_matrix.at[1, 1].set(1e-3) +mass_matrix = mass_matrix.at[9, 9].set(1e-3) +local_sampler_arg = {"step_size": mass_matrix * 1e-3} + +Adam_optimizer = optimization_Adam(n_steps=3000, learning_rate=0.01, noise_level=1) + +n_epochs = 30 +n_loop_training = 100 +learning_rate = 1e-4 + + +jim = Jim( + likelihood, + prior, + sample_transforms=sample_transforms, + likelihood_transforms=likelihood_transforms, + n_loop_training=n_loop_training, + n_loop_production=20, + n_local_steps=10, + n_global_steps=1000, + n_chains=500, + n_epochs=n_epochs, + learning_rate=learning_rate, + n_max_examples=30000, + n_flow_samples=100000, + momentum=0.9, + batch_size=30000, + use_global=True, + train_thinning=1, + output_thinning=10, + local_sampler_arg=local_sampler_arg, + strategies=[Adam_optimizer, "default"], +) + +jim.sample(jax.random.PRNGKey(42)) +jim.get_samples() +jim.print_summary() + +########################################### +########## Visualize the Data ############# +########################################### +import corner +import matplotlib.pyplot as plt +import numpy as np + +production_summary = jim.sampler.get_sampler_state(training=False) +production_chain = production_summary["chains"].reshape(-1, len(jim.parameter_names)).T +if jim.sample_transforms: + transformed_chain = jim.add_name(production_chain) + for transform in reversed(jim.sample_transforms): + transformed_chain = transform.backward(transformed_chain) +result = transformed_chain +labels = list(transformed_chain.keys()) + +samples = np.array(list(result.values())).reshape(int(len(labels)), -1) # flatten the array +transposed_array = samples.T # transpose the array +figure = corner.corner(transposed_array, labels=labels, plot_datapoints=False, title_quantiles=[0.16, 0.5, 0.84], show_titles=True, title_fmt='g', use_math_text=True) +plt.savefig("GW1500914_Pv2.jpeg") + +########################################### +############# Save the Run ################ +########################################### +# import pickle +# pickle.dump(result, open("GW150914_pv2.pkl", "wb"), protocol=pickle.HIGHEST_PROTOCOL) diff --git a/example/GW150914_heterodyne.py b/example/GW150914_heterodyne.py deleted file mode 100644 index c1faed03..00000000 --- a/example/GW150914_heterodyne.py +++ /dev/null @@ -1,158 +0,0 @@ -import time - -import jax -import jax.numpy as jnp - -from jimgw.jim import Jim -from jimgw.prior import Composite, Unconstrained_Uniform -from jimgw.single_event.detector import H1, L1 -from jimgw.single_event.likelihood import ( - HeterodynedTransientLikelihoodFD, - TransientLikelihoodFD, -) -from jimgw.single_event.waveform import RippleIMRPhenomD -from flowMC.strategy.optimization import optimization_Adam - -jax.config.update("jax_enable_x64", True) - -########################################### -########## First we grab data ############# -########################################### - -total_time_start = time.time() - -# first, fetch a 4s segment centered on GW150914 -gps = 1126259462.4 -duration = 4 -post_trigger_duration = 2 -start_pad = duration - post_trigger_duration -end_pad = post_trigger_duration -fmin = 20.0 -fmax = 1024.0 - -ifos = ["H1", "L1"] - -H1.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) -L1.load_data(gps, start_pad, end_pad, fmin, fmax, psd_pad=16, tukey_alpha=0.2) - -Mc_prior = Unconstrained_Uniform(10.0, 80.0, naming=["M_c"]) -q_prior = Unconstrained_Uniform( - 0.125, - 1.0, - naming=["q"], - transforms={"q": ("eta", lambda params: params["q"] / (1 + params["q"]) ** 2)}, -) -s1z_prior = Unconstrained_Uniform(-1.0, 1.0, naming=["s1_z"]) -s2z_prior = Unconstrained_Uniform(-1.0, 1.0, naming=["s2_z"]) -dL_prior = Unconstrained_Uniform(0.0, 2000.0, naming=["d_L"]) -t_c_prior = Unconstrained_Uniform(-0.05, 0.05, naming=["t_c"]) -phase_c_prior = Unconstrained_Uniform(0.0, 2 * jnp.pi, naming=["phase_c"]) -cos_iota_prior = Unconstrained_Uniform( - -1.0, - 1.0, - naming=["cos_iota"], - transforms={ - "cos_iota": ( - "iota", - lambda params: jnp.arccos( - jnp.arcsin(jnp.sin(params["cos_iota"] / 2 * jnp.pi)) * 2 / jnp.pi - ), - ) - }, -) -psi_prior = Unconstrained_Uniform(0.0, jnp.pi, naming=["psi"]) -ra_prior = Unconstrained_Uniform(0.0, 2 * jnp.pi, naming=["ra"]) -sin_dec_prior = Unconstrained_Uniform( - -1.0, - 1.0, - naming=["sin_dec"], - transforms={ - "sin_dec": ( - "dec", - lambda params: jnp.arcsin( - jnp.arcsin(jnp.sin(params["sin_dec"] / 2 * jnp.pi)) * 2 / jnp.pi - ), - ) - }, -) - -prior = Composite( - [ - Mc_prior, - q_prior, - s1z_prior, - s2z_prior, - dL_prior, - t_c_prior, - phase_c_prior, - cos_iota_prior, - psi_prior, - ra_prior, - sin_dec_prior, - ] -) - -bounds = jnp.array( - [ - [10.0, 80.0], - [0.125, 1.0], - [-1.0, 1.0], - [-1.0, 1.0], - [0.0, 2000.0], - [-0.05, 0.05], - [0.0, 2 * jnp.pi], - [-1.0, 1.0], - [0.0, jnp.pi], - [0.0, 2 * jnp.pi], - [-1.0, 1.0], - ] -) - -likelihood = HeterodynedTransientLikelihoodFD( - [H1, L1], - prior=prior, - bounds=bounds, - waveform=RippleIMRPhenomD(), - trigger_time=gps, - duration=duration, - post_trigger_duration=post_trigger_duration, - n_steps=3000, -) - -mass_matrix = jnp.eye(11) -mass_matrix = mass_matrix.at[1, 1].set(1e-3) -mass_matrix = mass_matrix.at[5, 5].set(1e-3) -local_sampler_arg = {"step_size": mass_matrix * 3e-3} - -Adam_optimizer = optimization_Adam(n_steps=3000, learning_rate=0.01, noise_level=1) -import optax -n_epochs = 20 -n_loop_training = 100 -total_epochs = n_epochs * n_loop_training -start = total_epochs//10 -learning_rate = optax.polynomial_schedule( - 1e-3, 1e-4, 4.0, total_epochs - start, transition_begin=start -) - -jim = Jim( - likelihood, - prior, - n_loop_training=n_loop_training, - n_loop_production=20, - n_local_steps=10, - n_global_steps=1000, - n_chains=500, - n_epochs=n_epochs, - learning_rate=learning_rate, - n_max_examples=30000, - n_flow_sample=100000, - momentum=0.9, - batch_size=30000, - use_global=True, - keep_quantile=0.0, - train_thinning=1, - output_thinning=10, - local_sampler_arg=local_sampler_arg, - # strategies=[Adam_optimizer,"default"], -) -jim.sample(jax.random.PRNGKey(42)) diff --git a/test/integration/test_GW150914_D.py b/test/integration/test_GW150914_D.py index e1eee9ac..24dac718 100644 --- a/test/integration/test_GW150914_D.py +++ b/test/integration/test_GW150914_D.py @@ -67,7 +67,7 @@ BoundToUnbound(name_mapping = [["q"], ["q_unbounded"]], original_lower_bound=q_min, original_upper_bound=q_max), BoundToUnbound(name_mapping = [["s1_z"], ["s1_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), BoundToUnbound(name_mapping = [["s2_z"], ["s2_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), - BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2000.0), + BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=1.0, original_upper_bound=2000.0), BoundToUnbound(name_mapping = [["t_c"], ["t_c_unbounded"]] , original_lower_bound=-0.05, original_upper_bound=0.05), BoundToUnbound(name_mapping = [["phase_c"], ["phase_c_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), BoundToUnbound(name_mapping = [["iota"], ["iota_unbounded"]], original_lower_bound=0., original_upper_bound=jnp.pi), diff --git a/test/integration/test_GW150914_D_heterodyne.py b/test/integration/test_GW150914_D_heterodyne.py index bf97efdb..ff5cc57b 100644 --- a/test/integration/test_GW150914_D_heterodyne.py +++ b/test/integration/test_GW150914_D_heterodyne.py @@ -67,7 +67,7 @@ BoundToUnbound(name_mapping = [["q"], ["q_unbounded"]], original_lower_bound=q_min, original_upper_bound=q_max), BoundToUnbound(name_mapping = [["s1_z"], ["s1_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), BoundToUnbound(name_mapping = [["s2_z"], ["s2_z_unbounded"]] , original_lower_bound=-1.0, original_upper_bound=1.0), - BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2000.0), + BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=1.0, original_upper_bound=2000.0), BoundToUnbound(name_mapping = [["t_c"], ["t_c_unbounded"]] , original_lower_bound=-0.05, original_upper_bound=0.05), BoundToUnbound(name_mapping = [["phase_c"], ["phase_c_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), BoundToUnbound(name_mapping = [["iota"], ["iota_unbounded"]], original_lower_bound=0., original_upper_bound=jnp.pi), diff --git a/test/integration/test_GW150914_Pv2.py b/test/integration/test_GW150914_Pv2.py index c9d83a5e..392ef131 100644 --- a/test/integration/test_GW150914_Pv2.py +++ b/test/integration/test_GW150914_Pv2.py @@ -7,7 +7,7 @@ from jimgw.prior import CombinePrior, UniformPrior, CosinePrior, SinePrior, PowerLawPrior from jimgw.single_event.detector import H1, L1 from jimgw.single_event.likelihood import TransientLikelihoodFD -from jimgw.single_event.waveform import RippleIMRPhenomD +from jimgw.single_event.waveform import RippleIMRPhenomPv2 from jimgw.transforms import BoundToUnbound from jimgw.single_event.transforms import MassRatioToSymmetricMassRatioTransform, SpinToCartesianSpinTransform from flowMC.strategy.optimization import optimization_Adam @@ -29,6 +29,8 @@ fmin = 20.0 fmax = 1024.0 +f_ref = 20.0 + ifos = [H1, L1] for ifo in ifos: @@ -80,7 +82,7 @@ BoundToUnbound(name_mapping = [["phi_12"], ["phi_12_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), BoundToUnbound(name_mapping = [["a_1"], ["a_1_unbounded"]] , original_lower_bound=0.0, original_upper_bound=1.0), BoundToUnbound(name_mapping = [["a_2"], ["a_2_unbounded"]] , original_lower_bound=0.0, original_upper_bound=1.0), - BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2000.0), + BoundToUnbound(name_mapping = [["d_L"], ["d_L_unbounded"]] , original_lower_bound=10.0, original_upper_bound=2000.0), BoundToUnbound(name_mapping = [["t_c"], ["t_c_unbounded"]] , original_lower_bound=-0.05, original_upper_bound=0.05), BoundToUnbound(name_mapping = [["phase_c"], ["phase_c_unbounded"]] , original_lower_bound=0.0, original_upper_bound=2 * jnp.pi), BoundToUnbound(name_mapping = [["psi"], ["psi_unbounded"]], original_lower_bound=0.0, original_upper_bound=jnp.pi), @@ -89,13 +91,13 @@ ] likelihood_transforms = [ - SpinToCartesianSpinTransform(name_mapping=[["theta_jn", "phi_jl", "theta_1", "theta_2", "phi_12", "a_1", "a_2"], ["iota", "s1_x", "s1_y", "s1_z", "s2_x", "s2_y", "s2_z"]], freq_ref=20.0), + SpinToCartesianSpinTransform(name_mapping=[["theta_jn", "phi_jl", "theta_1", "theta_2", "phi_12", "a_1", "a_2"], ["iota", "s1_x", "s1_y", "s1_z", "s2_x", "s2_y", "s2_z"]], freq_ref=f_ref), MassRatioToSymmetricMassRatioTransform(name_mapping=[["q"], ["eta"]]), ] likelihood = TransientLikelihoodFD( ifos, - waveform=RippleIMRPhenomD(), + waveform=RippleIMRPhenomPv2(f_ref=f_ref), trigger_time=gps, duration=4, post_trigger_duration=2, diff --git a/test/unit/test_prior.py b/test/unit/test_prior.py index 852ded16..5fbcf3c3 100644 --- a/test/unit/test_prior.py +++ b/test/unit/test_prior.py @@ -43,11 +43,8 @@ def test_sine(self): log_prob = jax.vmap(p.log_prob)(samples) assert jnp.all(jnp.isfinite(log_prob)) # Check that the log_prob is correct in the support - x = trace_prior_parent(p, [])[0].add_name(jnp.linspace(-10.0, 10.0, 1000)[None]) - y = jax.vmap(p.base_prior.base_prior.transform)(x) - y = jax.vmap(p.base_prior.transform)(y) - y = jax.vmap(p.transform)(y) - assert jnp.allclose(jax.vmap(p.log_prob)(y), jnp.log(jnp.sin(y['x'])/2.0)) + samples = samples['x'] + assert jnp.allclose(log_prob, jnp.log(jnp.sin(samples)/2.0)) def test_cosine(self): p = CosinePrior(["x"]) @@ -57,11 +54,8 @@ def test_cosine(self): # Check that the log_prob is finite log_prob = jax.vmap(p.log_prob)(samples) assert jnp.all(jnp.isfinite(log_prob)) - # Check that the log_prob is correct in the support - x = trace_prior_parent(p, [])[0].add_name(jnp.linspace(-10.0, 10.0, 1000)[None]) - y = jax.vmap(p.base_prior.transform)(x) - y = jax.vmap(p.transform)(y) - assert jnp.allclose(jax.vmap(p.log_prob)(y), jnp.log(jnp.cos(y['x'])/2.0)) + samples = samples['x'] + assert jnp.allclose(log_prob, jnp.log(jnp.cos(samples)/2.0)) def test_uniform_sphere(self): p = UniformSpherePrior(["x"])