example.yaml

# Example Cosmopower v2 packaging (Jense et al, 2024).
#   Designed by Hidde Jense and Ian Harrison, for example usage.
#   Based on the original Cosmopower emulators trained for Spurio Mancini et al, 2021.
# As of 12/06/2023.
# 
# A simple alphanumeric name for this network.
# The networks are normally sotred under "path/filename", where
# filename is either a specified file name, or (default) the
# network name + the computed quantity (e.g. "cosmopower/trained_models/CP_paper_Cl_tt.pkl"
# for a ClTT network).
network_name: CP_paper
# Where the network is (locally) stored.
path: cosmopower/trained_models

# The original networks were stored as pickle (.pkl) files
allow_pickle: True

emulated_code:
  # Details about the code being emulated (in this case, camb)
  name: camb
  # The version of camb used to generate the model with.
  version: "1.4.0"
  extra_args:
    # Any extra arguments passed on to the emulated code (e.g. CAMB accuracy parameters).
    # (the specifics can be handled by the Boltzmann code handler).
    lens_potential_accuracy: 1
    kmax: 100.0

samples:
  # The number of samples for training.
  Ntraining: 100000

  # The training parameters used for the networks.
  parameters:
    ombh2:
      range: [0.001,0.04]
    omch2:
      range: [0.002,0.50]
    logA:
      # We want to use logA as an input to our networks,
      # but use As as an input to camb. The "drop: True" flag
      # marks this parameter as "do not give to boltzmann code"
      range: [2.0,5.0]
      drop: True
    As:
      # We want to use logA as an input to our networks,
      # but use As as an input to camb. The "derived: True" flag
      # marks this parameter as "compute this value from another parameter"
      derived: True
      value: "lambda logA: 1.e-10 * np.exp(logA)"
    tau:
      range: [0.01, 0.10]
    ns:
      range: [0.8, 1.2]
    h:
      range: [0.5,1.0]
      drop: True
    H0:
      derived: True
      value: "lambda h: h * 100.0"
    # For P(k,z), an extra input parameter z was used, e.g.
    # z:
    #   range: [0.0, 5.0]
    
    # If you want, parameter can be spaced in logspace by adding
    # either a "derived" flag or adding a "spacing: log" tag.

# Information on the trained networks
networks:
  # a ClTT Neural Network emulator trained on log(Cl) on 2 <= l <= 2508.
  # Note: we have not fully decided on the correct naming scheme for the "quantity" value.
  #       Should we change this in the future, this would be a trivial thing to change, however.
  - quantity: "Cl/tt" # ClTT
    filename: "CMB/cmb_TT_NN"
    # List which parameters are used as inputs for this network.
    inputs: [ ombh2, omch2, logA, tau, ns, h ]
    # Recognized types: NN or PCAplusNN (if more get created in the future, those could be added too).
    type: NN
    # Train this network on log(Cl) rather than raw Cl.
    log: True
    ell_factor: false # i.e. train this on log(Cl), not on log(l (l+1) Cl / 2 pi)
    modes:
      # The label is purely cosmetic for now.
      label: l
      range: [2,2508]
      # you can specify the number of steps by adding a "steps: N"
      # tag.
      # similar to parameter ranges, modes can be put in log space
      # by adding a "spacing: log" tag (this requires specifying the number
      # of steps as well).
      # if the number of steps is not specified, it assumes linear spacing with unit
      # spacing (i.e. doing "l = np.arange(lmin, lmax+1)".
    n_traits:
      # Cosmopower NN traits.
      n_hidden: [ 512, 512, 512, 512 ]
    training:
      # cosmopower training parameters.
      # All these parameters are passed on as **kwargs to the network.train() method.
      validation_split: 0.1
      learning_rates: [ 1.e-2, 1.e-3, 1.e-4, 1.e-5, 1.e-6 ]
      batch_sizes: [ 1024, 1024, 1024, 1024, 1024 ]
      gradient_accumulation_steps: [ 1.0, 1.0, 1.0, 1.0, 1.0 ]
      patience_values: [ 100, 100, 100, 100, 100 ]
      max_epochs: [ 1000, 1000, 1000, 1000, 1000 ]

  - quantity: "Cl/te" # ClTE
    filename: "CMB/cmb_TE_PCAplusNN"
    inputs: [ ombh2, omch2, logA, tau, ns, h ]
    type: PCAplusNN
    log: False # the zero-crossing of TE makes it impossible to train on log(Cl).
    ell_factor: false
    modes:
      label: l
      range: [2,2508]
    p_traits:
      # Cosmpower PCA traits.
      n_pcas: 512
    n_traits:
      n_hidden: [ 512, 512, 512, 512 ]
    training:
      validation_split: 0.1
      learning_rates: [ 1.e-2, 1.e-3, 1.e-4, 1.e-5, 1.e-6 ]
      batch_sizes: [ 1024, 1024, 1024, 1024, 1024 ]
      gradient_accumulation_steps: [ 1.0, 1.0, 1.0, 1.0, 1.0 ]
      patience_values: [ 100, 100, 100, 100, 100 ]
      max_epochs: [ 1000, 1000, 1000, 1000, 1000 ]

  - quantity: "Cl/ee" # (ClEE is similar to ClTT).
    filename: "CMB/cmb_EE_NN"
    inputs: [ ombh2, omch2, logA, tau, ns, h ]
    type: NN
    log: True
    ell_factor: false
    modes:
      label: l
      range: [2,2508]
    n_traits:
      n_hidden: [ 512, 512, 512, 512 ]
    training:
      validation_split: 0.2
      learning_rates: [ 1.e-2, 1.e-3, 1.e-4, 1.e-5, 1.e-6 ]
      batch_sizes: [ 1024, 1024, 1024, 1024, 1024 ]
      gradient_accumulation_steps: [ 1.0, 1.0, 1.0, 1.0, 1.0 ]
      patience_values: [ 100, 100, 100, 100, 100 ]
      max_epochs: [ 1000, 1000, 1000, 1000, 1000 ]

  - quantity: "Cl/pp" # Clpp (phiphi), lensing spectrum.
    filename: "CMB/cmb_PP_PCAplusNN"
    # note we do not use tau as an input for Clpp
    inputs: [ ombh2, omch2, logA, ns, h ]
    type: PCAplusNN
    log: True
    ell_factor: false
    modes:
      label: l
      range: [2,2508]
    p_traits:
      n_pcas: 64
    n_traits:
      n_hidden: [ 512, 512, 512, 512 ]
    training:
      validation_split: 0.2
      learning_rates: [ 1.e-2, 1.e-3, 1.e-4, 1.e-5, 1.e-6 ]
      batch_sizes: [ 1024, 1024, 1024, 1024, 1024 ]
      gradient_accumulation_steps: [ 1.0, 1.0, 1.0, 1.0, 1.0 ]
      patience_values: [ 100, 100, 100, 100, 100 ]
      max_epochs: [ 1000, 1000, 1000, 1000, 1000 ]
  
  # Other quantities that might be of interest:
  # Cl/bb
  # Cl_unlensed/xy (xy = tt, te, ee, bb, pp) for unlensed spectra
  # Pk/lin (linear matter power spectrum)
  # Pk/nonlin (non-linear matter power spectrum) or Pk/boost (= Pk_nonlin / Pk_lin)
  # Hubble (H(z) emulator)
  # D_A (angular diameter distance D_A(z) emulator)