Neville model
Creating a Model
A CVAE model can be created by calling the function below. The example below creates a model that predicts 4 parameters. this is the standard setup to predice the four doppler parameters of a CW source, also this takes in 362 time steps at 1 day long steps.
neville_model = soapcw.neville.CVAE(
input_dim=362, # input length of timeseries
par_dim=4, # number of parameters to estimate (4 doppler paramaeters)
latent_dim=16, # size of the latent dimension
fc_layers = [128,128,128], # fully connected MLP layers
conv_layers = [(4,4,2,1), (4,4,1,1)], # convolutional layers (num filters, filter size, maxpool size, stride)
device="cpu", # which device to put the model on
fdim=0, # number of extra dimensions to add after convolutional layers (default is one reffering to base frequency of band)
inchannels = 1, # number of channels input
dist_type = "gaussian" # type of distribution to use in the latent space (default gaussian)
)
Loading a Model
In this demonstration I will show how to load a pretrained model into Neville using a config file and some pretrained weights.
neville_model, neville_config = soapcw.neville.load_models.load_model_from_config(
config_file, # path to configuration file - example below
neville_file, # path to the model weights
device=device)
[general]
save_dir = /path/to/save/directory
sft_dirs = /path/to/H1/sfts, /path/to/L1/sfts
narrowband_sft_dir = /path/to/narrowbanded/sfts/
[condor]
accounting_group = accounting.group.
data_load_size = 4.1 # size of frequency band to load for a single job
root_dir = /path/to/condor/files # root directory to save the condor files
[data]
band_starts = [20] # list of start frequencies for band
band_ends = [500] # list of end frequencies for band
band_widths = [0.1] # list of widths of bands
strides = [1] # list of strides to use for each band
resize_image = false # resize the spectrogram/vitmap image
run = o3 # observing run name
type = train # type of data to save
snrmin= 40 # minimum SNR for injections
snrmax= 200 # maximum SNR for injections
n_summed_sfts = 48 # number of SFTs to sum
save_options=["pars","paths"] # only need the paths and parameters for training neville
nperband = 10 # number of times to repeat injections in a band
gen_noise_only = True # generate bands without any injections in
tstart = 1238166018 # start time of search
tend = 1269363618 # end time of search
[lookuptable]
type = power # type of lookup statistic to use power or amplitude
lookup_dir = /path/to/lookup/tables # path to load lookup table from
snr_width_line = 4 # width of prior of SNR for line model
snr_width_signal = 10 # width of prior of SNR for signal model
prob_line = 0.4 # ratio of model probabilities for line and noise
[model]
model_type = "vitmapspectrogram" # which data types to use
save_dir = /directory/to/save/model/to
learning_rate = 1e-4 # learning rate for cnn training
img_dim = (180, 362) # size of input img
conv_layers = [(32, 8, 2, 1),] # convolutional layersin model (num filter, filter size, max pool size, stride)
avg_pool_size = [10,2] # size of addaptive pooling layer
fc_layers = [64,32,2] # fully connected mlp layers,
n_epochs=100 # number of epochs to train cnn for
n_channels = 3 # number of channels as input to cnn
n_train_multi_size=30 # number of repeated training steps on subset of data
save_interval=2 # interval of epochs to save cnn model
band_types = even, odd # which band types to run one (odd or even)
[neville]
save_dir = /path/to/save/neville/model
learning_rate = 1e-4 # learning rate for neville training
input_dim = 362 # input dimension for cvae model
num_predict_params = 4 # number of parameters to predict (4 for doppler parameters)
latent_dim = 16 # size of cvae latent dimension
conv_layers = [(4, 4, 1, 4), (4, 4, 1, 4)] # convolutional layers in model (num filters, filter size, max pool size, stride)
fc_layers = [128, 128, 128, 128] # fully connected mlp layers
fdim=1 # dimensions to add after convolutional layer, default it 1 to add fmin of band
n_channels=1 # number of channels input to cvae
dist_type = "gaussian" # distribution to use in latent dimension of cvae
[code]
search_exec=soapcw-cnn-make-data # exec to create data
Using the Model
For testing the model has a test function that will generate samples from the posterior distribution
neville_model.eval() # put the model in eval mode
with torch.no_grad(): # dont compute gradients for this
samps = neville_model.test(
tracks_input, # torch tensor of input tracks (ntracks, nchannels, nsamples)
freqs=fmins, # torch tensor of base frequency for sub-bands corresponding to tracks
num_samples=5000, # number of posterior samples to generate
transform_func=None, # function to transform parameters, see below
return_latent = False, # whether to return latent space samples also
par=None # injected parameters so latent space from q dsitribution can be returned
)
- The output of this function are:
samps[0] = Normalised doppler parameters posterior samples
samps[1] = Transformed doppler parameters posterior samples (only if a transform function has beed specified)
samps[2] = Track element posterior samples (i.e. samples from binomial distribution corresponding to probability that signal is consistent with that track element)
- In return latent argument is used the outputs are:
samps[0] = Normalised doppler parameters posterior samples
samps[1] = Transformed doppler parameters posterior samples (only if a transform function has beed specified)
samps[2] = samples from the latent space of the r encoder
samps[3] = samples from the latent space of the q encoder
samps[4] = Track element posterior samples (i.e. samples from binomial distribution corresponding to probability that signal is consistent with that track element)
If using the transform function it should take in the samples and an index and output the transformed samples
def transform(samps, i): lon = lon*2*np.pi lat = lat*np.pi/2 f = f + 100.0 fdot = fdot*1e-9 return lon, lat, f, fdot