easy_tpp.model.torch_model.torch_thinning
Classes
|
Event Sequence Sampler based on thinning algorithm, which corresponds to Algorithm 2 of The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process, https://arxiv.org/abs/1612.09328. |
- class easy_tpp.model.torch_model.torch_thinning.EventSampler(num_sample, num_exp, over_sample_rate, num_samples_boundary, dtime_max, patience_counter, device)[source]
Event Sequence Sampler based on thinning algorithm, which corresponds to Algorithm 2 of The Neural Hawkes Process: A Neurally Self-Modulating Multivariate Point Process, https://arxiv.org/abs/1612.09328.
The implementation uses code from https://github.com/yangalan123/anhp-andtt/blob/master/anhp/esm/thinning.py.
- __init__(num_sample, num_exp, over_sample_rate, num_samples_boundary, dtime_max, patience_counter, device)[source]
Initialize the event sampler.
- Parameters:
num_sample (int) – number of sampled next event times via thinning algo for computing predictions.
num_exp (int) – number of i.i.d. Exp(intensity_bound) draws at one time in thinning algorithm
over_sample_rate (float) – multiplier for the intensity up bound.
num_samples_boundary (int) – number of sampled event times to compute the boundary of the intensity.
dtime_max (float) – max value of delta times in sampling
patience_counter (int) – the maximum iteration used in adaptive thinning.
device (torch.device) – torch device index to select.
- compute_intensity_upper_bound(time_seq, time_delta_seq, event_seq, intensity_fn, compute_last_step_only)[source]
Compute the upper bound of intensity at each event timestamp.
- Parameters:
time_seq (tensor) – [batch_size, seq_len], timestamp seqs.
time_delta_seq (tensor) – [batch_size, seq_len], time delta seqs.
event_seq (tensor) – [batch_size, seq_len], event type seqs.
intensity_fn (fn) – a function that computes the intensity.
compute_last_step_only (bool) – wheter to compute the last time step pnly.
- Returns:
[batch_size, seq_len]
- Return type:
tensor
- sample_exp_distribution(sample_rate)[source]
Sample an exponential distribution.
- Parameters:
sample_rate (tensor) – [batch_size, seq_len], intensity rate.
- Returns:
[batch_size, seq_len, num_exp], exp numbers at each event timestamp.
- Return type:
tensor
- sample_uniform_distribution(intensity_upper_bound)[source]
Sample an uniform distribution
- Parameters:
intensity_upper_bound (tensor) – upper bound intensity computed in the previous step.
- Returns:
[batch_size, seq_len, num_sample, num_exp]
- Return type:
tensor
- sample_accept(unif_numbers, sample_rate, total_intensities)[source]
Do the sample-accept process.
For each parallel draw, find its min criterion： if that < 1.0, the 1st (i.e. smallest) sampled time with cri < 1.0 is accepted; if none is accepted, use boundary / maxsampletime for that draw
- Parameters:
unif_numbers (tensor) – [batch_size, max_len, num_sample, num_exp], sampled uniform random number.
sample_rate (tensor) – [batch_size, max_len], sample rate (intensity).
total_intensities (tensor) – [batch_size, seq_len, num_sample, num_exp]
- Returns:
two tensors, criterion, [batch_size, max_len, num_sample, num_exp] who_has_accepted_times, [batch_size, max_len, num_sample]
- Return type:
list
- draw_next_time_one_step(time_seq, time_delta_seq, event_seq, dtime_boundary, intensity_fn, compute_last_step_only=False)[source]
Compute next event time based on Thinning algorithm.
- Parameters:
time_seq (tensor) – [batch_size, seq_len], timestamp seqs.
time_delta_seq (tensor) – [batch_size, seq_len], time delta seqs.
event_seq (tensor) – [batch_size, seq_len], event type seqs.
dtime_boundary (tensor) – [batch_size, seq_len], dtime upper bound.
intensity_fn (fn) – a function to compute the intensity.
compute_last_step_only (bool, optional) – whether to compute last event timestep only. Defaults to False.
- Returns:
next event time prediction and weight.
- Return type:
tuple