device |
str |
best |
No |
Device to use for model inference. best will use the best available device. |
memory_strategy |
str | None |
high |
No |
Memory strategy to use for model inference. high will use more memory, low will use less. Utilizing more memory will improve runtime performance. |
tqdm_enabled |
bool |
True |
No |
Whether to enable tqdm progress bars. |
n_workers_for_norm_param_estimation |
int |
8 |
No |
Number of workers for norm parameter estimation from the baseline. Utilizing more workers will improve runtime performance and utilize more memory. Does not work with model compilation or MPS/GPU devices. |
batch_size_for_norm_param_estimation |
int |
32 |
No |
Batch size for norm parameter estimation from the baseline. Utilizing a larger batch size will improve runtime performance and utilize more memory. |
stride_for_norm_param_estimation |
int |
16 |
No |
Stride for norm parameter estimation from the baseline. Utilizing a larger stride will improve metric accuracy and utilize more memory.Memory usage scales inverse quadratically with stride. That is, If stride=16 consumes N bytes of memory, then stride=4 consumes 16N bytes of memory. |
apply_logit_to_inputs |
bool |
True |
No |
Whether to apply logit transform to the input data. |
n_workers_for_despeckling |
int |
8 |
No |
Number of workers for despeckling. Utilizing more workers will improve runtime performance and utilize more memory. |
lookback_strategy |
str |
multi_window |
No |
Lookback strategy to use for data curation of the baseline. multi_window will use a multi-window lookback strategy and is default for OEPRA DIST-S1, immediate_lookback will use an immediate lookback strategy using acquisitions preceding the post-date. immediate_lookback is not supported yet. |
post_date_buffer_days |
int |
1 |
No |
Buffer days around post-date for data collection to create acqusition image to compare baseline to. |
model_compilation |
bool |
False |
No |
Whether to compile the model for CPU or GPU. False, use the model as is. True, load the model and compile for CPU or GPU optimizations. |
max_pre_imgs_per_burst_mw |
tuple[int, ...] | None |
None |
Yes |
Max number of pre-images per burst within each window. If None, the value will be calculated based on the model context length and the number of anniversaries. Specifically, the value will be context_length // n_anniversaries with remainder added to the first window. |
delta_lookback_days_mw |
tuple[int, ...] | None |
None |
Yes |
Delta lookback days for each window relative to post-image acquisition date. If None, the value will be calculated based on the number of anniversaries (default is 3). |
low_confidence_alert_threshold |
float |
3.5 |
No |
Low confidence alert threshold for detecting disturbance between baseline and post-image. |
high_confidence_alert_threshold |
float |
5.5 |
No |
High confidence alert threshold for detecting disturbance between baseline and post-image. |
no_day_limit |
int |
30 |
No |
Number of days to limit confirmation process logic to. Confirmation must occur within first observance of disturbance and no_day_limit days after first disturbance. |
exclude_consecutive_no_dist |
int |
True |
No |
Boolean activation of consecutive no disturbance tracking during confirmation. True will apply this logic: after 2 no disturbances within product sequence, the disturbance must finish or be reset. False will not apply this logic. |
percent_reset_thresh |
int |
10 |
No |
Precentage number threshold to reset disturbance. Values below percent_reset_thresh will reset disturbance. |
no_count_reset_thresh |
int |
7 |
No |
If the number of non-disturbed observations prevnocount is above nocount_reset_thresh disturbance will reset. |
max_obs_num_year |
int |
253 |
No |
Max observation number per year. If observations exceeds this number, then the confirmation must conclude and be reset. |
confidence_upper_lim |
int |
32000 |
No |
Confidence upper limit for confirmation. Confidence is an accumulation of the metric over time. |
confirmation_confidence_threshold |
float |
31.5 |
No |
This is the threshold for the confirmation process to determine if a disturbance is confirmed. |
metric_value_upper_lim |
float |
100.0 |
No |
Metric upper limit set during confirmation |
model_source |
str |
transformer_optimized |
No |
Model source. If external, use externally supplied paths for weights and config. Otherwise, use distmetrics.model_load.ALLOWED_MODELS for other models. |
model_cfg_path |
Path | str | None |
None |
Yes |
Path to model config file. If external, use externally supplied path. Otherwise, use distmetrics.model_load.ALLOWED_MODELS for other models. |
model_wts_path |
Path | str | None |
None |
Yes |
Path to model weights file. If external, use externally supplied path. Otherwise, use distmetrics.model_load.ALLOWED_MODELS for other models. |
apply_despeckling |
bool |
True |
No |
Whether to apply despeckling to the input data. |
interpolation_method |
str |
bilinear |
No |
Interpolation method to use for despeckling. nearest will use nearest neighbor interpolation, bilinear will use bilinear interpolation, and none will not apply despeckling. |
model_dtype |
str |
float32 |
No |
Data type for model inference. Note: bfloat16 is only supported on GPU devices. |
use_date_encoding |
bool |
False |
No |
Whether to use acquisition date encoding in model application (currently not supported) |
n_anniversaries_for_mw |
int |
3 |
No |
Number of anniversaries to use for multi-window |