license: mit
Quakeflow_NC
Introduction
This dataset is part of the data from NCEDC (Northern California Earthquake Data Center) and is organized as several HDF5 files. The dataset structure is shown below: (File ncedc_event_dataset_000.h5.txt shows the structure of the firsr shard of the dataset, and you can find more information about the format at AI4EPS)
Cite the NCEDC: "NCEDC (2014), Northern California Earthquake Data Center. UC Berkeley Seismological Laboratory. Dataset. doi:10.7932/NCEDC."
Acknowledge the NCEDC: "Waveform data, metadata, or data products for this study were accessed through the Northern California Earthquake Data Center (NCEDC), doi:10.7932/NCEDC."
Group: / len:10000
|- Group: /nc100012 len:5
| |-* begin_time = 1987-05-08T00:15:48.890
| |-* depth_km = 7.04
| |-* end_time = 1987-05-08T00:17:48.890
| |-* event_id = nc100012
| |-* event_time = 1987-05-08T00:16:14.700
| |-* event_time_index = 2581
| |-* latitude = 37.5423
| |-* longitude = -118.4412
| |-* magnitude = 1.1
| |-* magnitude_type = D
| |-* num_stations = 5
| |- Dataset: /nc100012/NC.MRS..EH (shape:(3, 12000))
| | |- (dtype=float32)
| | | |-* azimuth = 265.0
| | | |-* component = ['Z']
| | | |-* distance_km = 39.1
| | | |-* dt_s = 0.01
| | | |-* elevation_m = 3680.0
| | | |-* emergence_angle = 93.0
| | | |-* event_id = ['nc100012' 'nc100012']
| | | |-* latitude = 37.5107
| | | |-* location =
| | | |-* longitude = -118.8822
| | | |-* network = NC
| | | |-* phase_index = [3274 3802]
| | | |-* phase_polarity = ['U' 'N']
| | | |-* phase_remark = ['IP' 'S']
| | | |-* phase_score = [1 1]
| | | |-* phase_time = ['1987-05-08T00:16:21.630' '1987-05-08T00:16:26.920']
| | | |-* phase_type = ['P' 'S']
| | | |-* snr = [0. 0. 1.98844361]
| | | |-* station = MRS
| | | |-* unit = 1e-6m/s
| |- Dataset: /nc100012/NN.BEN.N1.EH (shape:(3, 12000))
| | |- (dtype=float32)
| | | |-* azimuth = 329.0
| | | |-* component = ['Z']
| | | |-* distance_km = 22.5
| | | |-* dt_s = 0.01
| | | |-* elevation_m = 2476.0
| | | |-* emergence_angle = 102.0
| | | |-* event_id = ['nc100012' 'nc100012']
| | | |-* latitude = 37.7154
| | | |-* location = N1
| | | |-* longitude = -118.5741
| | | |-* network = NN
| | | |-* phase_index = [3010 3330]
| | | |-* phase_polarity = ['U' 'N']
| | | |-* phase_remark = ['IP' 'S']
| | | |-* phase_score = [0 0]
| | | |-* phase_time = ['1987-05-08T00:16:18.990' '1987-05-08T00:16:22.190']
| | | |-* phase_type = ['P' 'S']
| | | |-* snr = [0. 0. 7.31356192]
| | | |-* station = BEN
| | | |-* unit = 1e-6m/s
......
How to use
Requirements
- datasets
- h5py
- torch (for PyTorch)
Usage
Import the necessary packages:
import h5py
import numpy as np
import torch
from torch.utils.data import Dataset, IterableDataset, DataLoader
from datasets import load_dataset
We have 2 configurations for the dataset: NCEDC and NCEDC_full_size. They all return event-based samples one by one. But NCEDC returns samples with 10 stations each, while NCEDC_full_size return samples with stations same as the original data.
The sample of NCEDC is a dictionary with the following keys:
waveform: the waveform with shape(3, nt, n_sta), the first dimension is 3 components, the second dimension is the number of time samples, the third dimension is the number of stationsphase_pick: the probability of the phase pick with shape(3, nt, n_sta), the first dimension is noise, P and Sevent_location: the event location with shape(4,), including latitude, longitude, depth and timestation_location: the station location with shape(n_sta, 3), the first dimension is latitude, longitude and depth
Because Huggingface datasets only support dynamic size on first dimension, so the sample of NCEDC_full_size is a dictionary with the following keys:
waveform: the waveform with shape(n_sta, 3, nt),phase_pick: the probability of the phase pick with shape(n_sta, 3, nt)event_location: the event location with shape(4,)station_location: the station location with shape(n_sta, 3), the first dimension is latitude, longitude and depth
The default configuration is NCEDC. You can specify the configuration by argument name. For example:
# load dataset
# ATTENTION: Streaming(Iterable Dataset) is difficult to support because of the feature of HDF5
# So we recommend to directly load the dataset and convert it into iterable later
# The dataset is very large, so you need to wait for some time at the first time
# to load "NCEDC"
quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", split="train")
# or
quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", name="NCEDC", split="train")
# to load "NCEDC_full_size"
quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", name="NCEDC_full_size", split="train")
If you want to use the first several shards of the dataset, you can download the script quakeflow_nc.py and change the code as below:
# change the 37 to the number of shards you want
_URLS = {
"NCEDC": [f"{_REPO}/ncedc_event_dataset_{i:03d}.h5" for i in range(37)]
}
Then you can use the dataset like this (Don't forget to specify the argument name):
# don't forget to specify the script path
quakeflow_nc = datasets.load_dataset("path_to_script/quakeflow_nc.py", split="train")
quakeflow_nc
Usage for NCEDC
Then you can change the dataset into PyTorch format iterable dataset, and view the first sample:
quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", name="NCEDC", split="train")
quakeflow_nc = quakeflow_nc.to_iterable_dataset()
# because add examples formatting to get tensors when using the "torch" format
# has not been implemented yet, we need to manually add the formatting
quakeflow_nc = quakeflow_nc.map(lambda x: {key: torch.from_numpy(np.array(value, dtype=np.float32)) for key, value in x.items()})
try:
isinstance(quakeflow_nc, torch.utils.data.IterableDataset)
except:
raise Exception("quakeflow_nc is not an IterableDataset")
# print the first sample of the iterable dataset
for example in quakeflow_nc:
print("\nIterable test\n")
print(example.keys())
for key in example.keys():
print(key, example[key].shape, example[key].dtype)
break
dataloader = DataLoader(quakeflow_nc, batch_size=4)
for batch in dataloader:
print("\nDataloader test\n")
print(batch.keys())
for key in batch.keys():
print(key, batch[key].shape, batch[key].dtype)
break
Usage for NCEDC_full_size
Then you can change the dataset into PyTorch format dataset, and view the first sample (Don't forget to reorder the keys):
quakeflow_nc = datasets.load_dataset("AI4EPS/quakeflow_nc", split="train", name="NCEDC_full_size")
# for PyTorch DataLoader, we need to divide the dataset into several shards
num_workers=4
quakeflow_nc = quakeflow_nc.to_iterable_dataset(num_shards=num_workers)
# because add examples formatting to get tensors when using the "torch" format
# has not been implemented yet, we need to manually add the formatting
quakeflow_nc = quakeflow_nc.map(lambda x: {key: torch.from_numpy(np.array(value, dtype=np.float32)) for key, value in x.items()})
def reorder_keys(example):
example["waveform"] = example["waveform"].permute(1,2,0).contiguous()
example["phase_pick"] = example["phase_pick"].permute(1,2,0).contiguous()
return example
quakeflow_nc = quakeflow_nc.map(reorder_keys)
try:
isinstance(quakeflow_nc, torch.utils.data.IterableDataset)
except:
raise Exception("quakeflow_nc is not an IterableDataset")
data_loader = DataLoader(
quakeflow_nc,
batch_size=1,
num_workers=num_workers,
)
for batch in quakeflow_nc:
print("\nIterable test\n")
print(batch.keys())
for key in batch.keys():
print(key, batch[key].shape, batch[key].dtype)
break
for batch in data_loader:
print("\nDataloader test\n")
print(batch.keys())
for key in batch.keys():
batch[key] = batch[key].squeeze(0)
print(key, batch[key].shape, batch[key].dtype)
break