kylewhy commited on
Commit
fa3c8cd
1 Parent(s): d5333b8

add batched configuration

Browse files
Files changed (2) hide show
  1. README.md +67 -7
  2. quakeflow_nc.py +112 -32
README.md CHANGED
@@ -76,36 +76,49 @@ Group: / len:10000
76
  - torch (for PyTorch)
77
 
78
  ### Usage
 
79
  ```python
80
  import h5py
81
  import numpy as np
82
  import torch
83
  from torch.utils.data import Dataset, IterableDataset, DataLoader
84
  from datasets import load_dataset
85
-
 
 
86
  # load dataset
87
  # ATTENTION: Streaming(Iterable Dataset) is difficult to support because of the feature of HDF5
88
  # So we recommend to directly load the dataset and convert it into iterable later
89
  # The dataset is very large, so you need to wait for some time at the first time
90
- quakeflow_nc = datasets.load_dataset("AI4EPS/quakeflow_nc", split="train")
91
- quakeflow_nc
 
 
 
 
 
 
92
  ```
93
- If you want to use the first several shards of the dataset, you can download the script `quakeflow_nc.py` and change the code below:
 
94
  ```python
95
  # change the 37 to the number of shards you want
96
  _URLS = {
97
  "NCEDC": [f"{_REPO}/ncedc_event_dataset_{i:03d}.h5" for i in range(37)]
98
  }
99
  ```
100
- Then you can use the dataset like this:
101
  ```python
102
- quakeflow_nc = datasets.load_dataset("./quakeflow_nc.py", split="train")
 
103
  quakeflow_nc
104
  ```
 
 
105
  Then you can change the dataset into PyTorch format iterable dataset, and view the first sample:
106
  ```python
 
107
  quakeflow_nc = quakeflow_nc.to_iterable_dataset()
108
- quakeflow_nc = quakeflow_nc.with_format("torch")
109
  # because add examples formatting to get tensors when using the "torch" format
110
  # has not been implemented yet, we need to manually add the formatting
111
  quakeflow_nc = quakeflow_nc.map(lambda x: {key: torch.from_numpy(np.array(value, dtype=np.float32)) for key, value in x.items()})
@@ -130,4 +143,51 @@ for batch in dataloader:
130
  for key in batch.keys():
131
  print(key, batch[key].shape, batch[key].dtype)
132
  break
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
133
  ```
 
76
  - torch (for PyTorch)
77
 
78
  ### Usage
79
+ Import the necessary packages:
80
  ```python
81
  import h5py
82
  import numpy as np
83
  import torch
84
  from torch.utils.data import Dataset, IterableDataset, DataLoader
85
  from datasets import load_dataset
86
+ ```
87
+ We have 2 configurations for the dataset: `NCEDC` and `NCEDC_Batch`. `NCEDC` return event-based samples(with 10 stations each) one by one, while `NCEDC_Batch` return event-based samples in specified batch size and specified stations number(default:`batch_size=16`,`num_stations_list=[5, 10, 20]`). The default configuration is `NCEDC`. You can specify the configuration by argument `name`. For example:
88
+ ```python
89
  # load dataset
90
  # ATTENTION: Streaming(Iterable Dataset) is difficult to support because of the feature of HDF5
91
  # So we recommend to directly load the dataset and convert it into iterable later
92
  # The dataset is very large, so you need to wait for some time at the first time
93
+
94
+ # to load "NCEDC"
95
+ quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", split="train")
96
+ # or
97
+ quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", name="NCEDC", split="train")
98
+
99
+ # to load "NCEDC_Batch"
100
+ quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", name="NCEDC_Batch", split="train")
101
  ```
102
+
103
+ 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:
104
  ```python
105
  # change the 37 to the number of shards you want
106
  _URLS = {
107
  "NCEDC": [f"{_REPO}/ncedc_event_dataset_{i:03d}.h5" for i in range(37)]
108
  }
109
  ```
110
+ Then you can use the dataset like this (Don't forget to specify the argument `name`):
111
  ```python
112
+ # don't forget to specify the script path
113
+ quakeflow_nc = datasets.load_dataset("path_to_script/quakeflow_nc.py", split="train")
114
  quakeflow_nc
115
  ```
116
+
117
+ #### Usage for `NCEDC`
118
  Then you can change the dataset into PyTorch format iterable dataset, and view the first sample:
119
  ```python
120
+ quakeflow_nc = load_dataset("AI4EPS/quakeflow_nc", name="NCEDC", split="train")
121
  quakeflow_nc = quakeflow_nc.to_iterable_dataset()
 
122
  # because add examples formatting to get tensors when using the "torch" format
123
  # has not been implemented yet, we need to manually add the formatting
124
  quakeflow_nc = quakeflow_nc.map(lambda x: {key: torch.from_numpy(np.array(value, dtype=np.float32)) for key, value in x.items()})
 
143
  for key in batch.keys():
144
  print(key, batch[key].shape, batch[key].dtype)
145
  break
146
+ ```
147
+
148
+ #### Usage for `NCEDC_Batch`
149
+ Then you can change the dataset into PyTorch format iterable dataset, and view the first sample:
150
+ ```python
151
+ quakeflow_batch = datasets.load_dataset("AI4EPS/quakeflow_nc", split="train", name="NCEDC_Batch", batch_size=16, num_stations_list=[5,10,20])
152
+
153
+ # for PyTorch DataLoader, we need to divide the dataset into several shards
154
+ num_workers=4
155
+ quakeflow_batch = quakeflow_batch.to_iterable_dataset(num_shards=num_workers)
156
+ # because add examples formatting to get tensors when using the "torch" format
157
+ # has not been implemented yet, we need to manually add the formatting
158
+ quakeflow_batch = quakeflow_batch.map(lambda x: {key: torch.from_numpy(np.array(value, dtype=np.float32)) for key, value in x.items()})
159
+ def reorder_keys(example):
160
+ example["waveform"] = example["waveform"].permute(1,2,3,0).contiguous()
161
+ example["phase_pick"] = example["phase_pick"].permute(1,2,3,0).contiguous()
162
+ example["station_location"] = example["station_location"].permute(1,0,2).contiguous()
163
+ return example
164
+
165
+ quakeflow_batch = quakeflow_batch.map(reorder_keys)
166
+
167
+ try:
168
+ isinstance(quakeflow_batch, torch.utils.data.IterableDataset)
169
+ except:
170
+ raise Exception("quakeflow_nc is not an IterableDataset")
171
+
172
+ data_loader = DataLoader(
173
+ quakeflow_batch,
174
+ sampler=None,
175
+ collate_fn=None,
176
+ num_workers=num_workers,
177
+ )
178
+
179
+ for batch in quakeflow_batch:
180
+ print("\nIterable test\n")
181
+ print(batch.keys())
182
+ for key in batch.keys():
183
+ print(key, batch[key].shape, batch[key].dtype)
184
+ break
185
+
186
+ for batch in data_loader:
187
+ print("\nDataloader test\n")
188
+ print(batch.keys())
189
+ for key in batch.keys():
190
+ batch[key] = batch[key].squeeze(0)
191
+ print(key, batch[key].shape, batch[key].dtype)
192
+ break
193
  ```
quakeflow_nc.py CHANGED
@@ -26,6 +26,7 @@ import torch
26
  import fsspec
27
  from glob import glob
28
  from typing import Dict, List, Optional, Tuple, Union
 
29
 
30
  import datasets
31
 
@@ -58,14 +59,35 @@ _LICENSE = ""
58
  # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
59
  _REPO = "https://huggingface.co/datasets/AI4EPS/quakeflow_nc/resolve/main/data"
60
  _URLS = {
61
- "NCEDC": [f"{_REPO}/ncedc_event_dataset_{i:03d}.h5" for i in range(37)]
 
62
  }
63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
64
 
65
  # TODO: Name of the dataset usually matches the script name with CamelCase instead of snake_case
66
  class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
67
  """QuakeFlow_NC: A dataset of earthquake waveforms organized by earthquake events and based on the HDF5 format."""
68
 
 
 
 
 
 
 
 
69
  VERSION = datasets.Version("1.1.0")
70
 
71
  # This is an example of a dataset with multiple configurations.
@@ -79,22 +101,36 @@ class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
79
  # You will be able to load one or the other configurations in the following list with
80
  # data = datasets.load_dataset('my_dataset', 'first_domain')
81
  # data = datasets.load_dataset('my_dataset', 'second_domain')
 
 
82
  BUILDER_CONFIGS = [
83
- datasets.BuilderConfig(name="NCEDC", version=VERSION, description="This part of my dataset covers a first domain"),
 
84
  ]
85
 
86
- DEFAULT_CONFIG_NAME = "NCEDC" # It's not mandatory to have a default configuration. Just use one if it make sense.
87
 
88
  def _info(self):
89
  # TODO: This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
90
- features=datasets.Features(
91
- {
92
- "waveform": datasets.Array3D(shape=(3, self.nt, self.num_stations), dtype='float32'),
93
- "phase_pick": datasets.Array3D(shape=(3, self.nt, self.num_stations), dtype='float32'),
94
- "event_location": [datasets.Value("float32")],
95
- "station_location": datasets.Array2D(shape=(self.num_stations, 3), dtype="float32"),
96
- }
97
- )
 
 
 
 
 
 
 
 
 
 
 
98
  return datasets.DatasetInfo(
99
  # This is the description that will appear on the datasets page.
100
  description=_DESCRIPTION,
@@ -130,6 +166,8 @@ class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
130
  gen_kwargs={
131
  "filepath": files,
132
  "split": "train",
 
 
133
  },
134
  ),
135
  # datasets.SplitGenerator(
@@ -150,18 +188,20 @@ class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
150
  # ),
151
  ]
152
 
153
- degree2km = 111.32
154
- nt = 8192
155
- feature_nt = 512
156
- feature_scale = int(nt / feature_nt)
157
- sampling_rate=100.0
158
- num_stations = 10
159
 
160
  # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
161
- def _generate_examples(self, filepath, split):
162
  # TODO: This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
163
  # The `key` is for legacy reasons (tfds) and is not important in itself, but must be unique for each example.
164
- num_stations = self.num_stations
 
 
 
 
 
 
 
165
 
166
  for file in filepath:
167
  with h5py.File(file, "r") as fp:
@@ -169,11 +209,17 @@ class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
169
  for event_id in fp.keys():
170
  event = fp[event_id]
171
  station_ids = list(event.keys())
172
- if len(station_ids) < num_stations:
 
173
  continue
174
- else:
175
- station_ids = np.random.choice(station_ids, num_stations, replace=False)
176
-
 
 
 
 
 
177
  waveforms = np.zeros([3, self.nt, len(station_ids)])
178
  phase_pick = np.zeros_like(waveforms)
179
  attrs = event.attrs
@@ -182,13 +228,11 @@ class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
182
 
183
  for i, sta_id in enumerate(station_ids):
184
  # trace_id = event_id + "/" + sta_id
185
-
186
  waveforms[:, :, i] = event[sta_id][:,:self.nt]
187
  attrs = event[sta_id].attrs
188
  p_picks = attrs["phase_index"][attrs["phase_type"] == "P"]
189
  s_picks = attrs["phase_index"][attrs["phase_type"] == "S"]
190
  phase_pick[:, :, i] = generate_label([p_picks, s_picks], nt=self.nt)
191
-
192
  station_location.append([attrs["longitude"], attrs["latitude"], -attrs["elevation_m"]/1e3])
193
 
194
  std = np.std(waveforms, axis=1, keepdims=True)
@@ -196,13 +240,49 @@ class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
196
  waveforms = (waveforms - np.mean(waveforms, axis=1, keepdims=True)) / std
197
  waveforms = waveforms.astype(np.float32)
198
 
199
- yield event_id, {
200
- "waveform": torch.from_numpy(waveforms).float(),
201
- "phase_pick": torch.from_numpy(phase_pick).float(),
202
- "event_location": event_location,
203
- "station_location": station_location,
204
- }
205
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
206
 
207
 
208
  def generate_label(phase_list, label_width=[150, 150], nt=8192):
 
26
  import fsspec
27
  from glob import glob
28
  from typing import Dict, List, Optional, Tuple, Union
29
+ from collections import defaultdict
30
 
31
  import datasets
32
 
 
59
  # This can be an arbitrary nested dict/list of URLs (see below in `_split_generators` method)
60
  _REPO = "https://huggingface.co/datasets/AI4EPS/quakeflow_nc/resolve/main/data"
61
  _URLS = {
62
+ "NCEDC": [f"{_REPO}/ncedc_event_dataset_{i:03d}.h5" for i in range(37)],
63
+ "NCEDC_Batch": [f"{_REPO}/ncedc_event_dataset_{i:03d}.h5" for i in range(37)],
64
  }
65
 
66
+ class BatchBuilderConfig(datasets.BuilderConfig):
67
+ """
68
+ yield a batch of event-based sample, so the number of sample stations can vary among batches
69
+ Batch Config for QuakeFlow_NC
70
+ :param batch_size: number of samples in a batch
71
+ :param num_stations_list: possible number of stations in a batch
72
+ """
73
+ def __init__(self, batch_size: int, num_stations_list: List, **kwargs):
74
+ super().__init__(**kwargs)
75
+ self.batch_size = batch_size
76
+ self.num_stations_list = num_stations_list
77
+ #batch_size: int=None
78
+ #num_stations_list: List=None
79
 
80
  # TODO: Name of the dataset usually matches the script name with CamelCase instead of snake_case
81
  class QuakeFlow_NC(datasets.GeneratorBasedBuilder):
82
  """QuakeFlow_NC: A dataset of earthquake waveforms organized by earthquake events and based on the HDF5 format."""
83
 
84
+ degree2km = 111.32
85
+ nt = 8192
86
+ feature_nt = 512
87
+ feature_scale = int(nt / feature_nt)
88
+ sampling_rate=100.0
89
+ num_stations = 10
90
+
91
  VERSION = datasets.Version("1.1.0")
92
 
93
  # This is an example of a dataset with multiple configurations.
 
101
  # You will be able to load one or the other configurations in the following list with
102
  # data = datasets.load_dataset('my_dataset', 'first_domain')
103
  # data = datasets.load_dataset('my_dataset', 'second_domain')
104
+
105
+ # default config, you can change batch_size and num_stations_list when use `datasets.load_dataset`
106
  BUILDER_CONFIGS = [
107
+ BatchBuilderConfig(name="NCEDC", version=VERSION, description="yield event-based samples one by one, the", batch_size=1, num_stations_list=[num_stations]),
108
+ BatchBuilderConfig(name="NCEDC_Batch", version=VERSION, description="This part of my dataset covers a first domain", batch_size=16, num_stations_list=[5, 10, 20]),
109
  ]
110
 
111
+ DEFAULT_CONFIG_NAME = "NCEDC_Batch" # It's not mandatory to have a default configuration. Just use one if it make sense.
112
 
113
  def _info(self):
114
  # TODO: This method specifies the datasets.DatasetInfo object which contains informations and typings for the dataset
115
+ if self.config.name=="NCEDC":
116
+ features=datasets.Features(
117
+ {
118
+ "waveform": datasets.Array3D(shape=(3, self.nt, self.num_stations), dtype='float32'),
119
+ "phase_pick": datasets.Array3D(shape=(3, self.nt, self.num_stations), dtype='float32'),
120
+ "event_location": datasets.Sequence(datasets.Value("float32")),
121
+ "station_location": datasets.Array2D(shape=(self.num_stations, 3), dtype="float32"),
122
+ })
123
+
124
+ elif self.config.name=="NCEDC_Batch":
125
+ features=datasets.Features(
126
+ {
127
+ "waveform": datasets.Array4D(shape=(None, self.config.batch_size, 3, self.nt), dtype='float32'),# datasets.Array4D(shape=(self.config.batch_size, 3, self.nt, self.num_stations), dtype='float32'),
128
+ "phase_pick": datasets.Array4D(shape=(None, self.config.batch_size, 3, self.nt), dtype='float32'),
129
+ "event_location": datasets.Array2D(shape=(self.config.batch_size, 4), dtype="float32"),
130
+ "station_location": datasets.Array3D(shape=(None, self.config.batch_size, 3), dtype="float32"),
131
+ }
132
+ )
133
+
134
  return datasets.DatasetInfo(
135
  # This is the description that will appear on the datasets page.
136
  description=_DESCRIPTION,
 
166
  gen_kwargs={
167
  "filepath": files,
168
  "split": "train",
169
+ "batch_size": self.config.batch_size,
170
+ "num_stations_list": self.config.num_stations_list,
171
  },
172
  ),
173
  # datasets.SplitGenerator(
 
188
  # ),
189
  ]
190
 
191
+
 
 
 
 
 
192
 
193
  # method parameters are unpacked from `gen_kwargs` as given in `_split_generators`
194
+ def _generate_examples(self, filepath, split, batch_size, num_stations_list):
195
  # TODO: This method handles input defined in _split_generators to yield (key, example) tuples from the dataset.
196
  # The `key` is for legacy reasons (tfds) and is not important in itself, but must be unique for each example.
197
+ num_batches = 0
198
+
199
+ num_stations_list = np.array(num_stations_list)
200
+ if self.config.name=="NCEDC_Batch":
201
+ waveform_buffer_per_group = defaultdict(list)
202
+ phase_pick_buffer_per_group = defaultdict(list)
203
+ event_location_buffer_per_group = defaultdict(list)
204
+ station_location_buffer_per_group = defaultdict(list)
205
 
206
  for file in filepath:
207
  with h5py.File(file, "r") as fp:
 
209
  for event_id in fp.keys():
210
  event = fp[event_id]
211
  station_ids = list(event.keys())
212
+
213
+ if len(station_ids) < num_stations_list[0]:
214
  continue
215
+
216
+ if batch_size==1 and self.config.name=="NCEDC":
217
+ num_stations = num_stations_list[0]
218
+ elif self.config.name=="NCEDC_Batch":
219
+ num_stations=num_stations_list[num_stations_list<=len(station_ids)][-1]
220
+
221
+ station_ids = np.random.choice(station_ids, num_stations, replace=False)
222
+
223
  waveforms = np.zeros([3, self.nt, len(station_ids)])
224
  phase_pick = np.zeros_like(waveforms)
225
  attrs = event.attrs
 
228
 
229
  for i, sta_id in enumerate(station_ids):
230
  # trace_id = event_id + "/" + sta_id
 
231
  waveforms[:, :, i] = event[sta_id][:,:self.nt]
232
  attrs = event[sta_id].attrs
233
  p_picks = attrs["phase_index"][attrs["phase_type"] == "P"]
234
  s_picks = attrs["phase_index"][attrs["phase_type"] == "S"]
235
  phase_pick[:, :, i] = generate_label([p_picks, s_picks], nt=self.nt)
 
236
  station_location.append([attrs["longitude"], attrs["latitude"], -attrs["elevation_m"]/1e3])
237
 
238
  std = np.std(waveforms, axis=1, keepdims=True)
 
240
  waveforms = (waveforms - np.mean(waveforms, axis=1, keepdims=True)) / std
241
  waveforms = waveforms.astype(np.float32)
242
 
243
+ if batch_size==1 and self.config.name=="NCEDC":
244
+ yield event_id, {
245
+ "waveform": torch.from_numpy(waveforms).float(),
246
+ "phase_pick": torch.from_numpy(phase_pick).float(),
247
+ "event_location": event_location,
248
+ "station_location": station_location,
249
+ }
250
+ elif self.config.name=="NCEDC_Batch":
251
+
252
+ waveform_buffer_per_group[num_stations].append(waveforms)
253
+ phase_pick_buffer_per_group[num_stations].append(phase_pick)
254
+ event_location_buffer_per_group[num_stations].append(event_location)
255
+ station_location_buffer_per_group[num_stations].append(station_location)
256
+
257
+ if len(waveform_buffer_per_group[num_stations])==batch_size:
258
+ yield num_batches, {
259
+ "waveform": torch.from_numpy(np.stack(waveform_buffer_per_group[num_stations], axis=0)).float().permute(3,0,1,2),
260
+ "phase_pick": torch.from_numpy(np.stack(phase_pick_buffer_per_group[num_stations], axis=0)).float().permute(3,0,1,2),
261
+ "event_location": torch.from_numpy(np.stack(event_location_buffer_per_group[num_stations], axis=0)).float(),
262
+ "station_location": torch.from_numpy(np.stack(station_location_buffer_per_group[num_stations], axis=0)).float().permute(1,0,2),
263
+ }
264
+ del waveform_buffer_per_group[num_stations]
265
+ del phase_pick_buffer_per_group[num_stations]
266
+ del event_location_buffer_per_group[num_stations]
267
+ del station_location_buffer_per_group[num_stations]
268
+ num_batches += 1
269
+ assert len(waveform_buffer_per_group[num_stations])<batch_size, "batch size is not correct"
270
+
271
+ '''
272
+ # drop_last=False
273
+ if self.config.name=="NCEDC_Batch":
274
+ for num_stations in waveform_buffer_per_group:
275
+ yield event_id, {
276
+ "waveform": torch.from_numpy(np.stack(waveform_buffer_per_group, axis=0)).float(),
277
+ "phase_pick": torch.from_numpy(np.stack(phase_pick_buffer_per_group, axis=0)).float(),
278
+ "event_location": np.stack(event_location_buffer_per_group, axis=0),
279
+ "station_location": np.stack(station_location_buffer_per_group, axis=0),
280
+ }
281
+ del waveform_buffer_per_group[num_stations]
282
+ del phase_pick_buffer_per_group[num_stations]
283
+ del event_location_buffer_per_group[num_stations]
284
+ del station_location_buffer_per_group[num_stations]
285
+ '''
286
 
287
 
288
  def generate_label(phase_list, label_width=[150, 150], nt=8192):