Failed to run the training example

The Seismology Benchmark collection (SeisBench) is an open-source python toolbox for machine learning in seismology. It provides a unified API for accessing seismic datasets and both training and applying machine learning algorithms to seismic data. SeisBench has been built to reduce the overhead when applying or developing machine learning techniques for seismological tasks.

Getting started

SeisBench offers three core modules, data, models, and generate. data provides access to benchmark datasets and offers functionality for loading datasets. models offers a collection of machine learning models for seismology. You can easily create models, load pretrained models or train models on any dataset. generate contains tools for building data generation pipelines. They bridge the gap between data and models.

The easiest way of getting started is through our colab notebooks.

Examples
Dataset basics
Model API
Generator Pipelines
Applied picking
Using DeepDenoiser
Depth phases and earthquake depth
Training PhaseNet (advanced)
Creating a dataset (advanced)
Building an event catalog with GaMMA (advanced)
Building an event catalog with PyOcto (advanced)

Alternatively, you can clone the repository and run the same examples locally.

For more detailed information on Seisbench check out the SeisBench documentation.

Installation

SeisBench can be installed in two ways. In both cases, you might consider installing SeisBench in a virtual environment, for example using conda.

The recommended way is installation through pip. Simply run:

pip install seisbench

Alternatively, you can install the latest version from source. For this approach, clone the repository, switch to the repository root and run:

pip install .

which will install SeisBench in your current python environment.

CPU only installation

SeisBench is built on pytorch, which in turn runs on CUDA for GPU acceleration. Sometimes, it might be preferable to install pytorch without CUDA, for example, because CUDA will not be used and the CUDA binaries are rather large. To install such a pure CPU version, the easiest way is to follow a two-step installation. First, install pytorch in a pure CPU version as explained here. Second, install SeisBench the regular way through pip. Example instructions would be:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
pip install seisbench

Contributing

There are many ways to contribute to SeisBench and we are always looking forward to your contributions. Check out the contribution guidelines for details on how to contribute.

Known issues

Some institutions and internet providers are blocking access to our data and model repository, as it is running on a non-standard port (2880). This usually manifests in timeouts when trying to download data or model weights. To verify the issue, try accessing https://hifis-storage.desy.de:2880/ directly from the same machine. As a mitigation, you can use our backup repository. Just run seisbench.use_backup_repository(). Please note that the backup repository will usually show lower download speeds. We recommend contacting your network administrator to allow outgoing access to TCP port 2880 on our server as a higher performance solution.

We've recently changed the URL of the SeisBench repository. To use the new URL update to SeisBench 0.4.1. It this is not possible, you can use the following commands within your runtime to update the URL manually:

import seisbench
from urllib.parse import urljoin

seisbench.remote_root = "https://hifis-storage.desy.de:2880/Helmholtz/HelmholtzAI/SeisBench/"
seisbench.remote_data_root = urljoin(seisbench.remote_root, "datasets/")
seisbench.remote_model_root = urljoin(seisbench.remote_root, "models/v3/")

On the Apple M1 and M2 chips, pytorch seems to not always work when installed directly within pip install seisbench. As a workaround, follow the instructions at (https://pytorch.org/) to install pytorch and then install SeisBench as usual through pip.
EQTransformer model weights "original" in version 1 and 2 are incompatible with SeisBench >=0.2.3. Simply use from_pretrained("original", version="3") or from_pretrained("original", update=True). The weights will not differ in their predictions.

References

Reference publications for SeisBench:

SeisBench - A Toolbox for Machine Learning in Seismology

Reference publication for software.

Which picker fits my data? A quantitative evaluation of deep learning based seismic pickers

Example of in-depth bencharking study of deep learning-based picking routines using the SeisBench framework.

Acknowledgement

The initial version of SeisBench has been developed at GFZ Potsdam and KIT with funding from Helmholtz AI. The SeisBench repository is hosted by HIFIS - Helmholtz Federated IT Services.

	def export_model(row):
	output_base = Path("seisbench_models")
	weights = Path("weights") / row["experiment"]

	version = sorted(weights.iterdir())[-1]
	config_path = version / "hparams.yaml"
	with open(config_path, "r") as f:
	# config = yaml.safe_load(f)
	config = yaml.full_load(f)

	model_cls = models.__getattribute__(config["model"] + "Lit")
	model = load_best_model(model_cls, weights, version.name)

	output_path = output_base / row["model"] / f"{row['data']}.pt.v1"
	json_path = output_base / row["model"] / f"{row['data']}.json.v1"
	output_path.parent.mkdir(parents=True, exist_ok=True)
	torch.save(model.model.state_dict(), output_path)

	meta = generate_metadata(row)
	with open(json_path, "w") as f:
	json.dump(meta, f, indent=4)

	def predict_step(self, batch, batch_idx=None, dataloader_idx=None):
	x = batch["X"]
	window_borders = batch["window_borders"]

	pred = self.model(x)

	score_detection = torch.zeros(pred.shape[0])
	score_p_or_s = torch.zeros(pred.shape[0])
	p_sample = torch.zeros(pred.shape[0], dtype=int)
	s_sample = torch.zeros(pred.shape[0], dtype=int)

	for i in range(pred.shape[0]):
	start_sample, end_sample = window_borders[i]
	local_pred = pred[i, :, start_sample:end_sample]

	score_detection[i] = torch.max(1 - local_pred[-1]) # 1 - noise
	score_p_or_s[i] = torch.max(local_pred[0]) / torch.max(
	local_pred[1]
	) # most likely P by most likely S

	p_sample[i] = torch.argmax(local_pred[0])
	s_sample[i] = torch.argmax(local_pred[1])

	return score_detection, score_p_or_s, p_sample, s_sample

	generator = sbg.SteeredGenerator(split, task_targets)
	generator.add_augmentations(model.get_eval_augmentations())

seisbench / pick-benchmark Goto Github PK

pick-benchmark's Introduction

Getting started

Installation

CPU only installation

Contributing

Known issues

References

Acknowledgement

pick-benchmark's People

Contributors

Stargazers

Watchers

Forkers

pick-benchmark's Issues

Problem

My try

Training outputs

Evaluation

Recommend Projects

Recommend Topics

Recommend Org