Hyrax default configuration#
The following is the complete default configuration file for Hyrax. This file is used when no custom configuration file is provided. See Configuration for how to override individual values and The Hyrax Configuration System for how Hyrax layers multiple config sources.
1[general]
2# Set to `true` during development to skip checking for default config values
3# in external libraries. Use `false` otherwise.
4dev_mode = false
5
6# Destination of log messages. Options: 'stderr', 'stdout' specify the console,
7# "path/to/hyrax.log" specifies a file.
8log_destination = "stderr"
9
10# Lowest log level to emit. Options: "critical", "error", "warning", "info", "debug".
11log_level = "info"
12
13# Directory where data is stored.
14data_dir = "./data"
15
16# Top level directory for writing results.
17results_dir = "./results"
18
19# Use multiple GPUs if available
20# Set to `false` by default for development purposes for now
21distributed = false
22
23[download]
24# Cut out width in arcseconds.
25sw = "22asec"
26
27# Cut out height in arcseconds.
28sh = "22asec"
29
30# The filters to download.
31filter = ["HSC-G", "HSC-R", "HSC-I", "HSC-Z", "HSC-Y"]
32
33# The type of data to download.
34type = "coadd"
35
36# The data release to download from.
37rerun = "pdr3_wide"
38
39# Path to credentials.ini file for the downloader. File contents should be:
40# username = "<your username>"
41# password = "<your password>"
42credentials_file = "./credentials.ini"
43
44# Alternate way to pass credentials to the downloader. Users should prefer a
45# credentials.ini file to avoid exposing credentials with source control.
46username = false
47password = false
48
49# The number of sources to download from the catalog. Default is -1, which
50# downloads all sources in the catalog.
51num_sources = -1
52
53# The number of concurrent connections to use when downloading data.
54concurrent_connections = 4
55
56# The number of seconds between printing download statistics.
57stats_print_interval = 60
58
59# The path to the catalog file that defines which cutouts to download.
60fits_file = "./catalog.fits"
61
62# The number of seconds to wait before retrying a failed HTTP request in seconds.
63retry_wait = 30
64
65# How many times to retry a failed HTTP request before moving on to the next one.
66retries = 3
67
68# Number of seconds to wait for a full HTTP response from the server.
69timeout = 3600
70
71# The number of sky location rectangles should we request in a single request.
72chunk_size = 990
73
74# Request the image layer from the cutout service
75image = true
76
77# Request the variance layer from the cutout service
78variance = false
79
80# Request the mask layer from the cutout service
81mask = false
82
83
84[model]
85# NOTE: All parameters are NOT used by all models. Check the model code before training.
86
87# The name of the model to use. Option are a built-in model class name or import path
88# to an external model. e.g. "HyraxAutoencoder", "user_pkg.model.ExternalModel"
89name = ""
90
91
92[model.HyraxAutoencoder]
93# The shape of one input image as [channels, height, width]. The layers are sized from
94# this when the model is built, so it must match what the dataset produces (see
95# [data_set].crop_to). The default matches HyraxCifarDataset.
96input_shape = [3, 32, 32]
97
98# The number of output channels from the first layer.
99base_channel_size = 32
100
101# The length of the latent space vector.
102latent_dim = 64
103
104
105[model.HyraxAutoencoderV2]
106# The shape of one input image as [channels, height, width]. The layers are sized from
107# this when the model is built, so it must match what the dataset produces (see
108# [data_set].crop_to). The default matches HyraxCifarDataset.
109input_shape = [3, 32, 32]
110
111# The number of output channels from the first layer.
112base_channel_size = 32
113
114# The length of the latent space vector.
115latent_dim = 64
116
117# The activation function of the final layer.
118final_layer = "tanh"
119
120[model.ImageDCAE]
121# The shape of one input image as [channels, height, width]. The layers are sized from
122# this when the model is built, so it must match what the dataset produces (see
123# [data_set].crop_to). The default matches HyraxCifarDataset.
124input_shape = [3, 32, 32]
125
126# The number of output channels from the first layer.
127base_channel_size = 32
128
129# The length of the latent space vector.
130latent_dim = 512
131
132# The activation function of the final layer.
133final_layer = "identity"
134
135
136[model.SimCLR]
137# The dimension of the projection head for SimCLR
138projection_dimension = 128
139
140# The scalar temperature parameter for its loss function, NTXentLoss, for SimCLR
141temperature = 0.5
142
143# The probability of applying horizontal flip augmentation for SimCLR
144horizontal_flip_probability = 0.5
145
146# The parameters for color jitter augmentation for SimCLR
147# [brightness, contrast, saturation, hue]
148color_jitter_params = [0.8, 0.8, 0.8, 0.2]
149
150# The probability of applying color jitter augmentation for SimCLR
151color_jitter_probability = 0.8
152
153# The probability of applying grayscale augmentation for SimCLR
154grayscale_probability = 0.2
155
156# The kernel size of Gaussian blur augmentation for SimCLR
157gaussian_blur_kernel_size = 9
158
159# The sigma range used in Gaussian blur augmentation for SimCLR
160gaussian_blur_sigma_range = [0.1, 2.0]
161
162
163[model.HyraxCNN]
164# The shape of one input image as [channels, height, width]. The layers are sized from
165# this when the model is built, so it must match what the dataset produces (see
166# [data_set].crop_to). The default matches HyraxCifarDataset.
167input_shape = [3, 32, 32]
168
169# The number of classes to predict as the output of the model. i.e. 2 would be a
170# binary classifer, 10 would predict the 10 classes in the CiFAR dataset.
171output_classes = 10
172
173
174[model.HyraxTs2Vec]
175# The length of the representation vector produced for each object. This is the
176# width of the latent space that `reduce_dimensions` will consume.
177output_dims = 320
178
179# The channel width inside the dilated convolution encoder.
180hidden_dims = 64
181
182# The number of residual blocks in the encoder. Dilation doubles each block, so the
183# receptive field grows as 2**depth; keep 2**depth at or above the sequence length.
184depth = 10
185
186# How timestamps are masked during training, the augmentation the contrastive loss
187# learns from. Options: "binomial", "continuous", "all_true", "all_false", "mask_last".
188# Masking is disabled automatically during inference.
189mask_mode = "binomial"
190
191# The hierarchy level below which the temporal contrastive term is skipped. Also sets the
192# shortest allowed crop, 2**(temporal_unit + 1), and so the shortest usable sequence.
193temporal_unit = 0
194
195# Weight on the instance-contrastive loss term; (1 - alpha) weights the temporal term.
196alpha = 0.5
197
198# Learning rate for the AdamW optimizer this model builds for itself. HyraxTs2Vec does not
199# use [optimizer]; TS2Vec is specified on AdamW, not the Hyrax default of SGD.
200learning_rate = 0.001
201
202# How per-timestamp representations are pooled into the inference output.
203# "full_series" -> max-pool the whole series to one vector per object, shape
204# (batch, output_dims). This is what `reduce_dimensions` expects.
205# "multiscale" -> concatenate max-pools at several scales, keeping the time axis.
206encoding_window = "full_series"
207
208# --- Light-curve input encoding -------------------------------------------------------
209# The keys below describe the data rather than the architecture: prepare_inputs reads them
210# to turn padded light-curve columns into the (batch, time, channels) event sequence the
211# encoder consumes. They live here, with the model, because the encoding is part of this
212# model's input contract - any dataset that supplies the named columns will work.
213
214# Nested light-curve columns are exposed as "<column>_<subcolumn>", so these are the
215# flattened names, not the dotted ones. Defaults follow LSST naming; change them to match
216# whatever catalog you stream, and list the same names in the data request's `fields`.
217time_field = "lightcurve_midpointMjdTai"
218flux_field = "lightcurve_psfFlux"
219flux_err_field = "lightcurve_psfFluxErr"
220
221# Set to `false` for a single-band survey such as TESS or Kepler, which has no band column.
222# The one-hot band block is then dropped and `bands` is ignored, leaving 3 channels.
223band_field = "lightcurve_band"
224
225# The bands to encode, in one-hot channel order. Sets the width of the band block, so the
226# model's input width is 3 + len(bands). Observations in any other band are kept but get an
227# all-zero band indicator and a warning. Integer band codes are read as indices into this
228# list. Extend this when a cross-match brings in additional filters.
229bands = ["u", "g", "r", "i", "z", "y"]
230
231# Observations kept per object. Curves longer than this are truncated to their earliest
232# observations; shorter ones are padded and masked off. This fixes the time axis, so every
233# batch has the same shape.
234max_sequence_length = 256
235
236# Per-object flux normalization, applied to the flux and its error on a shared scale.
237# "median_mad" -> subtract the median, divide by 1.4826 * MAD (robust to outbursts)
238# "zscore" -> subtract the mean, divide by the standard deviation
239# false -> leave raw fluxes alone
240# Raw fluxes span orders of magnitude across bands and objects, and TS2Vec applies no
241# normalization of its own, so disabling this will usually stall training.
242normalize = "median_mad"
243
244
245[criterion]
246# The name of the built-in criterion to use or the import path to an external criterion
247name = "torch.nn.CrossEntropyLoss"
248
249# Whether to "sum" or "mean" loss across channels. Only used by HyraxAutoencoderV2
250band_loss_reduction = "mean"
251
252
253[optimizer]
254# The name of the built-in optimizer to use or the import path to an external optimizer
255name = "torch.optim.SGD"
256
257
258["torch.optim.SGD"]
259# learning rate for torch.optim.SGD optimizer.
260lr = 0.01
261
262# momentum for torch.optim.SGD optimizer.
263momentum = 0.9
264
265["torch.optim.Adam"]
266# learning rate for torch.optim.SGD optimizer.
267lr = 0.01
268
269[scheduler]
270# name of the learning rate scheduler
271# With gamma=1, ExponentialLR will keep the learning rate constant
272# https://docs.pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.ExponentialLR.html
273name = "torch.optim.lr_scheduler.ExponentialLR"
274
275["torch.optim.lr_scheduler.ExponentialLR"]
276# the decay multipler on each epoch
277gamma = 1
278
279["torch.optim.lr_scheduler.ConstantLR"]
280
281last_epoch = -1
282
283[train]
284# The name of the file were the model weights will be saved after training.
285weights_filename = "example_model.pth"
286
287#The number of epochs to train for.
288epochs = 10
289
290# If resuming from a check point, set to the path of the checkpoint file.
291# Otherwise set to `false` to start training from the beginning.
292resume = false
293
294# Path to a pre-trained model weights file to use as the starting point for fine-tuning.
295# If `false`, training starts from randomly initialized weights. Cannot be set with `resume`.
296model_weights_file = false
297
298# The data_set split to use when training a model.
299split = "train"
300
301# The name of the experiment when logging training results to mlflow
302experiment_name = "notebook"
303
304# The name of the run when logging training results to mlflow.
305# If false, uses result directory string, <timestamp>-train-<uid>, as run name.
306run_name = false
307
308# If true, shuffle training samples each epoch. Only the train verb uses this option.
309shuffle = true
310
311
312[test]
313# The path to the model weights file to use for testing. If `false`, the most recent
314# training results (from [train]) will be used.
315model_weights_file = false
316
317# The name of the experiment when logging test results to mlflow.
318# If not set, uses the experiment name from [train].
319experiment_name = "notebook"
320
321# The name of the run when logging test results to mlflow.
322# If false, uses result directory string, <timestamp>-test-<uid>, as run name.
323run_name = false
324
325
326
327[onnx]
328
329# The operator set version to use when exporting a model. See the following for info:
330# https://onnxruntime.ai/docs/reference/compatibility.html#onnx-opset-support
331opset_version = 20
332
333# The directory to find input model files to convert to ONNX. ONNX-ified models
334# will be written to this directory as well.
335input_model_directory = false
336
337
338# [data_request]
339# Top-level table that defines the dataset(s) used for training, validation, and inference.
340# Configure with [data_request.train], [data_request.validate], or [data_request.infer] sections.
341
342
343[data_set]
344# Crop pixel dimensions for images, e.g., [100, 100]. If false, scans for the
345# smallest image size in [general].data_dir and uses it.
346crop_to = false
347
348# Used by HSCDataset, LSSTDataset, and DownloadedLSSTDataset.
349# Limit to only particular filters. When `false`, use all filters.
350# Options: ["HSC-G", "HSC-R", "HSC-I", "HSC-Z", "HSC-Y"] for HSC
351# Options: ["u", "g", "r", "i", "z" , "y"] for LSST
352filters = false
353
354# Path to a fits file that specifies object IDs to use from the data stored in
355# [general].data_dir. Implementation is data_set class dependent. Use `false` for no filtering.
356filter_catalog = false
357
358# The transformation to be applied to images before being passed on to the model
359# This must be a valid Numpy function. Passing false will result in no transformations
360# (other than cropping) be applied to the images.
361transform = "tanh"
362
363# Number to seed with for generating a random split. Use `false` to seed from a
364# system source at runtime.
365seed = false
366
367# If `true`, cache samples in memory during training to reduce runtime after the
368# first epoch. Set to `false` when running inference or on memory-constrained systems.
369use_cache = true
370
371
372# Override the name of the object_id column for FitsImageDataset, HSCDataset and DownloadedLSSTDataset
373object_id_column_name = false
374
375# Override the name of the filter column for FitsImageDataset and HSCDataset
376filter_column_name = false
377
378# Override the name of the filename column for FitsImageDataset and HSCDataset
379filename_column_name = false
380
381# Replace NaN in input data with a value, modes are false for no replacement or "quantile" to replace with a
382# defined quantile of the non-NaN data, see nan_quantile.
383nan_mode = false
384
385# When replacing NaN values with a quantile, which quantile in the non-nan tensor should be used.
386nan_quantile = 0.05
387
388# The astropy table to use as a catalog in LSSTDataset and friends
389astropy_table = false
390
391# Semi width in degrees of cutouts made from the butler (17 arcsec)
392semi_width_deg = 0.00472
393
394# Semi height in degrees of cutouts made from the butler (17 arcsec)
395semi_height_deg = 0.00472
396
397
398[data_set.HyraxCifarDataset]
399# If `true`, download CIFAR10 training set, otherwise download test set.
400use_training_data = true
401
402
403[data_set.HyraxRandomDataset]
404# Total number of samples produced by the random dataset
405size = 100
406
407# The dimensions of the numpy arrays that will be produced for each sample represented
408# as a list where each element is the size of dimension.
409shape = [2,5,5]
410
411# Seed to use for random number generation
412seed = 42
413
414# If a list is provided, the data will have randomly labeled with values from the list
415# If set to false, no labels will be included with the data.
416provided_labels = [0, 1, 2]
417
418# List of metadata field names. These will be populated with dummy data.
419metadata_fields = ["meta_field_1", "meta_field_2"]
420
421# Set this to a positive integer to randomly replace some values with an "invalid" value.
422number_invalid_values = 0
423
424# The value to use for invalid values in the data. Must be one of the following:
425# "nan", "inf", "-inf", "none" or a float value.
426invalid_value_type = "nan"
427
428[data_set.HyraxHATSDataset]
429# Extra keyword arguments passed directly to lsdb.open_catalog
430[data_set.HyraxHATSDataset.open_catalog_kwargs]
431
432[data_set.NestedPandasDataset]
433# Extra keyword arguments passed directly to the selected nested_pandas reader.
434[data_set.NestedPandasDataset.read_kwargs]
435# engine = "auto" # pandas default, but can be "pyarrow", "fastparquet"
436# See pandas.read_parquet API reference for all possible arguments.
437
438[data_set.LanceDBDataset]
439# Table name to open inside the LanceDB database.
440table_name = false
441
442# Extra keyword arguments passed directly to lancedb.connect.
443[data_set.LanceDBDataset.connect_kwargs]
444
445# Extra keyword arguments passed directly to db.open_table.
446[data_set.LanceDBDataset.open_table_kwargs]
447
448[data_set.MultimodalUniverseDataset]
449# Hugging Face split name to load (for example: "train", "validation", or "test").
450split = "train"
451
452# Maximum number of rows to load from the split. Set to false for no explicit limit.
453max_samples = 100
454
455# Stream rows from Hugging Face instead of downloading full files.
456# If true, `max_samples` must be set.
457streaming = true
458
459
460[data_set.KafkaStreamDataset]
461# Kafka bootstrap broker address.
462bootstrap_servers = "localhost:9092"
463
464# List of Kafka topics to consume. Must be set, `[]` raises at runtime. A single
465# topic string is accepted for convenience, and will be converted to a list of one topic.
466topics = []
467
468# Kafka consumer group id. NOTE this default is shared: two runs against the same broker
469# join the same group and split its partitions between them, so on a single-partition
470# topic one of them is assigned nothing and sees no messages at all. Give concurrent runs
471# distinct group ids. A group also remembers its committed offsets, so a fresh group_id is
472# one way to replay a topic this one has already consumed.
473group_id = "hyrax-infer-stream"
474
475# Where to start consuming when the group has no committed offset: "latest" or "earliest".
476# This has no effect once the group has committed an offset -- see "Recovering / replaying
477# a stream" in the hyrax.datasets.kafka_stream_dataset docs.
478auto_offset_reset = "earliest"
479
480# Once a batch of messages has been processed downstream, commit the offsets for that batch.
481# Setting to false disables committing - used for testing to replay the stream.
482commit_after_batch = true
483
484# Seconds to block on each Kafka poll while waiting for a message.
485poll_timeout = 1.0
486
487# Maximum seconds to wait before flushing a partial (smaller than batch_size) batch.
488batch_flush_timeout = 5.0
489
490# Provided path credentials file for a `confluent_kafka` consumer instance
491# e.g.
492# "sasl.username" = <USER_NAME>
493# "sasl.password" = <PASSWORD>
494# "security.protocol" = "SASL_PLAINTEXT"
495# "sasl.mechanism" = "SCRAM-SHA-512"
496credentials_file = false
497
498# Arbitrary librdkafka settings merged into the consumer configuration, applied after
499# `credentials_file` so they win over it. Use this to tune client behavior, e.g.
500# [data_set.KafkaStreamDataset.consumer_options]
501# "max.poll.interval.ms" = 600000
502# "session.timeout.ms" = 45000
503#
504# `enable.auto.commit` is forced to false and should be left alone: offsets are committed
505# by the stream only once a batch has been fully processed downstream, so that a failure
506# mid-pipeline replays those alerts instead of silently dropping them.
507#
508# NOTE: While it is technically possible to put credentials (e.g. "sasl.username",
509# "sasl.password") here, you SHOULD NOT do so. Use `credentials_file` instead to avoid
510# accidentally committing secrets into your configuration files.
511[data_set.KafkaStreamDataset.consumer_options]
512
513
514[data_set.LSDBStreamDataset]
515# Which lsdb stream to build over the catalog:
516# "catalog" -> lsdb.streams.CatalogStream, a single finite pass over every partition.
517# Best for inference: every object is visited exactly once and the run
518# ends on its own when the catalog is exhausted.
519# "infinite" -> lsdb.streams.InfiniteStream, endless random resampling of partitions.
520# Best for training: batches keep arriving until you stop the session,
521# so training is not bounded by a single pass over the catalog.
522stream_type = "catalog"
523
524# Number of HATS partitions computed per chunk. Larger values amortize per-chunk dask
525# overhead but hold more rows in memory; lsdb clips this to the catalog's partition count.
526# With a dask client attached, set this to at least twice the worker count so a single
527# chunk can occupy the whole cluster.
528partitions_per_chunk = 1
529
530# Hand lsdb an active dask.distributed Client. lsdb submits the next chunk's future before
531# returning the current one, so with a client the next chunk computes in the background
532# while Hyrax trains on this one; with no client it computes inline and every fetch blocks.
533# Hyrax attaches to whatever client is already active, so create one before starting the
534# run. Client(processes=False) is usually the right choice on a single machine: it still
535# gives real futures, but keeps results in-process instead of serializing every chunk back
536# from a worker.
537use_dask_client = true
538
539# Explicit dask scheduler address, e.g. "tcp://127.0.0.1:8786". When false, the active
540# client is used if there is one, and the stream runs synchronously if there is not.
541dask_client_address = false
542
543# Shuffle the partition order and the rows within each chunk. Ignored (and always on)
544# when stream_type = "infinite"; lsdb's InfiniteStream does not accept this option.
545shuffle = true
546
547# Random seed for partition/row shuffling. Set to false for fresh entropy each run.
548seed = false
549
550# Extra keyword arguments passed directly to lsdb.open_catalog. Ignored when
551# data_location is an "lsdb://<name>" reference to a pre-registered catalog.
552[data_set.LSDBStreamDataset.open_catalog_kwargs]
553
554
555[data_loader]
556# The number of data points to load at once.
557batch_size = 512
558
559
560[infer]
561# The path to the model weights file to use for inference.
562model_weights_file = false
563
564
565[infer_stream]
566# Path to model weights. If false, uses the most recent training results.
567model_weights_file = false
568
569# Save the results of inference in the timestamped result directory.
570save_model_output = false
571
572
573[train_stream]
574# The name of the file where the model weights will be saved.
575weights_filename = "example_model.pth"
576
577# Path to a pre-trained model weights file to warm-start training. If `false`,
578# training starts from randomly initialized weights.
579model_weights_file = false
580
581# The name of the experiment when logging training results to mlflow.
582experiment_name = "notebook"
583
584# The name of the run when logging training results to mlflow.
585# If false, uses result directory string, <timestamp>-train_stream-<uid>, as run name.
586run_name = false
587
588# Save model weights every N processed batches. If `false`, weights are only saved on close().
589save_weights_every = false
590
591# Batches smaller than this are skipped rather than trained on. If `false`, no minimum.
592min_batch_size = false
593
594
595[vector_db]
596# The type of vector db to use. Use "false" to disable vector database.
597name = "chromadb"
598
599# The directory where the vector database will be stored. Use "false" to create
600# a new vector database in a timestamped directory. Otherwise set to a path.
601vector_db_dir = false
602
603# The path to inference results. Setting to "false" will use the most recent
604# inference results.
605infer_results_dir = false
606
607
608[vector_db.chromadb]
609# The approximate maximum size of a shard before creating a new one. A smaller
610# value will decrease insert times while increasing search times.
611shard_size_limit = 65536
612
613# Inserting vectors with more than this many elements logs a warning message. ChromaDB
614# performance degrades with vectors of this size. Set to "false" to disable warning.
615vector_size_warning = 10000
616
617
618[vector_db.qdrant]
619# The number of elements in the vectors that will be stored in the vector database.
620# This must be the same as the size of the vectors produced by the model.
621vector_size = 64
622
623
624[results]
625# Path to inference results to use for visualization and lookups. Uses latest inference run if none provided.
626inference_dir = false
627
628
629[reduce_dimensions]
630# Name of the reduction algorithm to use
631algorithm = "umap"
632
633# Save the fitted reducer model as a pickle file
634save_fit_model = true
635
636# The number of data points to use when transforming with reduction algorithm at once
637batch_size = 1024
638
639# Use multiprocessing during transforming with redudction algorithm (More memory intensive)
640parallel = false
641
642
643[reduce_dimensions.umap]
644# Number of data points used to fit the umap model.
645fit_sample_size = 1024
646
647# Path to a pre-existing umap reducer model
648model_path = false
649
650
651[reduce_dimensions.umap.kwargs]
652# Specify any parameter accepted by https://umap-learn.readthedocs.io/en/latest/api.html#umap
653# Dimension of the embedded space
654n_components = 2
655
656# Controls how UMAP balances local versus global structure in the data.
657# See official documentation for details.
658n_neighbors = 15
659
660
661[reduce_dimensions.tsne]
662# Placeholder for config values of tsne model
663
664
665[reduce_dimensions.tsne.kwargs]
666# Specify any parameter accepted by https://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html
667# Dimension of the embedded space
668n_components = 2
669
670# Number of nearest neighbors that is used in other manifold learning algorithms
671# See official documentation for details.
672perplexity = 30.0
673
674
675[reduce_dimensions.pca]
676# Number of data points used to fit the pca model.
677fit_sample_size = 1024
678
679# Path to a pre-existing pca reducer model
680model_path = false
681
682
683[reduce_dimensions.pca.kwargs]
684# Specify any parameter accepted by https://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html#
685# Dimension of the embedded space
686n_components=2
687
688
689[visualize]
690
691# List of metadata field names to use in visualizer. Must be available as metadata in your dataset
692fields = []
693
694# Whether to display a panel of randomly chosen images corresponding to the selected points
695display_images = false
696
697# Name of catalog column to use for coloring points in the scatter plot. Use false for no coloring.
698color_column = false
699
700# Colormap to use for coloring points in the scatter plot when color_column is specified
701cmap = "viridis"
702
703# Only valid for .pt tensor images. Which bands should be loaded for display
704# [0,3,5] would map bands in that order to R,G,B. Single band will be grayscale.
705torch_tensor_bands = [3]
706
707# Whether to rasterize plot. Will break coloring (Haloviews Bug)
708# Helpful to reduce lag in large datasets.
709rasterize_plot = false
710
711
712[visualize_v2]
713
714# Number of hexagonal bins across the x-axis at any zoom level
715target_bins = 50
716
717# Extra padding around the viewport as a fraction of viewport width/height
718buffer_factor = 0.2
719
720# Plot dimensions in pixels
721plot_width = 1000
722plot_height = 1000
723
724# Colormap for hexbin density
725cmap = "Viridis"
726
727# Maximum rows rendered in the selection table
728max_table_rows = 1000
729
730# Number of detail plots per tab row
731num_detail_plots = 6
732
733
734
735[engine]
736
737# The directory containing the ONNX model used for inference in production
738model_directory = false
739
740
741[split]
742# Per-group split configuration. Keys correspond to groups in [data_request].
743# Values are a fraction of the group's primary dataset to use — (0.0, 1.0]
744# or a path to a previously generated <group>_split.npz file.
745# Groups absent from this table default to 1.0 (use the full dataset).
746
747# RNG seed for shuffling/partitioning. `false` uses config["data_set"]["seed"].
748# A non-empty value uses a dedicated generator and does affect global RNG state.
749rng_seed = false
750
751
752[balance]
753# Name of the field identifying each item's class. Must resolve to a
754# get_<field> getter on the primary dataset. `false` disables stratification.
755field = false
756
757# Data groups whose sampling weights are adjusted to the target distribution.
758# Empty list disables rebalancing. See [balance.distribution] for the target.
759groups = []
760
761[balance.distribution]
762# Maps class label → target fraction in (0.0, 1.0]. Empty table means equal
763# target across all observed classes (uniform). Values must sum to exactly 1.0.
764# Example:
765# cat = 0.25
766# dog = 0.75
767
768[label]
769# Optional. Maps human-readable string aliases to the raw values returned by
770# get_<balance.field>. Required when balance.distribution is used and
771# get_<balance.field> returns non-string values (e.g. integers),
772# All balance.distribution keys must appear here.
773# Example:
774# cat = 0
775# dog = 1
776# bird = 2