File size: 50,742 Bytes
9e9cca9 f75f5ac 9e9cca9 f75f5ac 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 eeba323 9e9cca9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 |
"""
Geneformer classifier.
**Input data:**
| Cell state classifier:
| Single-cell transcriptomes as Geneformer rank value encodings with cell state labels in Geneformer .dataset format (generated from single-cell RNAseq data by tokenizer.py)
| Gene classifier:
| Dictionary in format {Gene_label: list(genes)} for gene labels and single-cell transcriptomes as Geneformer rank value encodings in Geneformer .dataset format (generated from single-cell RNAseq data by tokenizer.py)
**Usage:**
.. code-block :: python
>>> from geneformer import Classifier
>>> cc = Classifier(classifier="cell", # example of cell state classifier
... cell_state_dict={"state_key": "disease", "states": "all"},
... filter_data={"cell_type":["Cardiomyocyte1","Cardiomyocyte2","Cardiomyocyte3"]},
... training_args=training_args,
... freeze_layers = 2,
... num_crossval_splits = 1,
... forward_batch_size=200,
... nproc=16)
>>> cc.prepare_data(input_data_file="path/to/input_data",
... output_directory="path/to/output_directory",
... output_prefix="output_prefix")
>>> all_metrics = cc.validate(model_directory="path/to/model",
... prepared_input_data_file=f"path/to/output_directory/{output_prefix}_labeled.dataset",
... id_class_dict_file=f"path/to/output_directory/{output_prefix}_id_class_dict.pkl",
... output_directory="path/to/output_directory",
... output_prefix="output_prefix",
... predict_eval=True)
>>> cc.plot_conf_mat(conf_mat_dict={"Geneformer": all_metrics["conf_matrix"]},
... output_directory="path/to/output_directory",
... output_prefix="output_prefix",
... custom_class_order=["healthy","disease1","disease2"])
>>> cc.plot_predictions(predictions_file=f"path/to/output_directory/datestamp_geneformer_cellClassifier_{output_prefix}/ksplit1/predictions.pkl",
... id_class_dict_file=f"path/to/output_directory/{output_prefix}_id_class_dict.pkl",
... title="disease",
... output_directory="path/to/output_directory",
... output_prefix="output_prefix",
... custom_class_order=["healthy","disease1","disease2"])
"""
import datetime
import logging
import os
import pickle
import subprocess
from pathlib import Path
import numpy as np
import pandas as pd
import seaborn as sns
from sklearn.model_selection import StratifiedKFold
from tqdm.auto import tqdm, trange
from transformers import Trainer
from transformers.training_args import TrainingArguments
from . import DataCollatorForCellClassification, DataCollatorForGeneClassification
from . import classifier_utils as cu
from . import evaluation_utils as eu
from . import perturber_utils as pu
from .tokenizer import TOKEN_DICTIONARY_FILE
sns.set()
logger = logging.getLogger(__name__)
class Classifier:
valid_option_dict = {
"classifier": {"cell", "gene"},
"cell_state_dict": {None, dict},
"gene_class_dict": {None, dict},
"filter_data": {None, dict},
"rare_threshold": {int, float},
"max_ncells": {None, int},
"max_ncells_per_class": {None, int},
"training_args": {None, dict},
"freeze_layers": {int},
"num_crossval_splits": {0, 1, 5},
"eval_size": {int, float},
"no_eval": {bool},
"stratify_splits_col": {None, str},
"forward_batch_size": {int},
"nproc": {int},
}
def __init__(
self,
classifier=None,
cell_state_dict=None,
gene_class_dict=None,
filter_data=None,
rare_threshold=0,
max_ncells=None,
max_ncells_per_class=None,
training_args=None,
freeze_layers=0,
num_crossval_splits=1,
eval_size=0.2,
stratify_splits_col=None,
no_eval=False,
forward_batch_size=100,
nproc=4,
):
"""
Initialize Geneformer classifier.
**Parameters:**
classifier : {"cell", "gene"}
| Whether to fine-tune a cell state or gene classifier.
cell_state_dict : None, dict
| Cell states to fine-tune model to distinguish.
| Two-item dictionary with keys: state_key and states
| state_key: key specifying name of column in .dataset that defines the states to model
| states: list of values in the state_key column that specifies the states to model
| Alternatively, instead of a list of states, can specify "all" to use all states in that state key from input data.
| Of note, if using "all", states will be defined after data is filtered.
| Must have at least 2 states to model.
| For example: {"state_key": "disease",
| "states": ["nf", "hcm", "dcm"]}
| or
| {"state_key": "disease",
| "states": "all"}
gene_class_dict : None, dict
| Gene classes to fine-tune model to distinguish.
| Dictionary in format: {Gene_label_A: list(geneA1, geneA2, ...),
| Gene_label_B: list(geneB1, geneB2, ...)}
| Gene values should be Ensembl IDs.
filter_data : None, dict
| Default is to fine-tune with all input data.
| Otherwise, dictionary specifying .dataset column name and list of values to filter by.
rare_threshold : float
| Threshold below which rare cell states should be removed.
| For example, setting to 0.05 will remove cell states representing
| < 5% of the total cells from the cell state classifier's possible classes.
max_ncells : None, int
| Maximum number of cells to use for fine-tuning.
| Default is to fine-tune with all input data.
max_ncells_per_class : None, int
| Maximum number of cells per cell class to use for fine-tuning.
| Of note, will be applied after max_ncells above.
| (Only valid for cell classification.)
training_args : None, dict
| Training arguments for fine-tuning.
| If None, defaults will be inferred for 6 layer Geneformer.
| Otherwise, will use the Hugging Face defaults:
| https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments
| Note: Hyperparameter tuning is highly recommended, rather than using defaults.
freeze_layers : int
| Number of layers to freeze from fine-tuning.
| 0: no layers will be frozen; 2: first two layers will be frozen; etc.
num_crossval_splits : {0, 1, 5}
| 0: train on all data without splitting
| 1: split data into train and eval sets by designated eval_size
| 5: split data into 5 folds of train and eval sets by designated eval_size
eval_size : None, float
| Proportion of data to hold out for evaluation (e.g. 0.2 if intending 80:20 train/eval split)
stratify_splits_col : None, str
| Name of column in .dataset to be used for stratified splitting.
| Proportion of each class in this column will be the same in the splits as in the original dataset.
no_eval : bool
| If True, will skip eval step and use all data for training.
| Otherwise, will perform eval during training.
forward_batch_size : int
| Batch size for forward pass (for evaluation, not training).
nproc : int
| Number of CPU processes to use.
"""
self.classifier = classifier
self.cell_state_dict = cell_state_dict
self.gene_class_dict = gene_class_dict
self.filter_data = filter_data
self.rare_threshold = rare_threshold
self.max_ncells = max_ncells
self.max_ncells_per_class = max_ncells_per_class
self.training_args = training_args
self.freeze_layers = freeze_layers
self.num_crossval_splits = num_crossval_splits
self.eval_size = eval_size
self.stratify_splits_col = stratify_splits_col
self.no_eval = no_eval
self.forward_batch_size = forward_batch_size
self.nproc = nproc
if self.training_args is None:
logger.warning(
"Hyperparameter tuning is highly recommended for optimal results. "
"No training_args provided; using default hyperparameters."
)
self.validate_options()
if self.filter_data is None:
self.filter_data = dict()
if self.classifier == "cell":
if self.cell_state_dict["states"] != "all":
self.filter_data[
self.cell_state_dict["state_key"]
] = self.cell_state_dict["states"]
# load token dictionary (Ensembl IDs:token)
with open(TOKEN_DICTIONARY_FILE, "rb") as f:
self.gene_token_dict = pickle.load(f)
self.token_gene_dict = {v: k for k, v in self.gene_token_dict.items()}
# filter genes for gene classification for those in token dictionary
if self.classifier == "gene":
all_gene_class_values = set(pu.flatten_list(self.gene_class_dict.values()))
missing_genes = [
gene
for gene in all_gene_class_values
if gene not in self.gene_token_dict.keys()
]
if len(missing_genes) == len(all_gene_class_values):
logger.error(
"None of the provided genes to classify are in token dictionary."
)
raise
elif len(missing_genes) > 0:
logger.warning(
f"Genes to classify {missing_genes} are not in token dictionary."
)
self.gene_class_dict = {
k: set([self.gene_token_dict.get(gene) for gene in v])
for k, v in self.gene_class_dict.items()
}
empty_classes = []
for k, v in self.gene_class_dict.items():
if len(v) == 0:
empty_classes += [k]
if len(empty_classes) > 0:
logger.error(
f"Class(es) {empty_classes} did not contain any genes in the token dictionary."
)
raise
def validate_options(self):
# confirm arguments are within valid options and compatible with each other
for attr_name, valid_options in self.valid_option_dict.items():
attr_value = self.__dict__[attr_name]
if not isinstance(attr_value, (list, dict)):
if attr_value in valid_options:
continue
valid_type = False
for option in valid_options:
if (option in [int, float, list, dict, bool]) and isinstance(
attr_value, option
):
valid_type = True
break
if valid_type:
continue
logger.error(
f"Invalid option for {attr_name}. "
f"Valid options for {attr_name}: {valid_options}"
)
raise
if self.filter_data is not None:
for key, value in self.filter_data.items():
if not isinstance(value, list):
self.filter_data[key] = [value]
logger.warning(
"Values in filter_data dict must be lists. "
f"Changing {key} value to list ([{value}])."
)
if self.classifier == "cell":
if set(self.cell_state_dict.keys()) != set(["state_key", "states"]):
logger.error(
"Invalid keys for cell_state_dict. "
"The cell_state_dict should have only 2 keys: state_key and states"
)
raise
if self.cell_state_dict["states"] != "all":
if not isinstance(self.cell_state_dict["states"], list):
logger.error(
"States in cell_state_dict should be list of states to model."
)
raise
if len(self.cell_state_dict["states"]) < 2:
logger.error(
"States in cell_state_dict should contain at least 2 states to classify."
)
raise
if self.classifier == "gene":
if len(self.gene_class_dict.keys()) < 2:
logger.error(
"Gene_class_dict should contain at least 2 gene classes to classify."
)
raise
def prepare_data(
self,
input_data_file,
output_directory,
output_prefix,
split_id_dict=None,
test_size=None,
attr_to_split=None,
attr_to_balance=None,
max_trials=100,
pval_threshold=0.1,
):
"""
Prepare data for cell state or gene classification.
**Parameters**
input_data_file : Path
| Path to directory containing .dataset input
output_directory : Path
| Path to directory where prepared data will be saved
output_prefix : str
| Prefix for output file
split_id_dict : None, dict
| Dictionary of IDs for train and test splits
| Three-item dictionary with keys: attr_key, train, test
| attr_key: key specifying name of column in .dataset that contains the IDs for the data splits
| train: list of IDs in the attr_key column to include in the train split
| test: list of IDs in the attr_key column to include in the test split
| For example: {"attr_key": "individual",
| "train": ["patient1", "patient2", "patient3", "patient4"],
| "test": ["patient5", "patient6"]}
test_size : None, float
| Proportion of data to be saved separately and held out for test set
| (e.g. 0.2 if intending hold out 20%)
| The training set will be further split to train / validation in self.validate
| Note: only available for CellClassifiers
attr_to_split : None, str
| Key for attribute on which to split data while balancing potential confounders
| e.g. "patient_id" for splitting by patient while balancing other characteristics
| Note: only available for CellClassifiers
attr_to_balance : None, list
| List of attribute keys on which to balance data while splitting on attr_to_split
| e.g. ["age", "sex"] for balancing these characteristics while splitting by patient
| Note: only available for CellClassifiers
max_trials : None, int
| Maximum number of trials of random splitting to try to achieve balanced other attributes
| If no split is found without significant (p<0.05) differences in other attributes, will select best
| Note: only available for CellClassifiers
pval_threshold : None, float
| P-value threshold to use for attribute balancing across splits
| E.g. if set to 0.1, will accept trial if p >= 0.1 for all attributes in attr_to_balance
"""
# prepare data and labels for classification
data = pu.load_and_filter(self.filter_data, self.nproc, input_data_file)
if self.classifier == "cell":
if "label" in data.features:
logger.error(
"Column name 'label' must be reserved for class IDs. Please rename column."
)
raise
elif self.classifier == "gene":
if "labels" in data.features:
logger.error(
"Column name 'labels' must be reserved for class IDs. Please rename column."
)
raise
if self.classifier == "cell":
# remove cell states representing < rare_threshold of cells
data = cu.remove_rare(
data, self.rare_threshold, self.cell_state_dict["state_key"], self.nproc
)
# downsample max cells and max per class
data = cu.downsample_and_shuffle(
data, self.max_ncells, self.max_ncells_per_class, self.cell_state_dict
)
# rename cell state column to "label"
data = cu.rename_cols(data, self.cell_state_dict["state_key"])
# convert classes to numerical labels and save as id_class_dict
# of note, will label all genes in gene_class_dict
# if (cross-)validating, genes will be relabeled in column "labels" for each split
# at the time of training with Classifier.validate
data, id_class_dict = cu.label_classes(
self.classifier, data, self.gene_class_dict, self.nproc
)
# save id_class_dict for future reference
id_class_output_path = (
Path(output_directory) / f"{output_prefix}_id_class_dict"
).with_suffix(".pkl")
with open(id_class_output_path, "wb") as f:
pickle.dump(id_class_dict, f)
if split_id_dict is not None:
data_dict = dict()
data_dict["train"] = pu.filter_by_dict(
data, {split_id_dict["attr_key"]: split_id_dict["train"]}, self.nproc
)
data_dict["test"] = pu.filter_by_dict(
data, {split_id_dict["attr_key"]: split_id_dict["test"]}, self.nproc
)
train_data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled_train"
).with_suffix(".dataset")
test_data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled_test"
).with_suffix(".dataset")
data_dict["train"].save_to_disk(train_data_output_path)
data_dict["test"].save_to_disk(test_data_output_path)
elif (test_size is not None) and (self.classifier == "cell"):
if 1 > test_size > 0:
if attr_to_split is None:
data_dict = data.train_test_split(
test_size=test_size,
stratify_by_column=self.stratify_splits_col,
seed=42,
)
train_data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled_train"
).with_suffix(".dataset")
test_data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled_test"
).with_suffix(".dataset")
data_dict["train"].save_to_disk(train_data_output_path)
data_dict["test"].save_to_disk(test_data_output_path)
else:
data_dict, balance_df = cu.balance_attr_splits(
data,
attr_to_split,
attr_to_balance,
test_size,
max_trials,
pval_threshold,
self.cell_state_dict["state_key"],
self.nproc,
)
balance_df.to_csv(
f"{output_directory}/{output_prefix}_train_test_balance_df.csv"
)
train_data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled_train"
).with_suffix(".dataset")
test_data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled_test"
).with_suffix(".dataset")
data_dict["train"].save_to_disk(train_data_output_path)
data_dict["test"].save_to_disk(test_data_output_path)
else:
data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled"
).with_suffix(".dataset")
data.save_to_disk(data_output_path)
print(data_output_path)
else:
data_output_path = (
Path(output_directory) / f"{output_prefix}_labeled"
).with_suffix(".dataset")
data.save_to_disk(data_output_path)
def train_all_data(
self,
model_directory,
prepared_input_data_file,
id_class_dict_file,
output_directory,
output_prefix,
save_eval_output=True,
):
"""
Train cell state or gene classifier using all data.
**Parameters**
model_directory : Path
| Path to directory containing model
prepared_input_data_file : Path
| Path to directory containing _labeled.dataset previously prepared by Classifier.prepare_data
id_class_dict_file : Path
| Path to _id_class_dict.pkl previously prepared by Classifier.prepare_data
| (dictionary of format: numerical IDs: class_labels)
output_directory : Path
| Path to directory where model and eval data will be saved
output_prefix : str
| Prefix for output files
save_eval_output : bool
| Whether to save cross-fold eval output
| Saves as pickle file of dictionary of eval metrics
**Output**
Returns trainer after fine-tuning with all data.
"""
##### Load data and prepare output directory #####
# load numerical id to class dictionary (id:class)
with open(id_class_dict_file, "rb") as f:
id_class_dict = pickle.load(f)
class_id_dict = {v: k for k, v in id_class_dict.items()}
# load previously filtered and prepared data
data = pu.load_and_filter(None, self.nproc, prepared_input_data_file)
data = data.shuffle(seed=42) # reshuffle in case users provide unshuffled data
# define output directory path
current_date = datetime.datetime.now()
datestamp = f"{str(current_date.year)[-2:]}{current_date.month:02d}{current_date.day:02d}"
if output_directory[-1:] != "/": # add slash for dir if not present
output_directory = output_directory + "/"
output_dir = f"{output_directory}{datestamp}_geneformer_{self.classifier}Classifier_{output_prefix}/"
subprocess.call(f"mkdir {output_dir}", shell=True)
# get number of classes for classifier
num_classes = cu.get_num_classes(id_class_dict)
if self.classifier == "gene":
targets = pu.flatten_list(self.gene_class_dict.values())
labels = pu.flatten_list(
[
[class_id_dict[label]] * len(targets)
for label, targets in self.gene_class_dict.items()
]
)
assert len(targets) == len(labels)
data = cu.prep_gene_classifier_all_data(
data, targets, labels, self.max_ncells, self.nproc
)
trainer = self.train_classifier(
model_directory, num_classes, data, None, output_dir
)
return trainer
def validate(
self,
model_directory,
prepared_input_data_file,
id_class_dict_file,
output_directory,
output_prefix,
split_id_dict=None,
attr_to_split=None,
attr_to_balance=None,
max_trials=100,
pval_threshold=0.1,
save_eval_output=True,
predict_eval=True,
predict_trainer=False,
):
"""
(Cross-)validate cell state or gene classifier.
**Parameters**
model_directory : Path
| Path to directory containing model
prepared_input_data_file : Path
| Path to directory containing _labeled.dataset previously prepared by Classifier.prepare_data
id_class_dict_file : Path
| Path to _id_class_dict.pkl previously prepared by Classifier.prepare_data
| (dictionary of format: numerical IDs: class_labels)
output_directory : Path
| Path to directory where model and eval data will be saved
output_prefix : str
| Prefix for output files
split_id_dict : None, dict
| Dictionary of IDs for train and eval splits
| Three-item dictionary with keys: attr_key, train, eval
| attr_key: key specifying name of column in .dataset that contains the IDs for the data splits
| train: list of IDs in the attr_key column to include in the train split
| eval: list of IDs in the attr_key column to include in the eval split
| For example: {"attr_key": "individual",
| "train": ["patient1", "patient2", "patient3", "patient4"],
| "eval": ["patient5", "patient6"]}
| Note: only available for CellClassifiers with 1-fold split (self.classifier="cell"; self.num_crossval_splits=1)
attr_to_split : None, str
| Key for attribute on which to split data while balancing potential confounders
| e.g. "patient_id" for splitting by patient while balancing other characteristics
| Note: only available for CellClassifiers with 1-fold split (self.classifier="cell"; self.num_crossval_splits=1)
attr_to_balance : None, list
| List of attribute keys on which to balance data while splitting on attr_to_split
| e.g. ["age", "sex"] for balancing these characteristics while splitting by patient
max_trials : None, int
| Maximum number of trials of random splitting to try to achieve balanced other attribute
| If no split is found without significant (p < pval_threshold) differences in other attributes, will select best
pval_threshold : None, float
| P-value threshold to use for attribute balancing across splits
| E.g. if set to 0.1, will accept trial if p >= 0.1 for all attributes in attr_to_balance
save_eval_output : bool
| Whether to save cross-fold eval output
| Saves as pickle file of dictionary of eval metrics
predict_eval : bool
| Whether or not to save eval predictions
| Saves as a pickle file of self.evaluate predictions
predict_trainer : bool
| Whether or not to save eval predictions from trainer
| Saves as a pickle file of trainer predictions
"""
if self.num_crossval_splits == 0:
logger.error("num_crossval_splits must be 1 or 5 to validate.")
raise
# ensure number of genes in each class is > 5 if validating model
if self.classifier == "gene":
insuff_classes = [k for k, v in self.gene_class_dict.items() if len(v) < 5]
if (self.num_crossval_splits > 0) and (len(insuff_classes) > 0):
logger.error(
f"Insufficient # of members in class(es) {insuff_classes} to (cross-)validate."
)
raise
##### Load data and prepare output directory #####
# load numerical id to class dictionary (id:class)
with open(id_class_dict_file, "rb") as f:
id_class_dict = pickle.load(f)
class_id_dict = {v: k for k, v in id_class_dict.items()}
# load previously filtered and prepared data
data = pu.load_and_filter(None, self.nproc, prepared_input_data_file)
data = data.shuffle(seed=42) # reshuffle in case users provide unshuffled data
# define output directory path
current_date = datetime.datetime.now()
datestamp = f"{str(current_date.year)[-2:]}{current_date.month:02d}{current_date.day:02d}"
if output_directory[-1:] != "/": # add slash for dir if not present
output_directory = output_directory + "/"
output_dir = f"{output_directory}{datestamp}_geneformer_{self.classifier}Classifier_{output_prefix}/"
subprocess.call(f"mkdir {output_dir}", shell=True)
# get number of classes for classifier
num_classes = cu.get_num_classes(id_class_dict)
##### (Cross-)validate the model #####
results = []
all_conf_mat = np.zeros((num_classes, num_classes))
iteration_num = 1
if self.classifier == "cell":
for i in trange(self.num_crossval_splits):
print(
f"****** Validation split: {iteration_num}/{self.num_crossval_splits} ******\n"
)
ksplit_output_dir = os.path.join(output_dir, f"ksplit{iteration_num}")
if self.num_crossval_splits == 1:
# single 1-eval_size:eval_size split
if split_id_dict is not None:
data_dict = dict()
data_dict["train"] = pu.filter_by_dict(
data,
{split_id_dict["attr_key"]: split_id_dict["train"]},
self.nproc,
)
data_dict["test"] = pu.filter_by_dict(
data,
{split_id_dict["attr_key"]: split_id_dict["eval"]},
self.nproc,
)
elif attr_to_split is not None:
data_dict, balance_df = cu.balance_attr_splits(
data,
attr_to_split,
attr_to_balance,
self.eval_size,
max_trials,
pval_threshold,
self.cell_state_dict["state_key"],
self.nproc,
)
balance_df.to_csv(
f"{output_dir}/{output_prefix}_train_valid_balance_df.csv"
)
else:
data_dict = data.train_test_split(
test_size=self.eval_size,
stratify_by_column=self.stratify_splits_col,
seed=42,
)
train_data = data_dict["train"]
eval_data = data_dict["test"]
else:
# 5-fold cross-validate
num_cells = len(data)
fifth_cells = num_cells * 0.2
num_eval = min((self.eval_size * num_cells), fifth_cells)
start = i * fifth_cells
end = start + num_eval
eval_indices = [j for j in range(start, end)]
train_indices = [
j for j in range(num_cells) if j not in eval_indices
]
eval_data = data.select(eval_indices)
train_data = data.select(train_indices)
trainer = self.train_classifier(
model_directory,
num_classes,
train_data,
eval_data,
ksplit_output_dir,
predict_trainer,
)
result = self.evaluate_model(
trainer.model,
num_classes,
id_class_dict,
eval_data,
predict_eval,
ksplit_output_dir,
output_prefix,
)
results += [result]
all_conf_mat = all_conf_mat + result["conf_mat"]
iteration_num = iteration_num + 1
elif self.classifier == "gene":
# set up (cross-)validation splits
targets = pu.flatten_list(self.gene_class_dict.values())
labels = pu.flatten_list(
[
[class_id_dict[label]] * len(targets)
for label, targets in self.gene_class_dict.items()
]
)
assert len(targets) == len(labels)
n_splits = int(1 / self.eval_size)
skf = StratifiedKFold(n_splits=n_splits, random_state=0, shuffle=True)
# (Cross-)validate
for train_index, eval_index in tqdm(skf.split(targets, labels)):
print(
f"****** Validation split: {iteration_num}/{self.num_crossval_splits} ******\n"
)
ksplit_output_dir = os.path.join(output_dir, f"ksplit{iteration_num}")
# filter data for examples containing classes for this split
# subsample to max_ncells and relabel data in column "labels"
train_data, eval_data = cu.prep_gene_classifier_split(
data,
targets,
labels,
train_index,
eval_index,
self.max_ncells,
iteration_num,
self.nproc,
)
trainer = self.train_classifier(
model_directory,
num_classes,
train_data,
eval_data,
ksplit_output_dir,
predict_trainer,
)
result = self.evaluate_model(
trainer.model,
num_classes,
id_class_dict,
eval_data,
predict_eval,
ksplit_output_dir,
output_prefix,
)
results += [result]
all_conf_mat = all_conf_mat + result["conf_mat"]
# break after 1 or 5 splits, each with train/eval proportions dictated by eval_size
if iteration_num == self.num_crossval_splits:
break
iteration_num = iteration_num + 1
all_conf_mat_df = pd.DataFrame(
all_conf_mat, columns=id_class_dict.values(), index=id_class_dict.values()
)
all_metrics = {
"conf_matrix": all_conf_mat_df,
"macro_f1": [result["macro_f1"] for result in results],
"acc": [result["acc"] for result in results],
}
all_roc_metrics = None # roc metrics not reported for multiclass
if num_classes == 2:
mean_fpr = np.linspace(0, 1, 100)
all_tpr = [result["roc_metrics"]["interp_tpr"] for result in results]
all_roc_auc = [result["roc_metrics"]["auc"] for result in results]
all_tpr_wt = [result["roc_metrics"]["tpr_wt"] for result in results]
mean_tpr, roc_auc, roc_auc_sd = eu.get_cross_valid_roc_metrics(
all_tpr, all_roc_auc, all_tpr_wt
)
all_roc_metrics = {
"mean_tpr": mean_tpr,
"mean_fpr": mean_fpr,
"all_roc_auc": all_roc_auc,
"roc_auc": roc_auc,
"roc_auc_sd": roc_auc_sd,
}
all_metrics["all_roc_metrics"] = all_roc_metrics
if save_eval_output is True:
eval_metrics_output_path = (
Path(output_dir) / f"{output_prefix}_eval_metrics_dict"
).with_suffix(".pkl")
with open(eval_metrics_output_path, "wb") as f:
pickle.dump(all_metrics, f)
return all_metrics
def train_classifier(
self,
model_directory,
num_classes,
train_data,
eval_data,
output_directory,
predict=False,
):
"""
Fine-tune model for cell state or gene classification.
**Parameters**
model_directory : Path
| Path to directory containing model
num_classes : int
| Number of classes for classifier
train_data : Dataset
| Loaded training .dataset input
| For cell classifier, labels in column "label".
| For gene classifier, labels in column "labels".
eval_data : None, Dataset
| (Optional) Loaded evaluation .dataset input
| For cell classifier, labels in column "label".
| For gene classifier, labels in column "labels".
output_directory : Path
| Path to directory where fine-tuned model will be saved
predict : bool
| Whether or not to save eval predictions from trainer
"""
##### Validate and prepare data #####
train_data, eval_data = cu.validate_and_clean_cols(
train_data, eval_data, self.classifier
)
if (self.no_eval is True) and (eval_data is not None):
logger.warning(
"no_eval set to True; model will be trained without evaluation."
)
eval_data = None
if (self.classifier == "gene") and (predict is True):
logger.warning(
"Predictions during training not currently available for gene classifiers; setting predict to False."
)
predict = False
# ensure not overwriting previously saved model
saved_model_test = os.path.join(output_directory, "pytorch_model.bin")
if os.path.isfile(saved_model_test) is True:
logger.error("Model already saved to this designated output directory.")
raise
# make output directory
subprocess.call(f"mkdir {output_directory}", shell=True)
##### Load model and training args #####
if self.classifier == "cell":
model_type = "CellClassifier"
elif self.classifier == "gene":
model_type = "GeneClassifier"
model = pu.load_model(model_type, num_classes, model_directory, "train")
def_training_args, def_freeze_layers = cu.get_default_train_args(
model, self.classifier, train_data, output_directory
)
if self.training_args is not None:
def_training_args.update(self.training_args)
logging_steps = round(
len(train_data) / def_training_args["per_device_train_batch_size"] / 10
)
def_training_args["logging_steps"] = logging_steps
def_training_args["output_dir"] = output_directory
if eval_data is None:
def_training_args["evaluation_strategy"] = "no"
def_training_args["load_best_model_at_end"] = False
training_args_init = TrainingArguments(**def_training_args)
if self.freeze_layers is not None:
def_freeze_layers = self.freeze_layers
if def_freeze_layers > 0:
modules_to_freeze = model.bert.encoder.layer[:def_freeze_layers]
for module in modules_to_freeze:
for param in module.parameters():
param.requires_grad = False
##### Fine-tune the model #####
# define the data collator
if self.classifier == "cell":
data_collator = DataCollatorForCellClassification()
elif self.classifier == "gene":
data_collator = DataCollatorForGeneClassification()
# create the trainer
trainer = Trainer(
model=model,
args=training_args_init,
data_collator=data_collator,
train_dataset=train_data,
eval_dataset=eval_data,
compute_metrics=cu.compute_metrics,
)
# train the classifier
trainer.train()
trainer.save_model(output_directory)
if predict is True:
# make eval predictions and save predictions and metrics
predictions = trainer.predict(eval_data)
prediction_output_path = f"{output_directory}/predictions.pkl"
with open(prediction_output_path, "wb") as f:
pickle.dump(predictions, f)
trainer.save_metrics("eval", predictions.metrics)
return trainer
def evaluate_model(
self,
model,
num_classes,
id_class_dict,
eval_data,
predict=False,
output_directory=None,
output_prefix=None,
):
"""
Evaluate the fine-tuned model.
**Parameters**
model : nn.Module
| Loaded fine-tuned model (e.g. trainer.model)
num_classes : int
| Number of classes for classifier
id_class_dict : dict
| Loaded _id_class_dict.pkl previously prepared by Classifier.prepare_data
| (dictionary of format: numerical IDs: class_labels)
eval_data : Dataset
| Loaded evaluation .dataset input
predict : bool
| Whether or not to save eval predictions
output_directory : Path
| Path to directory where eval data will be saved
output_prefix : str
| Prefix for output files
"""
##### Evaluate the model #####
labels = id_class_dict.keys()
y_pred, y_true, logits_list = eu.classifier_predict(
model, self.classifier, eval_data, self.forward_batch_size
)
conf_mat, macro_f1, acc, roc_metrics = eu.get_metrics(
y_pred, y_true, logits_list, num_classes, labels
)
if predict is True:
pred_dict = {
"pred_ids": y_pred,
"label_ids": y_true,
"predictions": logits_list,
}
pred_dict_output_path = (
Path(output_directory) / f"{output_prefix}_pred_dict"
).with_suffix(".pkl")
with open(pred_dict_output_path, "wb") as f:
pickle.dump(pred_dict, f)
return {
"conf_mat": conf_mat,
"macro_f1": macro_f1,
"acc": acc,
"roc_metrics": roc_metrics,
}
def evaluate_saved_model(
self,
model_directory,
id_class_dict_file,
test_data_file,
output_directory,
output_prefix,
predict=True,
):
"""
Evaluate the fine-tuned model.
**Parameters**
model_directory : Path
| Path to directory containing model
id_class_dict_file : Path
| Path to _id_class_dict.pkl previously prepared by Classifier.prepare_data
| (dictionary of format: numerical IDs: class_labels)
test_data_file : Path
| Path to directory containing test .dataset
output_directory : Path
| Path to directory where eval data will be saved
output_prefix : str
| Prefix for output files
predict : bool
| Whether or not to save eval predictions
"""
# load numerical id to class dictionary (id:class)
with open(id_class_dict_file, "rb") as f:
id_class_dict = pickle.load(f)
# get number of classes for classifier
num_classes = cu.get_num_classes(id_class_dict)
# load previously filtered and prepared data
test_data = pu.load_and_filter(None, self.nproc, test_data_file)
# load previously fine-tuned model
if self.classifier == "cell":
model_type = "CellClassifier"
elif self.classifier == "gene":
model_type = "GeneClassifier"
model = pu.load_model(model_type, num_classes, model_directory, "eval")
# evaluate the model
result = self.evaluate_model(
model,
num_classes,
id_class_dict,
test_data,
predict=predict,
output_directory=output_directory,
output_prefix=output_prefix,
)
all_conf_mat_df = pd.DataFrame(
result["conf_mat"],
columns=id_class_dict.values(),
index=id_class_dict.values(),
)
all_metrics = {
"conf_matrix": all_conf_mat_df,
"macro_f1": result["macro_f1"],
"acc": result["acc"],
}
all_roc_metrics = None # roc metrics not reported for multiclass
if num_classes == 2:
mean_fpr = np.linspace(0, 1, 100)
mean_tpr = result["roc_metrics"]["interp_tpr"]
all_roc_auc = result["roc_metrics"]["auc"]
all_roc_metrics = {
"mean_tpr": mean_tpr,
"mean_fpr": mean_fpr,
"all_roc_auc": all_roc_auc,
}
all_metrics["all_roc_metrics"] = all_roc_metrics
test_metrics_output_path = (
Path(output_directory) / f"{output_prefix}_test_metrics_dict"
).with_suffix(".pkl")
with open(test_metrics_output_path, "wb") as f:
pickle.dump(all_metrics, f)
return all_metrics
def plot_conf_mat(
self,
conf_mat_dict,
output_directory,
output_prefix,
custom_class_order=None,
):
"""
Plot confusion matrix results of evaluating the fine-tuned model.
**Parameters**
conf_mat_dict : dict
| Dictionary of model_name : confusion_matrix_DataFrame
| (all_metrics["conf_matrix"] from self.validate)
output_directory : Path
| Path to directory where plots will be saved
output_prefix : str
| Prefix for output file
custom_class_order : None, list
| List of classes in custom order for plots.
| Same order will be used for all models.
"""
for model_name in conf_mat_dict.keys():
eu.plot_confusion_matrix(
conf_mat_dict[model_name],
model_name,
output_directory,
output_prefix,
custom_class_order,
)
def plot_roc(
self,
roc_metric_dict,
model_style_dict,
title,
output_directory,
output_prefix,
):
"""
Plot ROC curve results of evaluating the fine-tuned model.
**Parameters**
roc_metric_dict : dict
| Dictionary of model_name : roc_metrics
| (all_metrics["all_roc_metrics"] from self.validate)
model_style_dict : dict[dict]
| Dictionary of model_name : dictionary of style_attribute : style
| where style includes color and linestyle
| e.g. {'Model_A': {'color': 'black', 'linestyle': '-'}, 'Model_B': ...}
title : str
| Title of plot (e.g. 'Dosage-sensitive vs -insensitive factors')
output_directory : Path
| Path to directory where plots will be saved
output_prefix : str
| Prefix for output file
"""
eu.plot_ROC(
roc_metric_dict, model_style_dict, title, output_directory, output_prefix
)
def plot_predictions(
self,
predictions_file,
id_class_dict_file,
title,
output_directory,
output_prefix,
custom_class_order=None,
kwargs_dict=None,
):
"""
Plot prediction results of evaluating the fine-tuned model.
**Parameters**
predictions_file : path
| Path of model predictions output to plot
| (saved output from self.validate if predict_eval=True)
| (or saved output from self.evaluate_saved_model)
id_class_dict_file : Path
| Path to _id_class_dict.pkl previously prepared by Classifier.prepare_data
| (dictionary of format: numerical IDs: class_labels)
title : str
| Title for legend containing class labels.
output_directory : Path
| Path to directory where plots will be saved
output_prefix : str
| Prefix for output file
custom_class_order : None, list
| List of classes in custom order for plots.
| Same order will be used for all models.
kwargs_dict : None, dict
| Dictionary of kwargs to pass to plotting function.
"""
# load predictions
with open(predictions_file, "rb") as f:
predictions = pickle.load(f)
# load numerical id to class dictionary (id:class)
with open(id_class_dict_file, "rb") as f:
id_class_dict = pickle.load(f)
if isinstance(predictions, dict):
if all(
[
key in predictions.keys()
for key in ["pred_ids", "label_ids", "predictions"]
]
):
# format is output from self.evaluate_saved_model
predictions_logits = np.array(predictions["predictions"])
true_ids = predictions["label_ids"]
else:
# format is output from self.validate if predict_eval=True
predictions_logits = predictions.predictions
true_ids = predictions.label_ids
num_classes = len(id_class_dict.keys())
num_predict_classes = predictions_logits.shape[1]
assert num_classes == num_predict_classes
classes = id_class_dict.values()
true_labels = [id_class_dict[idx] for idx in true_ids]
predictions_df = pd.DataFrame(predictions_logits, columns=classes)
if custom_class_order is not None:
predictions_df = predictions_df.reindex(columns=custom_class_order)
predictions_df["true"] = true_labels
custom_dict = dict(zip(classes, [i for i in range(len(classes))]))
if custom_class_order is not None:
custom_dict = dict(
zip(custom_class_order, [i for i in range(len(custom_class_order))])
)
predictions_df = predictions_df.sort_values(
by=["true"], key=lambda x: x.map(custom_dict)
)
eu.plot_predictions(
predictions_df, title, output_directory, output_prefix, kwargs_dict
)
|