Christina Theodoris
commited on
Commit
•
79a0c41
1
Parent(s):
b2aee1b
Add example for hyperparameter optimization for disease classifier
Browse files
examples/hyperparam_optimiz_for_disease_classifier.py
ADDED
@@ -0,0 +1,193 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
# coding: utf-8
|
3 |
+
|
4 |
+
# hyperparameter optimization with raytune for disease classification
|
5 |
+
|
6 |
+
# imports
|
7 |
+
import os
|
8 |
+
import subprocess
|
9 |
+
GPU_NUMBER = [0,1,2,3]
|
10 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = ",".join([str(s) for s in GPU_NUMBER])
|
11 |
+
os.environ["NCCL_DEBUG"] = "INFO"
|
12 |
+
os.environ["CONDA_OVERRIDE_GLIBC"] = "2.56"
|
13 |
+
os.environ["LD_LIBRARY_PATH"] = "/path/to/miniconda3/lib:/path/to/sw/lib:/path/to/sw/lib"
|
14 |
+
|
15 |
+
# initiate runtime environment for raytune
|
16 |
+
import pyarrow # must occur prior to ray import
|
17 |
+
import ray
|
18 |
+
from ray import tune
|
19 |
+
from ray.tune import ExperimentAnalysis
|
20 |
+
from ray.tune.suggest.hyperopt import HyperOptSearch
|
21 |
+
runtime_env = {"conda": "base",
|
22 |
+
"env_vars": {"LD_LIBRARY_PATH": "/path/to/miniconda3/lib:/path/to/sw/lib:/path/to/sw/lib"}}
|
23 |
+
ray.init(runtime_env=runtime_env)
|
24 |
+
|
25 |
+
import datetime
|
26 |
+
import numpy as np
|
27 |
+
import pandas as pd
|
28 |
+
import random
|
29 |
+
import seaborn as sns; sns.set()
|
30 |
+
from collections import Counter
|
31 |
+
from datasets import load_from_disk
|
32 |
+
from scipy.stats import ranksums
|
33 |
+
from sklearn.metrics import accuracy_score
|
34 |
+
from transformers import BertForSequenceClassification
|
35 |
+
from transformers import Trainer
|
36 |
+
from transformers.training_args import TrainingArguments
|
37 |
+
|
38 |
+
from geneformer import DataCollatorForCellClassification
|
39 |
+
|
40 |
+
# number of CPU cores
|
41 |
+
num_proc=30
|
42 |
+
|
43 |
+
# load train dataset with columns:
|
44 |
+
# cell_type (annotation of each cell's type)
|
45 |
+
# disease (healthy or disease state)
|
46 |
+
# individual (unique ID for each patient)
|
47 |
+
# length (length of that cell's rank value encoding)
|
48 |
+
train_dataset=load_from_disk("/path/to/disease_train_data.dataset")
|
49 |
+
|
50 |
+
# filter dataset for given cell_type
|
51 |
+
def if_cell_type(example):
|
52 |
+
return example["cell_type"].startswith("Cardiomyocyte")
|
53 |
+
|
54 |
+
trainset_v2 = train_dataset.filter(if_cell_type, num_proc=num_proc)
|
55 |
+
|
56 |
+
# create dictionary of disease states : label ids
|
57 |
+
target_names = ["healthy", "disease1", "disease2"]
|
58 |
+
target_name_id_dict = dict(zip(target_names,[i for i in range(len(target_names))]))
|
59 |
+
|
60 |
+
trainset_v3 = trainset_v2.rename_column("disease","label")
|
61 |
+
|
62 |
+
# change labels to numerical ids
|
63 |
+
def classes_to_ids(example):
|
64 |
+
example["label"] = target_name_id_dict[example["label"]]
|
65 |
+
return example
|
66 |
+
|
67 |
+
trainset_v4 = trainset_v3.map(classes_to_ids, num_proc=num_proc)
|
68 |
+
|
69 |
+
# separate into train, validation, test sets
|
70 |
+
indiv_list = trainset_v4["individual"]
|
71 |
+
random.seed(42)
|
72 |
+
train_indiv = random.sample(indiv_list,round(0.7*len(indiv_list)))
|
73 |
+
eval_indiv = [indiv for indiv in indiv_list if indiv not in train_indiv]
|
74 |
+
valid_indiv = random.sample(eval_indiv,round(0.5*len(eval_indiv)))
|
75 |
+
test_indiv = [indiv for indiv in eval_indiv if indiv not in valid_indiv]
|
76 |
+
|
77 |
+
def if_train(example):
|
78 |
+
return example["individual"] in train_indiv
|
79 |
+
|
80 |
+
classifier_trainset = trainset_v4.filter(if_train,num_proc=num_proc).shuffle(seed=42)
|
81 |
+
|
82 |
+
def if_valid(example):
|
83 |
+
return example["individual"] in valid_indiv
|
84 |
+
|
85 |
+
classifier_validset = trainset_v4.filter(if_valid,num_proc=num_proc).shuffle(seed=42)
|
86 |
+
|
87 |
+
# define output directory path
|
88 |
+
current_date = datetime.datetime.now()
|
89 |
+
datestamp = f"{str(current_date.year)[-2:]}{current_date.month:02d}{current_date.day:02d}"
|
90 |
+
output_dir = f"/path/to/models/{datestamp}_geneformer_DiseaseClassifier/"
|
91 |
+
|
92 |
+
# ensure not overwriting previously saved model
|
93 |
+
saved_model_test = os.path.join(output_dir, f"pytorch_model.bin")
|
94 |
+
if os.path.isfile(saved_model_test) == True:
|
95 |
+
raise Exception("Model already saved to this directory.")
|
96 |
+
|
97 |
+
# make output directory
|
98 |
+
subprocess.call(f'mkdir {output_dir}', shell=True)
|
99 |
+
|
100 |
+
# set training parameters
|
101 |
+
# how many pretrained layers to freeze
|
102 |
+
freeze_layers = 2
|
103 |
+
# batch size for training and eval
|
104 |
+
geneformer_batch_size = 12
|
105 |
+
# number of epochs
|
106 |
+
epochs = 1
|
107 |
+
# logging steps
|
108 |
+
logging_steps = round(len(classifier_trainset)/geneformer_batch_size/10)
|
109 |
+
|
110 |
+
# define function to initiate model
|
111 |
+
def model_init():
|
112 |
+
model = BertForSequenceClassification.from_pretrained("/path/to/pretrained_model/",
|
113 |
+
num_labels=len(target_names),
|
114 |
+
output_attentions = False,
|
115 |
+
output_hidden_states = False)
|
116 |
+
if freeze_layers is not None:
|
117 |
+
modules_to_freeze = model.bert.encoder.layer[:freeze_layers]
|
118 |
+
for module in modules_to_freeze:
|
119 |
+
for param in module.parameters():
|
120 |
+
param.requires_grad = False
|
121 |
+
|
122 |
+
model = model.to("cuda:0")
|
123 |
+
return model
|
124 |
+
|
125 |
+
# define metrics
|
126 |
+
def compute_metrics(pred):
|
127 |
+
labels = pred.label_ids
|
128 |
+
preds = pred.predictions.argmax(-1)
|
129 |
+
# calculate accuracy using sklearn's function
|
130 |
+
acc = accuracy_score(labels, preds)
|
131 |
+
return {
|
132 |
+
'accuracy': acc,
|
133 |
+
}
|
134 |
+
|
135 |
+
# set training arguments
|
136 |
+
training_args = {
|
137 |
+
"do_train": True,
|
138 |
+
"do_eval": True,
|
139 |
+
"evaluation_strategy": "steps",
|
140 |
+
"eval_steps": logging_steps,
|
141 |
+
"logging_steps": logging_steps,
|
142 |
+
"group_by_length": True,
|
143 |
+
"length_column_name": "length",
|
144 |
+
"disable_tqdm": True,
|
145 |
+
"skip_memory_metrics": True, # memory tracker causes errors in raytune
|
146 |
+
"per_device_train_batch_size": geneformer_batch_size,
|
147 |
+
"per_device_eval_batch_size": geneformer_batch_size,
|
148 |
+
"num_train_epochs": epochs,
|
149 |
+
"load_best_model_at_end": True,
|
150 |
+
"output_dir": output_dir,
|
151 |
+
}
|
152 |
+
|
153 |
+
training_args_init = TrainingArguments(**training_args)
|
154 |
+
|
155 |
+
# create the trainer
|
156 |
+
trainer = Trainer(
|
157 |
+
model_init=model_init,
|
158 |
+
args=training_args_init,
|
159 |
+
data_collator=DataCollatorForCellClassification(),
|
160 |
+
train_dataset=classifier_trainset,
|
161 |
+
eval_dataset=classifier_validset,
|
162 |
+
compute_metrics=compute_metrics,
|
163 |
+
)
|
164 |
+
|
165 |
+
# specify raytune hyperparameter search space
|
166 |
+
ray_config = {
|
167 |
+
"num_train_epochs": tune.choice([epochs]),
|
168 |
+
"learning_rate": tune.loguniform(1e-6, 1e-3),
|
169 |
+
"weight_decay": tune.uniform(0.0, 0.3),
|
170 |
+
"lr_scheduler_type": tune.choice(["linear","cosine","polynomial"]),
|
171 |
+
"warmup_steps": tune.uniform(100, 2000),
|
172 |
+
"seed": tune.uniform(0,100),
|
173 |
+
"per_device_train_batch_size": tune.choice([geneformer_batch_size])
|
174 |
+
}
|
175 |
+
|
176 |
+
hyperopt_search = HyperOptSearch(
|
177 |
+
metric="eval_accuracy", mode="max")
|
178 |
+
|
179 |
+
# optimize hyperparameters
|
180 |
+
trainer.hyperparameter_search(
|
181 |
+
direction="maximize",
|
182 |
+
backend="ray",
|
183 |
+
resources_per_trial={"cpu":8,"gpu":1},
|
184 |
+
hp_space=lambda _: ray_config,
|
185 |
+
search_alg=hyperopt_search,
|
186 |
+
n_trials=100, # number of trials
|
187 |
+
progress_reporter=tune.CLIReporter(max_report_frequency=600,
|
188 |
+
sort_by_metric=True,
|
189 |
+
max_progress_rows=100,
|
190 |
+
mode="max",
|
191 |
+
metric="eval_accuracy",
|
192 |
+
metric_columns=["loss", "eval_loss", "eval_accuracy"])
|
193 |
+
)
|