AmelieSchreiber
commited on
Commit
•
8551e6c
1
Parent(s):
4d0ba55
Upload 3 files
Browse files- lora_train.py +195 -0
- metrics_2.py +95 -0
- qlora_train.py +245 -0
lora_train.py
ADDED
@@ -0,0 +1,195 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import wandb
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
from datetime import datetime
|
7 |
+
from sklearn.model_selection import train_test_split
|
8 |
+
from sklearn.utils.class_weight import compute_class_weight
|
9 |
+
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, roc_auc_score, matthews_corrcoef
|
10 |
+
from transformers import AutoModelForTokenClassification, AutoTokenizer, DataCollatorForTokenClassification, TrainingArguments, Trainer
|
11 |
+
from datasets import Dataset
|
12 |
+
from accelerate import Accelerator
|
13 |
+
from peft import get_peft_config, PeftModel, PeftConfig, get_peft_model, LoraConfig, TaskType
|
14 |
+
import pickle
|
15 |
+
|
16 |
+
# Initialize accelerator and Weights & Biases
|
17 |
+
accelerator = Accelerator()
|
18 |
+
os.environ["WANDB_NOTEBOOK_NAME"] = 'train.py'
|
19 |
+
wandb.init(project='binding_site_prediction')
|
20 |
+
|
21 |
+
# Helper Functions and Data Preparation
|
22 |
+
def save_config_to_txt(config, filename):
|
23 |
+
"""Save the configuration dictionary to a text file."""
|
24 |
+
with open(filename, 'w') as f:
|
25 |
+
for key, value in config.items():
|
26 |
+
f.write(f"{key}: {value}\n")
|
27 |
+
|
28 |
+
def truncate_labels(labels, max_length):
|
29 |
+
return [label[:max_length] for label in labels]
|
30 |
+
|
31 |
+
def compute_metrics(p):
|
32 |
+
predictions, labels = p
|
33 |
+
predictions = np.argmax(predictions, axis=2)
|
34 |
+
predictions = predictions[labels != -100].flatten()
|
35 |
+
labels = labels[labels != -100].flatten()
|
36 |
+
accuracy = accuracy_score(labels, predictions)
|
37 |
+
precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average='binary')
|
38 |
+
auc = roc_auc_score(labels, predictions)
|
39 |
+
mcc = matthews_corrcoef(labels, predictions)
|
40 |
+
return {'accuracy': accuracy, 'precision': precision, 'recall': recall, 'f1': f1, 'auc': auc, 'mcc': mcc}
|
41 |
+
|
42 |
+
def compute_loss(model, logits, inputs):
|
43 |
+
# logits = model(**inputs).logits
|
44 |
+
labels = inputs["labels"]
|
45 |
+
loss_fct = nn.CrossEntropyLoss(weight=class_weights)
|
46 |
+
active_loss = inputs["attention_mask"].view(-1) == 1
|
47 |
+
active_logits = logits.view(-1, model.config.num_labels)
|
48 |
+
active_labels = torch.where(
|
49 |
+
active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
|
50 |
+
)
|
51 |
+
loss = loss_fct(active_logits, active_labels)
|
52 |
+
return loss
|
53 |
+
|
54 |
+
# Load data from pickle files
|
55 |
+
with open("770K_data/train_sequences_chunked_by_family.pkl", "rb") as f:
|
56 |
+
train_sequences = pickle.load(f)
|
57 |
+
|
58 |
+
with open("770K_data/test_sequences_chunked_by_family.pkl", "rb") as f:
|
59 |
+
test_sequences = pickle.load(f)
|
60 |
+
|
61 |
+
with open("770K_data/train_labels_chunked_by_family.pkl", "rb") as f:
|
62 |
+
train_labels = pickle.load(f)
|
63 |
+
|
64 |
+
with open("770K_data/test_labels_chunked_by_family.pkl", "rb") as f:
|
65 |
+
test_labels = pickle.load(f)
|
66 |
+
|
67 |
+
# Tokenization
|
68 |
+
tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t12_35M_UR50D")
|
69 |
+
|
70 |
+
# Set max_sequence_length to the tokenizer's max input length
|
71 |
+
max_sequence_length = tokenizer.model_max_length
|
72 |
+
|
73 |
+
train_tokenized = tokenizer(train_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
74 |
+
test_tokenized = tokenizer(test_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
75 |
+
|
76 |
+
# Directly truncate the entire list of labels
|
77 |
+
train_labels = truncate_labels(train_labels, max_sequence_length)
|
78 |
+
test_labels = truncate_labels(test_labels, max_sequence_length)
|
79 |
+
|
80 |
+
train_dataset = Dataset.from_dict({k: v for k, v in train_tokenized.items()}).add_column("labels", train_labels)
|
81 |
+
test_dataset = Dataset.from_dict({k: v for k, v in test_tokenized.items()}).add_column("labels", test_labels)
|
82 |
+
|
83 |
+
# Compute Class Weights
|
84 |
+
classes = [0, 1]
|
85 |
+
flat_train_labels = [label for sublist in train_labels for label in sublist]
|
86 |
+
class_weights = compute_class_weight(class_weight='balanced', classes=classes, y=flat_train_labels)
|
87 |
+
class_weights = torch.tensor(class_weights, dtype=torch.float32).to(accelerator.device)
|
88 |
+
|
89 |
+
# Define Custom Trainer Class
|
90 |
+
class WeightedTrainer(Trainer):
|
91 |
+
def compute_loss(self, model, inputs, return_outputs=False):
|
92 |
+
outputs = model(**inputs)
|
93 |
+
logits = outputs.logits
|
94 |
+
loss = compute_loss(model, logits, inputs)
|
95 |
+
return (loss, outputs) if return_outputs else loss
|
96 |
+
|
97 |
+
# Define and run training function
|
98 |
+
def train_function_no_sweeps(train_dataset, test_dataset):
|
99 |
+
|
100 |
+
# Directly set the config
|
101 |
+
config = {
|
102 |
+
"lora_alpha": 1,
|
103 |
+
"lora_dropout": 0.5,
|
104 |
+
"lr": 3.701568055793089e-04,
|
105 |
+
"lr_scheduler_type": "cosine_with_restarts",
|
106 |
+
"max_grad_norm": 0.5,
|
107 |
+
"num_train_epochs": 3,
|
108 |
+
"per_device_train_batch_size": 6,
|
109 |
+
"r": 2,
|
110 |
+
"weight_decay": 0.2,
|
111 |
+
# Add other hyperparameters as needed
|
112 |
+
}
|
113 |
+
|
114 |
+
# Log the config to W&B
|
115 |
+
wandb.config.update(config)
|
116 |
+
|
117 |
+
# Save the config to a text file
|
118 |
+
timestamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
|
119 |
+
config_filename = f"esm2_t12_35M_lora_config_{timestamp}.txt"
|
120 |
+
save_config_to_txt(config, config_filename)
|
121 |
+
|
122 |
+
model_checkpoint = "facebook/esm2_t12_35M_UR50D"
|
123 |
+
|
124 |
+
# Define labels and model
|
125 |
+
id2label = {0: "No binding site", 1: "Binding site"}
|
126 |
+
label2id = {v: k for k, v in id2label.items()}
|
127 |
+
model = AutoModelForTokenClassification.from_pretrained(model_checkpoint, num_labels=len(id2label), id2label=id2label, label2id=label2id)
|
128 |
+
|
129 |
+
# Convert the model into a PeftModel
|
130 |
+
peft_config = LoraConfig(
|
131 |
+
task_type=TaskType.TOKEN_CLS,
|
132 |
+
inference_mode=False,
|
133 |
+
r=config["r"],
|
134 |
+
lora_alpha=config["lora_alpha"],
|
135 |
+
target_modules=["query", "key", "value"],
|
136 |
+
lora_dropout=config["lora_dropout"],
|
137 |
+
bias="none", # or "all" or "lora_only"
|
138 |
+
modules_to_save=["classifier"]
|
139 |
+
)
|
140 |
+
model = get_peft_model(model, peft_config)
|
141 |
+
|
142 |
+
# Use the accelerator
|
143 |
+
model = accelerator.prepare(model)
|
144 |
+
train_dataset = accelerator.prepare(train_dataset)
|
145 |
+
test_dataset = accelerator.prepare(test_dataset)
|
146 |
+
|
147 |
+
timestamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
|
148 |
+
|
149 |
+
# Training setup
|
150 |
+
training_args = TrainingArguments(
|
151 |
+
output_dir=f"esm2_t12_35M_lora_binding_sites_{timestamp}",
|
152 |
+
learning_rate=config["lr"],
|
153 |
+
lr_scheduler_type=config["lr_scheduler_type"],
|
154 |
+
gradient_accumulation_steps=1,
|
155 |
+
max_grad_norm=config["max_grad_norm"],
|
156 |
+
per_device_train_batch_size=config["per_device_train_batch_size"],
|
157 |
+
per_device_eval_batch_size=config["per_device_train_batch_size"],
|
158 |
+
num_train_epochs=config["num_train_epochs"],
|
159 |
+
weight_decay=config["weight_decay"],
|
160 |
+
evaluation_strategy="epoch",
|
161 |
+
save_strategy="epoch",
|
162 |
+
load_best_model_at_end=True,
|
163 |
+
metric_for_best_model="f1",
|
164 |
+
greater_is_better=True,
|
165 |
+
push_to_hub=False,
|
166 |
+
logging_dir=None,
|
167 |
+
logging_first_step=False,
|
168 |
+
logging_steps=200,
|
169 |
+
save_total_limit=7,
|
170 |
+
no_cuda=False,
|
171 |
+
seed=8893,
|
172 |
+
fp16=True,
|
173 |
+
report_to='wandb'
|
174 |
+
)
|
175 |
+
|
176 |
+
# Initialize Trainer
|
177 |
+
trainer = WeightedTrainer(
|
178 |
+
model=model,
|
179 |
+
args=training_args,
|
180 |
+
train_dataset=train_dataset,
|
181 |
+
eval_dataset=test_dataset,
|
182 |
+
tokenizer=tokenizer,
|
183 |
+
data_collator=DataCollatorForTokenClassification(tokenizer=tokenizer),
|
184 |
+
compute_metrics=compute_metrics
|
185 |
+
)
|
186 |
+
|
187 |
+
# Train and Save Model
|
188 |
+
trainer.train()
|
189 |
+
save_path = os.path.join("lora_binding_sites", f"best_model_esm2_t12_35M_lora_{timestamp}")
|
190 |
+
trainer.save_model(save_path)
|
191 |
+
tokenizer.save_pretrained(save_path)
|
192 |
+
|
193 |
+
# Call the training function
|
194 |
+
if __name__ == "__main__":
|
195 |
+
train_function_no_sweeps(train_dataset, test_dataset)
|
metrics_2.py
ADDED
@@ -0,0 +1,95 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import wandb
|
3 |
+
import numpy as np
|
4 |
+
import pickle
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, roc_auc_score, matthews_corrcoef
|
8 |
+
from transformers import AutoModelForTokenClassification, AutoTokenizer, DataCollatorForTokenClassification, Trainer
|
9 |
+
from datasets import Dataset
|
10 |
+
from accelerate import Accelerator
|
11 |
+
from peft import PeftModel
|
12 |
+
|
13 |
+
# Helper functions and data preparation
|
14 |
+
def truncate_labels(labels, max_length):
|
15 |
+
"""Truncate labels to the specified max_length."""
|
16 |
+
return [label[:max_length] for label in labels]
|
17 |
+
|
18 |
+
def compute_metrics(p):
|
19 |
+
"""Compute metrics for evaluation."""
|
20 |
+
predictions, labels = p
|
21 |
+
predictions = np.argmax(predictions, axis=2)
|
22 |
+
|
23 |
+
# Remove padding (-100 labels)
|
24 |
+
predictions = predictions[labels != -100].flatten()
|
25 |
+
labels = labels[labels != -100].flatten()
|
26 |
+
|
27 |
+
# Compute accuracy
|
28 |
+
accuracy = accuracy_score(labels, predictions)
|
29 |
+
|
30 |
+
# Compute precision, recall, F1 score, and AUC
|
31 |
+
precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average='binary')
|
32 |
+
auc = roc_auc_score(labels, predictions)
|
33 |
+
|
34 |
+
# Compute MCC
|
35 |
+
mcc = matthews_corrcoef(labels, predictions)
|
36 |
+
|
37 |
+
return {'accuracy': accuracy, 'precision': precision, 'recall': recall, 'f1': f1, 'auc': auc, 'mcc': mcc}
|
38 |
+
|
39 |
+
class WeightedTrainer(Trainer):
|
40 |
+
def compute_loss(self, model, inputs, return_outputs=False):
|
41 |
+
"""Custom compute_loss function."""
|
42 |
+
outputs = model(**inputs)
|
43 |
+
loss_fct = nn.CrossEntropyLoss()
|
44 |
+
active_loss = inputs["attention_mask"].view(-1) == 1
|
45 |
+
active_logits = outputs.logits.view(-1, model.config.num_labels)
|
46 |
+
active_labels = torch.where(
|
47 |
+
active_loss, inputs["labels"].view(-1), torch.tensor(loss_fct.ignore_index).type_as(inputs["labels"])
|
48 |
+
)
|
49 |
+
loss = loss_fct(active_logits, active_labels)
|
50 |
+
return (loss, outputs) if return_outputs else loss
|
51 |
+
|
52 |
+
if __name__ == "__main__":
|
53 |
+
# Environment setup
|
54 |
+
accelerator = Accelerator()
|
55 |
+
wandb.init(project='binding_site_prediction')
|
56 |
+
|
57 |
+
# Load data and labels
|
58 |
+
with open("1111K_data/train_sequences_chunked_by_family.pkl", "rb") as f:
|
59 |
+
train_sequences = pickle.load(f)
|
60 |
+
with open("1111K_data/test_sequences_chunked_by_family.pkl", "rb") as f:
|
61 |
+
test_sequences = pickle.load(f)
|
62 |
+
with open("1111K_data/train_labels_chunked_by_family.pkl", "rb") as f:
|
63 |
+
train_labels = pickle.load(f)
|
64 |
+
with open("1111K_data/test_labels_chunked_by_family.pkl", "rb") as f:
|
65 |
+
test_labels = pickle.load(f)
|
66 |
+
|
67 |
+
# Tokenization and dataset creation
|
68 |
+
tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t12_35M_UR50D")
|
69 |
+
max_sequence_length = tokenizer.model_max_length
|
70 |
+
train_tokenized = tokenizer(train_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
71 |
+
test_tokenized = tokenizer(test_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
72 |
+
train_labels = truncate_labels(train_labels, max_sequence_length)
|
73 |
+
test_labels = truncate_labels(test_labels, max_sequence_length)
|
74 |
+
train_dataset = Dataset.from_dict({k: v for k, v in train_tokenized.items()}).add_column("labels", train_labels)
|
75 |
+
test_dataset = Dataset.from_dict({k: v for k, v in test_tokenized.items()}).add_column("labels", test_labels)
|
76 |
+
|
77 |
+
# Load the pre-trained LoRA model
|
78 |
+
base_model_path = "facebook/esm2_t12_35M_UR50D"
|
79 |
+
lora_model_path = "esm2_t12_35M_lora_binding_sites_2023-09-23_03-04-43/checkpoint-102604" # Replace with the correct path to your LoRA model
|
80 |
+
base_model = AutoModelForTokenClassification.from_pretrained(base_model_path)
|
81 |
+
model = PeftModel.from_pretrained(base_model, lora_model_path)
|
82 |
+
model = accelerator.prepare(model)
|
83 |
+
|
84 |
+
# Define a function to compute metrics and get the train/test metrics
|
85 |
+
data_collator = DataCollatorForTokenClassification(tokenizer)
|
86 |
+
trainer = Trainer(model=model, data_collator=data_collator, compute_metrics=compute_metrics)
|
87 |
+
train_metrics = trainer.evaluate(train_dataset)
|
88 |
+
test_metrics = trainer.evaluate(test_dataset)
|
89 |
+
|
90 |
+
# Print the metrics
|
91 |
+
print(f"Train metrics: {train_metrics}")
|
92 |
+
print(f"Test metrics: {test_metrics}")
|
93 |
+
|
94 |
+
# Log metrics to W&B
|
95 |
+
wandb.log({"Train metrics": train_metrics, "Test metrics": test_metrics})
|
qlora_train.py
ADDED
@@ -0,0 +1,245 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import wandb
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
from datetime import datetime
|
7 |
+
from sklearn.model_selection import train_test_split
|
8 |
+
from sklearn.utils.class_weight import compute_class_weight
|
9 |
+
from sklearn.metrics import accuracy_score, precision_recall_fscore_support, roc_auc_score, matthews_corrcoef
|
10 |
+
from transformers import (
|
11 |
+
AutoModelForTokenClassification,
|
12 |
+
AutoTokenizer,
|
13 |
+
DataCollatorForTokenClassification,
|
14 |
+
TrainingArguments,
|
15 |
+
Trainer,
|
16 |
+
BitsAndBytesConfig
|
17 |
+
)
|
18 |
+
from datasets import Dataset
|
19 |
+
from accelerate import Accelerator
|
20 |
+
from peft import get_peft_config, PeftModel, PeftConfig, get_peft_model, LoraConfig, TaskType, prepare_model_for_kbit_training
|
21 |
+
import pickle
|
22 |
+
|
23 |
+
# Initialize accelerator and Weights & Biases
|
24 |
+
accelerator = Accelerator()
|
25 |
+
os.environ["WANDB_NOTEBOOK_NAME"] = 'train.py'
|
26 |
+
wandb.init(project='binding_site_prediction')
|
27 |
+
|
28 |
+
# Helper Functions and Data Preparation
|
29 |
+
#-----------------------------------------------------------------------------
|
30 |
+
# Added this first function in
|
31 |
+
def print_trainable_parameters(model):
|
32 |
+
"""
|
33 |
+
Prints the number of trainable parameters in the model.
|
34 |
+
"""
|
35 |
+
trainable_params = 0
|
36 |
+
all_param = 0
|
37 |
+
for _, param in model.named_parameters():
|
38 |
+
all_param += param.numel()
|
39 |
+
if param.requires_grad:
|
40 |
+
trainable_params += param.numel()
|
41 |
+
print(
|
42 |
+
f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
|
43 |
+
)
|
44 |
+
#-----------------------------------------------------------------------------
|
45 |
+
|
46 |
+
def save_config_to_txt(config, filename):
|
47 |
+
"""Save the configuration dictionary to a text file."""
|
48 |
+
with open(filename, 'w') as f:
|
49 |
+
for key, value in config.items():
|
50 |
+
f.write(f"{key}: {value}\n")
|
51 |
+
|
52 |
+
def truncate_labels(labels, max_length):
|
53 |
+
return [label[:max_length] for label in labels]
|
54 |
+
|
55 |
+
def compute_metrics(p):
|
56 |
+
predictions, labels = p
|
57 |
+
predictions = np.argmax(predictions, axis=2)
|
58 |
+
predictions = predictions[labels != -100].flatten()
|
59 |
+
labels = labels[labels != -100].flatten()
|
60 |
+
accuracy = accuracy_score(labels, predictions)
|
61 |
+
precision, recall, f1, _ = precision_recall_fscore_support(labels, predictions, average='binary')
|
62 |
+
auc = roc_auc_score(labels, predictions)
|
63 |
+
mcc = matthews_corrcoef(labels, predictions)
|
64 |
+
return {'accuracy': accuracy, 'precision': precision, 'recall': recall, 'f1': f1, 'auc': auc, 'mcc': mcc}
|
65 |
+
|
66 |
+
def compute_loss(model, logits, inputs):
|
67 |
+
# logits = model(**inputs).logits
|
68 |
+
labels = inputs["labels"]
|
69 |
+
loss_fct = nn.CrossEntropyLoss(weight=class_weights)
|
70 |
+
active_loss = inputs["attention_mask"].view(-1) == 1
|
71 |
+
active_logits = logits.view(-1, model.config.num_labels)
|
72 |
+
active_labels = torch.where(
|
73 |
+
active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
|
74 |
+
)
|
75 |
+
loss = loss_fct(active_logits, active_labels)
|
76 |
+
return loss
|
77 |
+
|
78 |
+
|
79 |
+
# Load data from pickle files
|
80 |
+
with open("data/600K_data/train_sequences_chunked_by_family.pkl", "rb") as f:
|
81 |
+
train_sequences = pickle.load(f)
|
82 |
+
|
83 |
+
with open("data/600K_data/test_sequences_chunked_by_family.pkl", "rb") as f:
|
84 |
+
test_sequences = pickle.load(f)
|
85 |
+
|
86 |
+
with open("data/600K_data/train_labels_chunked_by_family.pkl", "rb") as f:
|
87 |
+
train_labels = pickle.load(f)
|
88 |
+
|
89 |
+
with open("data/600K_data/test_labels_chunked_by_family.pkl", "rb") as f:
|
90 |
+
test_labels = pickle.load(f)
|
91 |
+
|
92 |
+
# Tokenization
|
93 |
+
tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t6_8M_UR50D")
|
94 |
+
|
95 |
+
# Set max_sequence_length to the tokenizer's max input length
|
96 |
+
max_sequence_length = tokenizer.model_max_length
|
97 |
+
|
98 |
+
train_tokenized = tokenizer(train_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
99 |
+
test_tokenized = tokenizer(test_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
100 |
+
|
101 |
+
# Directly truncate the entire list of labels
|
102 |
+
train_labels = truncate_labels(train_labels, max_sequence_length)
|
103 |
+
test_labels = truncate_labels(test_labels, max_sequence_length)
|
104 |
+
|
105 |
+
train_dataset = Dataset.from_dict({k: v for k, v in train_tokenized.items()}).add_column("labels", train_labels)
|
106 |
+
test_dataset = Dataset.from_dict({k: v for k, v in test_tokenized.items()}).add_column("labels", test_labels)
|
107 |
+
|
108 |
+
# Compute Class Weights
|
109 |
+
classes = [0, 1]
|
110 |
+
flat_train_labels = [label for sublist in train_labels for label in sublist]
|
111 |
+
class_weights = compute_class_weight(class_weight='balanced', classes=classes, y=flat_train_labels)
|
112 |
+
class_weights = torch.tensor(class_weights, dtype=torch.float32).to(accelerator.device)
|
113 |
+
|
114 |
+
# Define Custom Trainer Class
|
115 |
+
class WeightedTrainer(Trainer):
|
116 |
+
def compute_loss(self, model, inputs, return_outputs=False):
|
117 |
+
outputs = model(**inputs)
|
118 |
+
logits = outputs.logits
|
119 |
+
loss = compute_loss(model, logits, inputs)
|
120 |
+
return (loss, outputs) if return_outputs else loss
|
121 |
+
|
122 |
+
# Configure the quantization settings
|
123 |
+
bnb_config = BitsAndBytesConfig(
|
124 |
+
load_in_4bit=True,
|
125 |
+
bnb_4bit_use_double_quant=True,
|
126 |
+
bnb_4bit_quant_type="nf4",
|
127 |
+
bnb_4bit_compute_dtype=torch.bfloat16
|
128 |
+
)
|
129 |
+
|
130 |
+
def train_function_no_sweeps(train_dataset, test_dataset):
|
131 |
+
|
132 |
+
|
133 |
+
# Directly set the config
|
134 |
+
config = {
|
135 |
+
"lora_alpha": 1,
|
136 |
+
"lora_dropout": 0.5,
|
137 |
+
"lr": 3.701568055793089e-04,
|
138 |
+
"lr_scheduler_type": "cosine",
|
139 |
+
"max_grad_norm": 0.5,
|
140 |
+
"num_train_epochs": 4,
|
141 |
+
"per_device_train_batch_size": 64,
|
142 |
+
"r": 2,
|
143 |
+
"weight_decay": 0.2,
|
144 |
+
# Add other hyperparameters as needed
|
145 |
+
}
|
146 |
+
|
147 |
+
# Log the config to W&B
|
148 |
+
wandb.config.update(config)
|
149 |
+
|
150 |
+
# Save the config to a text file
|
151 |
+
timestamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
|
152 |
+
config_filename = f"esm2_t6_8M_qlora_config_{timestamp}.txt"
|
153 |
+
save_config_to_txt(config, config_filename)
|
154 |
+
|
155 |
+
|
156 |
+
model_checkpoint = "facebook/esm2_t6_8M_UR50D"
|
157 |
+
|
158 |
+
# Define labels and model
|
159 |
+
id2label = {0: "No binding site", 1: "Binding site"}
|
160 |
+
label2id = {v: k for k, v in id2label.items()}
|
161 |
+
|
162 |
+
model = AutoModelForTokenClassification.from_pretrained(
|
163 |
+
model_checkpoint,
|
164 |
+
num_labels=len(id2label),
|
165 |
+
id2label=id2label,
|
166 |
+
label2id=label2id,
|
167 |
+
quantization_config=bnb_config # Apply quantization here
|
168 |
+
)
|
169 |
+
|
170 |
+
# Prepare the model for 4-bit quantization training
|
171 |
+
model.gradient_checkpointing_enable()
|
172 |
+
model = prepare_model_for_kbit_training(model)
|
173 |
+
|
174 |
+
# Convert the model into a PeftModel
|
175 |
+
peft_config = LoraConfig(
|
176 |
+
task_type=TaskType.TOKEN_CLS,
|
177 |
+
inference_mode=False,
|
178 |
+
r=config["r"],
|
179 |
+
lora_alpha=config["lora_alpha"],
|
180 |
+
target_modules=["query", "key", "value"],
|
181 |
+
lora_dropout=config["lora_dropout"],
|
182 |
+
bias="none", # or "all" or "lora_only"
|
183 |
+
modules_to_save=["classifier"]
|
184 |
+
)
|
185 |
+
model = get_peft_model(model, peft_config)
|
186 |
+
print_trainable_parameters(model) # added this in
|
187 |
+
|
188 |
+
# Use the accelerator
|
189 |
+
model = accelerator.prepare(model)
|
190 |
+
train_dataset = accelerator.prepare(train_dataset)
|
191 |
+
test_dataset = accelerator.prepare(test_dataset)
|
192 |
+
|
193 |
+
timestamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
|
194 |
+
|
195 |
+
# Training setup
|
196 |
+
training_args = TrainingArguments(
|
197 |
+
output_dir=f"esm2_t6_8M_qlora_binding_sites_{timestamp}",
|
198 |
+
learning_rate=config["lr"],
|
199 |
+
lr_scheduler_type=config["lr_scheduler_type"],
|
200 |
+
gradient_accumulation_steps=2, # changed from 1 to 4
|
201 |
+
# warmup_steps=2, # added this in
|
202 |
+
max_grad_norm=config["max_grad_norm"],
|
203 |
+
per_device_train_batch_size=config["per_device_train_batch_size"],
|
204 |
+
per_device_eval_batch_size=config["per_device_train_batch_size"],
|
205 |
+
num_train_epochs=config["num_train_epochs"],
|
206 |
+
weight_decay=config["weight_decay"],
|
207 |
+
evaluation_strategy="epoch",
|
208 |
+
save_strategy="epoch",
|
209 |
+
load_best_model_at_end=True,
|
210 |
+
metric_for_best_model="f1",
|
211 |
+
greater_is_better=True,
|
212 |
+
push_to_hub=False,
|
213 |
+
logging_dir=None,
|
214 |
+
logging_first_step=False,
|
215 |
+
logging_steps=200,
|
216 |
+
save_total_limit=7,
|
217 |
+
no_cuda=False,
|
218 |
+
seed=8893,
|
219 |
+
fp16=True,
|
220 |
+
report_to='wandb',
|
221 |
+
optim="paged_adamw_8bit" # added this in
|
222 |
+
|
223 |
+
)
|
224 |
+
|
225 |
+
# Initialize Trainer
|
226 |
+
trainer = WeightedTrainer(
|
227 |
+
model=model,
|
228 |
+
args=training_args,
|
229 |
+
train_dataset=train_dataset,
|
230 |
+
eval_dataset=test_dataset,
|
231 |
+
tokenizer=tokenizer,
|
232 |
+
data_collator=DataCollatorForTokenClassification(tokenizer=tokenizer),
|
233 |
+
compute_metrics=compute_metrics
|
234 |
+
)
|
235 |
+
|
236 |
+
# Train and Save Model
|
237 |
+
trainer.train()
|
238 |
+
save_path = os.path.join("qlora_binding_sites", f"best_model_esm2_t6_8M_qlora_{timestamp}")
|
239 |
+
trainer.save_model(save_path)
|
240 |
+
tokenizer.save_pretrained(save_path)
|
241 |
+
|
242 |
+
# Call the training function
|
243 |
+
if __name__ == "__main__":
|
244 |
+
train_function_no_sweeps(train_dataset, test_dataset)
|
245 |
+
|