Commit
·
97a01f0
1
Parent(s):
9d54d52
Upload 2 files
Browse files- finetune.py +170 -0
- metrics (2).py +96 -0
finetune.py
ADDED
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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 precision_recall_fscore_support, roc_auc_score, accuracy_score, matthews_corrcoef
|
10 |
+
from transformers import (
|
11 |
+
AutoModelForTokenClassification,
|
12 |
+
AutoTokenizer,
|
13 |
+
DataCollatorForTokenClassification,
|
14 |
+
TrainingArguments,
|
15 |
+
Trainer
|
16 |
+
)
|
17 |
+
from datasets import Dataset
|
18 |
+
from accelerate import Accelerator
|
19 |
+
import pickle
|
20 |
+
|
21 |
+
# Initialize Weights & Biases logging
|
22 |
+
os.environ["WANDB_NOTEBOOK_NAME"] = 'esm2_t6_8M_finetune_600K.ipynb'
|
23 |
+
wandb.init(project='binding_site_prediction')
|
24 |
+
|
25 |
+
# Helper Functions
|
26 |
+
def truncate_labels(labels, max_length):
|
27 |
+
"""Truncate labels to the specified max_length."""
|
28 |
+
return [label[:max_length] for label in labels]
|
29 |
+
|
30 |
+
def compute_metrics(p):
|
31 |
+
"""Compute metrics for evaluation."""
|
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, inputs):
|
43 |
+
"""Custom compute_loss function."""
|
44 |
+
logits = model(**inputs).logits
|
45 |
+
labels = inputs["labels"]
|
46 |
+
loss_fct = nn.CrossEntropyLoss(weight=class_weights)
|
47 |
+
active_loss = inputs["attention_mask"].view(-1) == 1
|
48 |
+
active_logits = logits.view(-1, model.config.num_labels)
|
49 |
+
active_labels = torch.where(
|
50 |
+
active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
|
51 |
+
)
|
52 |
+
loss = loss_fct(active_logits, active_labels)
|
53 |
+
return loss
|
54 |
+
|
55 |
+
# Custom Trainer Class
|
56 |
+
class WeightedTrainer(Trainer):
|
57 |
+
def compute_loss(self, model, inputs, return_outputs=False):
|
58 |
+
outputs = model(**inputs)
|
59 |
+
loss = compute_loss(model, inputs)
|
60 |
+
return (loss, outputs) if return_outputs else loss
|
61 |
+
|
62 |
+
# Load data
|
63 |
+
with open("600K_data/train_sequences_chunked_by_family.pkl", "rb") as f:
|
64 |
+
train_sequences = pickle.load(f)
|
65 |
+
|
66 |
+
with open("600K_data/test_sequences_chunked_by_family.pkl", "rb") as f:
|
67 |
+
test_sequences = pickle.load(f)
|
68 |
+
|
69 |
+
with open("600K_data/train_labels_chunked_by_family.pkl", "rb") as f:
|
70 |
+
train_labels = pickle.load(f)
|
71 |
+
|
72 |
+
with open("600K_data/test_labels_chunked_by_family.pkl", "rb") as f:
|
73 |
+
test_labels = pickle.load(f)
|
74 |
+
|
75 |
+
# Tokenization
|
76 |
+
tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t6_8M_UR50D")
|
77 |
+
max_sequence_length = 1000
|
78 |
+
train_tokenized = tokenizer(train_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
79 |
+
test_tokenized = tokenizer(test_sequences, padding=True, truncation=True, max_length=max_sequence_length, return_tensors="pt", is_split_into_words=False)
|
80 |
+
train_labels = truncate_labels(train_labels, max_sequence_length)
|
81 |
+
test_labels = truncate_labels(test_labels, max_sequence_length)
|
82 |
+
train_dataset = Dataset.from_dict({k: v for k, v in train_tokenized.items()}).add_column("labels", train_labels)
|
83 |
+
test_dataset = Dataset.from_dict({k: v for k, v in test_tokenized.items()}).add_column("labels", test_labels)
|
84 |
+
|
85 |
+
# Compute Class Weights
|
86 |
+
classes = [0, 1]
|
87 |
+
flat_train_labels = [label for sublist in train_labels for label in sublist]
|
88 |
+
class_weights = compute_class_weight(class_weight='balanced', classes=classes, y=flat_train_labels)
|
89 |
+
accelerator = Accelerator()
|
90 |
+
class_weights = torch.tensor(class_weights, dtype=torch.float32).to(accelerator.device)
|
91 |
+
|
92 |
+
# Training Function
|
93 |
+
def train_function_no_sweeps(train_dataset, test_dataset):
|
94 |
+
# Initialize wandb
|
95 |
+
wandb.init()
|
96 |
+
|
97 |
+
# Configurations
|
98 |
+
config = {
|
99 |
+
"lr": 5.701568055793089e-04,
|
100 |
+
"lr_scheduler_type": "cosine",
|
101 |
+
"max_grad_norm": 0.5,
|
102 |
+
"num_train_epochs": 1,
|
103 |
+
"per_device_train_batch_size": 12,
|
104 |
+
"weight_decay": 0.2
|
105 |
+
}
|
106 |
+
|
107 |
+
# Model Setup
|
108 |
+
model_checkpoint = "facebook/esm2_t6_8M_UR50D"
|
109 |
+
id2label = {0: "No binding site", 1: "Binding site"}
|
110 |
+
label2id = {v: k for k, v in id2label.items()}
|
111 |
+
model = AutoModelForTokenClassification.from_pretrained(
|
112 |
+
model_checkpoint,
|
113 |
+
num_labels=len(id2label),
|
114 |
+
id2label=id2label,
|
115 |
+
label2id=label2id,
|
116 |
+
hidden_dropout_prob=0.5, # Add this line for hidden dropout
|
117 |
+
attention_probs_dropout_prob=0.5 # Add this line for attention dropout
|
118 |
+
)
|
119 |
+
model = accelerator.prepare(model)
|
120 |
+
train_dataset = accelerator.prepare(train_dataset)
|
121 |
+
test_dataset = accelerator.prepare(test_dataset)
|
122 |
+
timestamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
|
123 |
+
|
124 |
+
# Training setup
|
125 |
+
training_args = TrainingArguments(
|
126 |
+
output_dir=f"esm2_t6_8M_finetune_{timestamp}",
|
127 |
+
learning_rate=config["lr"],
|
128 |
+
lr_scheduler_type=config["lr_scheduler_type"],
|
129 |
+
gradient_accumulation_steps=1,
|
130 |
+
max_grad_norm=config["max_grad_norm"],
|
131 |
+
per_device_train_batch_size=config["per_device_train_batch_size"],
|
132 |
+
per_device_eval_batch_size=config["per_device_train_batch_size"],
|
133 |
+
num_train_epochs=config["num_train_epochs"],
|
134 |
+
weight_decay=config["weight_decay"],
|
135 |
+
evaluation_strategy="epoch",
|
136 |
+
save_strategy="epoch",
|
137 |
+
load_best_model_at_end=True,
|
138 |
+
metric_for_best_model="f1",
|
139 |
+
greater_is_better=True,
|
140 |
+
push_to_hub=False,
|
141 |
+
logging_dir=None,
|
142 |
+
logging_first_step=False,
|
143 |
+
logging_steps=200,
|
144 |
+
save_total_limit=7,
|
145 |
+
no_cuda=False,
|
146 |
+
seed=42,
|
147 |
+
fp16=True,
|
148 |
+
report_to='wandb'
|
149 |
+
)
|
150 |
+
|
151 |
+
# Initialize Trainer
|
152 |
+
trainer = WeightedTrainer(
|
153 |
+
model=model,
|
154 |
+
args=training_args,
|
155 |
+
train_dataset=train_dataset,
|
156 |
+
eval_dataset=test_dataset,
|
157 |
+
tokenizer=tokenizer,
|
158 |
+
data_collator=DataCollatorForTokenClassification(tokenizer=tokenizer),
|
159 |
+
compute_metrics=compute_metrics
|
160 |
+
)
|
161 |
+
|
162 |
+
# Train and Save Model
|
163 |
+
trainer.train()
|
164 |
+
save_path = os.path.join("binding_sites", f"best_model_esm2_t6_8M_{timestamp}")
|
165 |
+
trainer.save_model(save_path)
|
166 |
+
tokenizer.save_pretrained(save_path)
|
167 |
+
|
168 |
+
# Call the training function
|
169 |
+
if __name__ == "__main__":
|
170 |
+
train_function_no_sweeps(train_dataset, test_dataset)
|
metrics (2).py
ADDED
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
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("600K_data/train_sequences_chunked_by_family.pkl", "rb") as f:
|
59 |
+
train_sequences = pickle.load(f)
|
60 |
+
with open("600K_data/test_sequences_chunked_by_family.pkl", "rb") as f:
|
61 |
+
test_sequences = pickle.load(f)
|
62 |
+
with open("600K_data/train_labels_chunked_by_family.pkl", "rb") as f:
|
63 |
+
train_labels = pickle.load(f)
|
64 |
+
with open("600K_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 = "esm2_t6_8M_finetune_2023-10-08_00-58-24/checkpoint-42015"
|
79 |
+
# lora_model_path = "AmelieSchreiber/esm2_t12_35M_qlora_binding_2600K_cp1" # Replace with the correct path to your LoRA model
|
80 |
+
# base_model = AutoModelForTokenClassification.from_pretrained(base_model_path) # use this for LoRA
|
81 |
+
model = AutoModelForTokenClassification.from_pretrained(base_model_path) # remove this for LoRA
|
82 |
+
# model = PeftModel.from_pretrained(base_model, lora_model_path) # use this for LoRA
|
83 |
+
model = accelerator.prepare(model)
|
84 |
+
|
85 |
+
# Define a function to compute metrics and get the train/test metrics
|
86 |
+
data_collator = DataCollatorForTokenClassification(tokenizer)
|
87 |
+
trainer = Trainer(model=model, data_collator=data_collator, compute_metrics=compute_metrics)
|
88 |
+
train_metrics = trainer.evaluate(train_dataset)
|
89 |
+
test_metrics = trainer.evaluate(test_dataset)
|
90 |
+
|
91 |
+
# Print the metrics
|
92 |
+
print(f"Train metrics: {train_metrics}")
|
93 |
+
print(f"Test metrics: {test_metrics}")
|
94 |
+
|
95 |
+
# Log metrics to W&B
|
96 |
+
# wandb.log({"Train metrics": train_metrics, "Test metrics": test_metrics})
|