Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -73,6 +73,93 @@ def compute_loss(model, inputs):
|
|
73 |
loss = loss_fct(active_logits, active_labels)
|
74 |
return loss
|
75 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
76 |
# Load the data from pickle files (replace with your local paths)
|
77 |
with open("./datasets/train_sequences_chunked_by_family.pkl", "rb") as f:
|
78 |
train_sequences = pickle.load(f)
|
|
|
73 |
loss = loss_fct(active_logits, active_labels)
|
74 |
return loss
|
75 |
|
76 |
+
# fine-tuning function
|
77 |
+
def train_function_no_sweeps(base_model_path, train_dataset, test_dataset):
|
78 |
+
|
79 |
+
# Set the LoRA config
|
80 |
+
config = {
|
81 |
+
"lora_alpha": 1, #try 0.5, 1, 2, ..., 16
|
82 |
+
"lora_dropout": 0.2,
|
83 |
+
"lr": 5.701568055793089e-04,
|
84 |
+
"lr_scheduler_type": "cosine",
|
85 |
+
"max_grad_norm": 0.5,
|
86 |
+
"num_train_epochs": 3,
|
87 |
+
"per_device_train_batch_size": 12,
|
88 |
+
"r": 2,
|
89 |
+
"weight_decay": 0.2,
|
90 |
+
# Add other hyperparameters as needed
|
91 |
+
}
|
92 |
+
# The base model you will train a LoRA on top of
|
93 |
+
base_model_path = "facebook/esm2_t12_35M_UR50D"
|
94 |
+
|
95 |
+
# Define labels and model
|
96 |
+
id2label = {0: "No binding site", 1: "Binding site"}
|
97 |
+
label2id = {v: k for k, v in id2label.items()}
|
98 |
+
base_model = AutoModelForTokenClassification.from_pretrained(base_model_path, num_labels=len(id2label), id2label=id2label, label2id=label2id)
|
99 |
+
|
100 |
+
# Convert the model into a PeftModel
|
101 |
+
peft_config = LoraConfig(
|
102 |
+
task_type=TaskType.TOKEN_CLS,
|
103 |
+
inference_mode=False,
|
104 |
+
r=config["r"],
|
105 |
+
lora_alpha=config["lora_alpha"],
|
106 |
+
target_modules=["query", "key", "value"], # also try "dense_h_to_4h" and "dense_4h_to_h"
|
107 |
+
lora_dropout=config["lora_dropout"],
|
108 |
+
bias="none" # or "all" or "lora_only"
|
109 |
+
)
|
110 |
+
base_model = get_peft_model(base_model, peft_config)
|
111 |
+
|
112 |
+
# Use the accelerator
|
113 |
+
base_model = accelerator.prepare(base_model)
|
114 |
+
train_dataset = accelerator.prepare(train_dataset)
|
115 |
+
test_dataset = accelerator.prepare(test_dataset)
|
116 |
+
|
117 |
+
timestamp = datetime.now().strftime('%Y-%m-%d_%H-%M-%S')
|
118 |
+
|
119 |
+
# Training setup
|
120 |
+
training_args = TrainingArguments(
|
121 |
+
output_dir=f"esm2_t12_35M-lora-binding-sites_{timestamp}",
|
122 |
+
learning_rate=config["lr"],
|
123 |
+
lr_scheduler_type=config["lr_scheduler_type"],
|
124 |
+
gradient_accumulation_steps=1,
|
125 |
+
max_grad_norm=config["max_grad_norm"],
|
126 |
+
per_device_train_batch_size=config["per_device_train_batch_size"],
|
127 |
+
per_device_eval_batch_size=config["per_device_train_batch_size"],
|
128 |
+
num_train_epochs=config["num_train_epochs"],
|
129 |
+
weight_decay=config["weight_decay"],
|
130 |
+
evaluation_strategy="epoch",
|
131 |
+
save_strategy="epoch",
|
132 |
+
load_best_model_at_end=True,
|
133 |
+
metric_for_best_model="f1",
|
134 |
+
greater_is_better=True,
|
135 |
+
push_to_hub=False,
|
136 |
+
logging_dir=None,
|
137 |
+
logging_first_step=False,
|
138 |
+
logging_steps=200,
|
139 |
+
save_total_limit=7,
|
140 |
+
no_cuda=False,
|
141 |
+
seed=8893,
|
142 |
+
fp16=True,
|
143 |
+
report_to='wandb'
|
144 |
+
)
|
145 |
+
|
146 |
+
# Initialize Trainer
|
147 |
+
trainer = WeightedTrainer(
|
148 |
+
model=base_model,
|
149 |
+
args=training_args,
|
150 |
+
train_dataset=train_dataset,
|
151 |
+
eval_dataset=test_dataset,
|
152 |
+
tokenizer=tokenizer,
|
153 |
+
data_collator=DataCollatorForTokenClassification(tokenizer=tokenizer),
|
154 |
+
compute_metrics=compute_metrics
|
155 |
+
)
|
156 |
+
|
157 |
+
# Train and Save Model
|
158 |
+
trainer.train()
|
159 |
+
save_path = os.path.join("lora_binding_sites", f"best_model_esm2_t12_35M_lora_{timestamp}")
|
160 |
+
trainer.save_model(save_path)
|
161 |
+
tokenizer.save_pretrained(save_path)
|
162 |
+
|
163 |
# Load the data from pickle files (replace with your local paths)
|
164 |
with open("./datasets/train_sequences_chunked_by_family.pkl", "rb") as f:
|
165 |
train_sequences = pickle.load(f)
|