delete first version of dist multitask example
Browse files
examples/distrbuted_multitask_cell_classification.ipynb
DELETED
|
@@ -1,149 +0,0 @@
|
|
| 1 |
-
{
|
| 2 |
-
"cells": [
|
| 3 |
-
{
|
| 4 |
-
"cell_type": "code",
|
| 5 |
-
"execution_count": null,
|
| 6 |
-
"id": "b3266a7b",
|
| 7 |
-
"metadata": {},
|
| 8 |
-
"outputs": [],
|
| 9 |
-
"source": [
|
| 10 |
-
"import os\n",
|
| 11 |
-
"import torch\n",
|
| 12 |
-
"from geneformer import MTLClassifier"
|
| 13 |
-
]
|
| 14 |
-
},
|
| 15 |
-
{
|
| 16 |
-
"cell_type": "code",
|
| 17 |
-
"execution_count": null,
|
| 18 |
-
"id": "3e12ac9f",
|
| 19 |
-
"metadata": {},
|
| 20 |
-
"outputs": [],
|
| 21 |
-
"source": [
|
| 22 |
-
"# Define paths\n",
|
| 23 |
-
"pretrained_path = \"/path/to/pretrained/Geneformer/model\" \n",
|
| 24 |
-
"# input data is tokenized rank value encodings generated by Geneformer tokenizer (see tokenizing_scRNAseq_data.ipynb)\n",
|
| 25 |
-
"train_path = \"/path/to/train/data.dataset\"\n",
|
| 26 |
-
"val_path = \"/path/to/val/data.dataset\"\n",
|
| 27 |
-
"test_path = \"/path/to/test/data.dataset\"\n",
|
| 28 |
-
"results_dir = \"/path/to/results/directory\"\n",
|
| 29 |
-
"model_save_path = \"/path/to/model/save/path\"\n",
|
| 30 |
-
"tensorboard_log_dir = \"/path/to/tensorboard/log/dir\"\n",
|
| 31 |
-
"\n",
|
| 32 |
-
"# Define tasks and hyperparameters\n",
|
| 33 |
-
"# task_columns should be a list of column names from your dataset\n",
|
| 34 |
-
"# Each column represents a specific classification task (e.g. cell type, disease state)\n",
|
| 35 |
-
"task_columns = [\"cell_type\", \"disease_state\"] # Example task columns"
|
| 36 |
-
]
|
| 37 |
-
},
|
| 38 |
-
{
|
| 39 |
-
"cell_type": "code",
|
| 40 |
-
"execution_count": null,
|
| 41 |
-
"id": "c9bd7562",
|
| 42 |
-
"metadata": {},
|
| 43 |
-
"outputs": [],
|
| 44 |
-
"source": [
|
| 45 |
-
"# Check GPU environment\n",
|
| 46 |
-
"num_gpus = torch.cuda.device_count()\n",
|
| 47 |
-
"use_distributed = num_gpus > 1\n",
|
| 48 |
-
"print(f\"Number of GPUs detected: {num_gpus}\")\n",
|
| 49 |
-
"print(f\"Using distributed training: {use_distributed}\")\n",
|
| 50 |
-
"\n",
|
| 51 |
-
"# Set environment variables for distributed training when multiple GPUs are available\n",
|
| 52 |
-
"if use_distributed:\n",
|
| 53 |
-
" os.environ[\"MASTER_ADDR\"] = \"localhost\" # hostname\n",
|
| 54 |
-
" os.environ[\"MASTER_PORT\"] = \"12355\" # Choose an available port\n",
|
| 55 |
-
" print(\"Distributed environment variables set.\")"
|
| 56 |
-
]
|
| 57 |
-
},
|
| 58 |
-
{
|
| 59 |
-
"cell_type": "code",
|
| 60 |
-
"execution_count": null,
|
| 61 |
-
"id": "b6ff3618",
|
| 62 |
-
"metadata": {},
|
| 63 |
-
"outputs": [],
|
| 64 |
-
"source": [
|
| 65 |
-
"#Define Hyperparameters for Optimization\n",
|
| 66 |
-
"hyperparameters = {\n",
|
| 67 |
-
" \"learning_rate\": {\"type\": \"float\", \"low\": 1e-5, \"high\": 1e-3, \"log\": True},\n",
|
| 68 |
-
" \"warmup_ratio\": {\"type\": \"float\", \"low\": 0.005, \"high\": 0.01},\n",
|
| 69 |
-
" \"weight_decay\": {\"type\": \"float\", \"low\": 0.01, \"high\": 0.1},\n",
|
| 70 |
-
" \"dropout_rate\": {\"type\": \"float\", \"low\": 0.0, \"high\": 0.7},\n",
|
| 71 |
-
" \"lr_scheduler_type\": {\"type\": \"categorical\", \"choices\": [\"cosine\"]},\n",
|
| 72 |
-
" \"task_weights\": {\"type\": \"float\", \"low\": 0.1, \"high\": 2.0},\n",
|
| 73 |
-
"}"
|
| 74 |
-
]
|
| 75 |
-
},
|
| 76 |
-
{
|
| 77 |
-
"cell_type": "code",
|
| 78 |
-
"execution_count": null,
|
| 79 |
-
"id": "f665c5a7",
|
| 80 |
-
"metadata": {},
|
| 81 |
-
"outputs": [],
|
| 82 |
-
"source": [
|
| 83 |
-
"mc = MTLClassifier(\n",
|
| 84 |
-
" task_columns=task_columns, # Our defined classification tasks\n",
|
| 85 |
-
" study_name=\"MTLClassifier_distributed\",\n",
|
| 86 |
-
" pretrained_path=pretrained_path,\n",
|
| 87 |
-
" train_path=train_path,\n",
|
| 88 |
-
" val_path=val_path,\n",
|
| 89 |
-
" test_path=test_path,\n",
|
| 90 |
-
" model_save_path=model_save_path,\n",
|
| 91 |
-
" results_dir=results_dir,\n",
|
| 92 |
-
" tensorboard_log_dir=tensorboard_log_dir,\n",
|
| 93 |
-
" hyperparameters=hyperparameters,\n",
|
| 94 |
-
" # Distributed training parameters\n",
|
| 95 |
-
" distributed_training=use_distributed, # Enable distributed training if multiple GPUs available\n",
|
| 96 |
-
" master_addr=\"localhost\" if use_distributed else None,\n",
|
| 97 |
-
" master_port=\"12355\" if use_distributed else None,\n",
|
| 98 |
-
" # Other training parameters\n",
|
| 99 |
-
" n_trials=15, # Number of trials for hyperparameter optimization\n",
|
| 100 |
-
" epochs=1, # Number of training epochs (1 suggested to prevent overfitting)\n",
|
| 101 |
-
" batch_size=8, # Adjust based on available GPU memory\n",
|
| 102 |
-
" gradient_accumulation_steps=4, # Accumulate gradients over multiple steps\n",
|
| 103 |
-
" gradient_clipping=True, # Enable gradient clipping for stability\n",
|
| 104 |
-
" max_grad_norm=1.0, # Set maximum gradient norm\n",
|
| 105 |
-
" seed=42\n",
|
| 106 |
-
")"
|
| 107 |
-
]
|
| 108 |
-
},
|
| 109 |
-
{
|
| 110 |
-
"cell_type": "code",
|
| 111 |
-
"execution_count": null,
|
| 112 |
-
"id": "f69f7b6a",
|
| 113 |
-
"metadata": {},
|
| 114 |
-
"outputs": [],
|
| 115 |
-
"source": [
|
| 116 |
-
"# Run Hyperparameter Optimization with Distributed Training\n",
|
| 117 |
-
"if __name__ == \"__main__\":\n",
|
| 118 |
-
" # This guard is required for distributed training to prevent\n",
|
| 119 |
-
" # infinite subprocess spawning when using torch.multiprocessing\n",
|
| 120 |
-
" mc.run_optuna_study()"
|
| 121 |
-
]
|
| 122 |
-
},
|
| 123 |
-
{
|
| 124 |
-
"cell_type": "code",
|
| 125 |
-
"execution_count": null,
|
| 126 |
-
"id": "3affd5dd",
|
| 127 |
-
"metadata": {},
|
| 128 |
-
"outputs": [],
|
| 129 |
-
"source": [
|
| 130 |
-
"# Evaluate the Model on Test Data\n",
|
| 131 |
-
"if __name__ == \"__main__\":\n",
|
| 132 |
-
" mc.load_and_evaluate_test_model()"
|
| 133 |
-
]
|
| 134 |
-
}
|
| 135 |
-
],
|
| 136 |
-
"metadata": {
|
| 137 |
-
"kernelspec": {
|
| 138 |
-
"display_name": "bio",
|
| 139 |
-
"language": "python",
|
| 140 |
-
"name": "python3"
|
| 141 |
-
},
|
| 142 |
-
"language_info": {
|
| 143 |
-
"name": "python",
|
| 144 |
-
"version": "3.12.8"
|
| 145 |
-
}
|
| 146 |
-
},
|
| 147 |
-
"nbformat": 4,
|
| 148 |
-
"nbformat_minor": 5
|
| 149 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|