Merge pull request #31 from boostcampaitech4lv23nlp2/feat/mlflow-tran…

…sformer

Feat/mlflow-transformer

requirements.txt

@@ -1,6 +1,6 @@
 pandas==1.1.5
 scikit-learn~=0.24.1
-transformers==4.10.0
+transformers==4.24.0
 pytorch-lightning==1.7.7
 pyyaml==6.0
 pysftp==0.2.9

src/train.py

@@ -5,14 +5,13 @@
 
 from src.data_loader import REDataset, data_loader
 from src.model import compute_metrics
-from src.utils import get_train_valid_split, label_to_num, save_model_remote, set_mlflow_logger, set_seed
+from src.utils import end_train, get_train_valid_split, label_to_num, save_model_remote, set_mlflow_logger, set_seed
 
 
 def train(model_args, data_args, training_args):
     # Using HfArgumentParser we can turn this class into argparse arguments to be able to specify them on the command line.
     # parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
     # model_args, data_args, training_args = parser.parse_args_into_dataclasses()
-    set_mlflow_logger()
 
     set_seed(data_args.seed)
 
@@ -56,6 +55,12 @@ def train(model_args, data_args, training_args):
     )
 
     # train model
+    special_word = data_args.task_name
+    tracking_uri = ""
+    experiment_name = ""
+    logging_step = 100
+
+    model_id = set_mlflow_logger(special_word, tracking_uri, experiment_name, logging_step)
     trainer.train()
     model.save_pretrained(data_args.best_model_dir_path)
-    save_model_remote(special_word=data_args.task_name)
+    save_model_remote(experiment_name, model_id)

src/utils/__init__.py

@@ -1,5 +1,5 @@
 from .arguments import DataTrainingArguments, ModelArguments, get_training_args
-from .control_mlflow import save_model_remote, set_mlflow_logger
+from .control_mlflow import end_train, save_model_remote, set_mlflow_logger
 from .get_train_valid_split import get_train_valid_split
 from .preprocess import replace_symbol
 from .representation import representation

src/utils/control_mlflow.py

@@ -1,20 +1,27 @@
-import json
+import os
 import pickle as pickle
 import uuid
 
-import mlflow.pytorch
+import mlflow
 import pysftp
 import yaml
+from transformers.integrations import MLflowCallback
 
 
-def set_mlflow_logger(tracking_uri="", experiment_name="", logging_step=0):
+def end_train():
+    mlflow.end_run()
+    print("Train Finished... The model will be saved on remote")
+
+
+def set_mlflow_logger(special_word="", tracking_uri="", experiment_name="", logging_step=0):
     """A function sets mlfow logger environments.
 
-    :param `tracking_uri`: A String Data that informs uri of the mlflow site.
+    Args:
+        tracking_uri (String): A String Data that informs uri of the mlflow site.
                            Usually uses port 5000.
-    :param `experiment_name`: A String Data that informs experiment name at mlflow.
+        experiment_name (String): A String Data that informs experiment name at mlflow.
                               If it doesn't exist at mlflow, it creates one using this name.
-    :param `logging_step`: An Integer Data sets how much steps
+        logging_step (String): An Integer Data sets per how much steps
     """
 
     try:
@@ -47,16 +54,18 @@ def set_mlflow_logger(tracking_uri="", experiment_name="", logging_step=0):
         print("Set Experiment Name:", experiment_name)
         print("Set Logging Step:", logging_step)
 
-        mlflow.set_tracking_uri(tracking_uri)
-        mlflow.set_experiment(experiment_name)
-        mlflow.pytorch.autolog(
-            log_every_n_step=logging_step,
-        )
+        model_id = special_word + "_" + uuid.uuid4().hex
+
+        os.environ["MLFLOW_TRACKING_URI"] = tracking_uri
+        os.environ["MLFLOW_EXPERIMENT_NAME"] = experiment_name
+        os.environ["MLFLOW_TAGS"] = '{"mlflow.runName": "' + model_id + '"}'
+        mlflow.doctor()
     finally:
         print("MLflow setup job finished")
+        return model_id
 
 
-def save_model_remote(experiment_name="", special_word=""):
+def save_model_remote(experiment_name="", model_id=""):
     """A function saves best model on remote storage.
 
     Args:
@@ -65,74 +74,81 @@ def save_model_remote(experiment_name="", special_word=""):
         special_word (String): A String Data that user can customize the name of the model.
                                User can add anything like hyper_parameter setting, user name, etc.
     """
-    with open("src/config/mlflow_config.yml") as f:
-        config_data = yaml.load(f, Loader=yaml.FullLoader)
-        print(config_data)
-    if experiment_name == "":
-        print("No input for experiment_name... import Default")
-        experiment_name = config_data["experiment_name"]
-
-    progressDict = {}
-    progressEveryPercent = 10
-
-    for i in range(0, 101):
-        if i % progressEveryPercent == 0:
-            progressDict[str(i)] = ""
-
-    def printProgressDecimal(x, y):
-        """A callback function for show sftp progress log.
-           Source: https://stackoverflow.com/questions/24278146/how-do-i-monitor-the-progress-of-a-file-transfer-through-pysftp
-
-        Args:
-            x (String): Represent for to-do-data size(e.g. remained file size of pulling of getting)
-            y (String): Represent for total file size
-        """
-        if (
-            int(100 * (int(x) / int(y))) % progressEveryPercent == 0
-            and progressDict[str(int(100 * (int(x) / int(y))))] == ""
-        ):
-            print("{}% ({} Transfered(B)/ {} Total File Size(B))".format(str("%.2f" % (100 * (int(x) / int(y)))), x, y))
-            progressDict[str(int(100 * (int(x) / int(y))))] = "1"
-
-    model_id = uuid.uuid4().hex
-
-    with open("src/config/sftp_config.yml") as f:
-        config_data = yaml.load(f, Loader=yaml.FullLoader)
-    host = config_data["host"]
-    port = config_data["port"]
-    username = config_data["username"]
-    password = config_data["password"]
-
-    cnopts = pysftp.CnOpts()
-    cnopts.hostkeys = None
-
-    mlflow.log_artifact("src/best_model/config.json")
-    with pysftp.Connection(host, port=port, username=username, password=password, cnopts=cnopts) as sftp:
-        print("connected!!")
-        sftp.chdir("./mlflow_models")
-        try:
-            sftp.chdir(experiment_name)
-        except IOError:
-            sftp.mkdir(experiment_name)
-            sftp.chdir(experiment_name)
-        sftp.mkdir(special_word + "_" + model_id)
-        sftp.chdir(special_word + "_" + model_id)
-
-        model_url = "/mlflow_models/" + experiment_name + "/" + special_word + "_" + model_id
-        model_url_json = {"model_url": model_url}
-
-        with open("src/best_model/model_url.json", "w") as json_file:
-            json.dump(model_url_json, json_file)
-        mlflow.log_artifact("src/best_model/model_url.json")
-        sftp.put(
-            localpath="best_model/pytorch_model.bin",
-            remotepath="pytorch_model.bin",
-            callback=lambda x, y: printProgressDecimal(x, y),
-        )
-        sftp.put(
-            localpath="best_model/config.json",
-            remotepath="config.json",
-            callback=lambda x, y: printProgressDecimal(x, y),
+    try:
+        if "/" in model_id:
+            raise TypeError("##### special_word cannot include a character '/'")
+    except Exception as e:
+        print(e)
+        raise TypeError(
+            "Plz Check your parameter from train.py. \nThe model you trained may saved locally. \nSo, you can put on remote manually"
         )
-        print("Model Saved on", model_url)
-    sftp.close()
+    else:
+        with open("src/config/mlflow_config.yml") as f:
+            config_data = yaml.load(f, Loader=yaml.FullLoader)
+            print(config_data)
+        if experiment_name == "":
+            print("No input for experiment_name... import Default")
+            experiment_name = config_data["experiment_name"]
+
+        progressDict = {}
+        progressEveryPercent = 10
+
+        for i in range(0, 101):
+            if i % progressEveryPercent == 0:
+                progressDict[str(i)] = ""
+
+        def printProgressDecimal(x, y):
+            """A callback function for show sftp progress log.
+            Source: https://stackoverflow.com/questions/24278146/how-do-i-monitor-the-progress-of-a-file-transfer-through-pysftp
+
+            Args:
+                x (String): Represent for to-do-data size(e.g. remained file size of pulling of getting)
+                y (String): Represent for total file size
+            """
+            if (
+                int(100 * (int(x) / int(y))) % progressEveryPercent == 0
+                and progressDict[str(int(100 * (int(x) / int(y))))] == ""
+            ):
+                print(
+                    "{}% ({} Transfered(B)/ {} Total File Size(B))".format(
+                        str("%.2f" % (100 * (int(x) / int(y)))), x, y
+                    )
+                )
+                progressDict[str(int(100 * (int(x) / int(y))))] = "1"
+
+        with open("src/config/sftp_config.yml") as f:
+            config_data = yaml.load(f, Loader=yaml.FullLoader)
+        host = config_data["host"]
+        port = config_data["port"]
+        username = config_data["username"]
+        password = config_data["password"]
+
+        cnopts = pysftp.CnOpts()
+        cnopts.hostkeys = None
+
+        with pysftp.Connection(host, port=port, username=username, password=password, cnopts=cnopts) as sftp:
+            print("connected!!")
+            sftp.chdir("./mlflow_models")
+            try:
+                sftp.chdir(experiment_name)
+            except IOError:
+                sftp.mkdir(experiment_name)
+                sftp.chdir(experiment_name)
+            sftp.mkdir(model_id)
+            sftp.chdir(model_id)
+
+            model_url = "/mlflow_models/" + experiment_name + "/" + model_id
+            sftp.put(
+                localpath="src/best_model/pytorch_model.bin",
+                remotepath="pytorch_model.bin",
+                callback=lambda x, y: printProgressDecimal(x, y),
+            )
+            sftp.put(
+                localpath="src/best_model/config.json",
+                remotepath="config.json",
+                callback=lambda x, y: printProgressDecimal(x, y),
+            )
+            print("Success!!! Model Saved on", model_url)
+        sftp.close()
+    finally:
+        print("Model Saving job Finished")