file to create the train/val/test splits

create_tar.py

@@ -0,0 +1,91 @@
+import tarfile
+import os
+import shutil
+import multiprocessing as mp
+from pathlib import Path
+from tqdm import tqdm
+import pandas as pd
+
+df = pd.read_parquet("df_train_v2.parquet")
+df["filename_full"] = "/home/fatrek/data_network/faton/riksdagen_anforanden/data/rixvox_v2/" + df["filename"]
+df = df.rename(columns={"sex": "gender"})
+
+# Group by intressent_id and count occurences of each id in the dataset, keep the columns speaker
+# and intressent_id and sort by the count of occurences
+
+df["speaker_total_hours"] = df.groupby(["speaker", "party"])["duration"].transform("sum") / 3600
+df_hours = df.groupby(["speaker", "party"]).first().sort_values("speaker_total_hours", ascending=False).reset_index()
+df_hours = df_hours.sample(frac=1, random_state=1337)  # Shuffle the rows
+
+# Set train equals to True until the cumulative sum of speaker_total_hours is 98% of the total
+df_hours["train"] = df_hours["speaker_total_hours"].cumsum() / df_hours["speaker_total_hours"].sum() < 0.98
+# Valid equals True until cumulative sum is 1% of the total
+df_hours["valid"] = False
+df_hours.loc[df_hours["train"] == False, "valid"] = (
+    df_hours[df_hours["train"] == 0]["speaker_total_hours"].cumsum() / df_hours["speaker_total_hours"].sum() < 0.01
+)
+df_hours["test"] = (df_hours["train"] == False) & (df_hours["valid"] == False)  # The rest is test
+
+# Create splits
+df_train = pd.merge(df, df_hours.loc[df_hours["train"], ["speaker", "party"]], on=["speaker", "party"], how="inner")
+df_valid = pd.merge(df, df_hours.loc[df_hours["valid"], ["speaker", "party"]], on=["speaker", "party"], how="inner")
+df_test = pd.merge(df, df_hours.loc[df_hours["test"], ["speaker", "party"]], on=["speaker", "party"], how="inner")
+
+
+def split_creator(df, observations_per_shard, shard_name):
+    df["shard"] = range(0, len(df))
+    df["shard"] = df["shard"] // observations_per_shard
+    df["shard"] = shard_name + "_" + df["shard"].astype(str)
+    return df["shard"]
+
+
+df_train["shard"] = split_creator(df_train, 6500, "train")
+df_valid["shard"] = split_creator(df_valid, 6500, "dev")
+df_test["shard"] = split_creator(df_test, 6500, "test")
+
+df_train["nr_words"] = df_train["text"].str.split().str.len()
+df_train = df_train[df_train["nr_words"] <= 160].reset_index(drop=True)
+df_train = df_train.drop(columns="nr_words")
+
+
+def create_tar(df, data_folder="/home/fatrek/data_network/faton/rixvox/data"):
+    shard_filename = df["shard"].reset_index(drop=True).values[0]
+    shard_filename = shard_filename + ".tar.gz"
+    split = df["shard"].reset_index(drop=True).str.extract(r"(.*)_")[0][0]  # train_0 -> train
+    os.makedirs(os.path.join(data_folder, split), exist_ok=True)
+
+    print(f"Creating tarfile: {os.path.join(data_folder, split, shard_filename)}")
+    with tarfile.open(os.path.join(data_folder, split, shard_filename), "w:gz") as tar:
+        for filename in df["filename_full"].values:
+            tar.add(Path(filename), arcname=Path(filename).relative_to(Path(filename).parent.parent), recursive=False)
+
+
+# Group by shard and split dataframes in to several dataframes in list
+groups = df_train.groupby("shard")
+df_train_list = [groups.get_group(x) for x in groups.groups]
+groups = df_valid.groupby("shard")
+df_valid_list = [groups.get_group(x) for x in groups.groups]
+groups = df_test.groupby("shard")
+df_test_list = [groups.get_group(x) for x in groups.groups]
+
+
+data_folder = "/home/fatrek/data_network/faton/RixVox/data"
+
+# for shard in df_train_list:
+#     create_tar(shard, data_folder)
+
+with mp.Pool(16) as pool:
+    pool.map(create_tar, df_train_list)
+
+with mp.Pool(1) as pool:
+    pool.map(create_tar, df_valid_list)
+    pool.map(create_tar, df_test_list)
+
+
+df_train = df_train.drop(columns=["shard", "filename_full", "file_size"])
+df_valid = df_valid.drop(columns=["shard", "filename_full", "file_size"])
+df_test = df_test.drop(columns=["shard", "filename_full", "file_size"])
+
+df_train.to_parquet(os.path.join("data", "train_metadata.parquet"), index=False)
+df_valid.to_parquet(os.path.join("data", "dev_metadata.parquet"), index=False)
+df_test.to_parquet(os.path.join("data", "test_metadata.parquet"), index=False)