Migrate amc4th training (#1316)

* Migrate amc4th training

* Refine RL example scripts

* Resolve PR comments

Co-authored-by: luocy16 <luocy16@mails.tsinghua.edu.cn>

examples/rl/scripts/collect_pickle_dataframe.py

@@ -0,0 +1,21 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import os
+import pickle
+import pandas as pd
+from tqdm import tqdm
+
+os.makedirs(os.path.join("data", "pickle_dataframe"), exist_ok=True)
+
+for tag in ("backtest", "feature"):
+    df = pickle.load(open(os.path.join("data", "pickle", f"{tag}.pkl"), "rb"))
+    df = pd.concat(list(df.values())).reset_index()
+    df["date"] = df["datetime"].dt.date.astype("datetime64")
+    instruments = sorted(set(df["instrument"]))
+    
+    os.makedirs(os.path.join("data", "pickle_dataframe", tag), exist_ok=True)
+    for instrument in tqdm(instruments):
+        cur = df[df["instrument"] == instrument].sort_values(by=["datetime"])
+        cur = cur.set_index(["instrument", "datetime", "date"])
+        pickle.dump(cur, open(os.path.join("data", "pickle_dataframe", tag, f"{instrument}.pkl"), "wb"))

examples/rl/scripts/data_pipeline.sh

@@ -0,0 +1,14 @@
+# Generate `bin` format data
+set -e
+python ../../scripts/dump_bin.py dump_all --csv_path ./data/csv --qlib_dir ./data/bin --include_fields open,close,high,low,vwap,volume --symbol_field_name symbol --date_field_name date --freq 1min
+
+# Generate pickle format data
+python scripts/gen_pickle_data.py -c scripts/pickle_data_config.yml
+if [ -e stat/ ]; then
+    rm -r stat/
+fi
+python scripts/collect_pickle_dataframe.py
+
+# Sample orders
+python scripts/gen_training_orders.py
+python scripts/gen_backtest_orders.py

examples/rl/scripts/gen_backtest_orders.py

@@ -0,0 +1,41 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import argparse
+import os
+import pandas as pd
+import numpy as np
+import pickle
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--seed", type=int, default=20220926)
+parser.add_argument("--num_order", type=int, default=10)
+args = parser.parse_args()
+
+np.random.seed(args.seed)
+
+path = os.path.join("data", "pickle", "backtesttest.pkl")  # TODO: rename file
+df = pickle.load(open(path, "rb")).reset_index()
+df["date"] = df["datetime"].dt.date.astype("datetime64")
+
+instruments = sorted(set(df["instrument"]))
+df_list = []
+for instrument in instruments:
+    print(instrument)
+    
+    cur_df = df[df["instrument"] == instrument]
+    
+    dates = sorted(set([str(d).split(" ")[0] for d in cur_df["date"]]))
+    
+    n = args.num_order
+    df_list.append(
+        pd.DataFrame({
+            "date": sorted(np.random.choice(dates, size=n, replace=False)),
+            "instrument": [instrument] * n,
+            "amount": np.random.randint(low=3, high=11, size=n) * 100.0,
+            "order_type": np.random.randint(low=0, high=2, size=n),
+        }).set_index(["date", "instrument"]),
+    )
+
+total_df = pd.concat(df_list)
+total_df.to_csv("data/backtest_orders.csv")

examples/rl/scripts/gen_pickle_data.py

@@ -0,0 +1,43 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import yaml
+import argparse
+import os
+from copy import deepcopy
+
+from qlib.contrib.data.highfreq_provider import HighFreqProvider
+
+loader = yaml.FullLoader
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-c", "--config", type=str, default="config.yml")
+    parser.add_argument("-d", "--dest", type=str, default=".")
+    parser.add_argument("-s", "--split", type=str, choices=["none", "date", "stock", "both"], default="stock")
+    args = parser.parse_args()
+
+    conf = yaml.load(open(args.config), Loader=loader)
+
+    for k, v in conf.items():
+        if isinstance(v, dict) and "path" in v:
+            v["path"] = os.path.join(args.dest, v["path"])
+    provider = HighFreqProvider(**conf)
+
+    # Gen dataframe
+    if "feature_conf" in conf:
+        feature = provider._gen_dataframe(deepcopy(provider.feature_conf))
+    if "backtest_conf" in conf:
+        backtest = provider._gen_dataframe(deepcopy(provider.backtest_conf))
+
+    provider.feature_conf['path'] = os.path.splitext(provider.feature_conf['path'])[0] + '/'
+    provider.backtest_conf['path'] = os.path.splitext(provider.backtest_conf['path'])[0] + '/'
+    # Split by date
+    if args.split == "date" or args.split == "both":
+        provider._gen_day_dataset(deepcopy(provider.feature_conf), "feature")
+        provider._gen_day_dataset(deepcopy(provider.backtest_conf), "backtest")
+
+    # Split by stock
+    if args.split == "stock" or args.split == "both":
+        provider._gen_stock_dataset(deepcopy(provider.feature_conf), "feature")
+        provider._gen_stock_dataset(deepcopy(provider.backtest_conf), "backtest")

examples/rl/scripts/gen_training_orders.py

@@ -0,0 +1,37 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+import argparse
+import os
+import pandas as pd
+import numpy as np
+import pickle
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--seed", type=int, default=20220926)
+parser.add_argument("--stock", type=str, default="AAPL")
+parser.add_argument("--train_size", type=int, default=10)
+parser.add_argument("--valid_size", type=int, default=2)
+parser.add_argument("--test_size", type=int, default=2)
+args = parser.parse_args()
+
+np.random.seed(args.seed)
+
+os.makedirs(os.path.join("data", "training_order_split"), exist_ok=True)
+
+for group, n in zip(("train", "valid", "test"), (args.train_size, args.valid_size, args.test_size)):
+    path = os.path.join("data", "pickle", f"backtest{group}.pkl")
+    df = pickle.load(open(path, "rb")).reset_index()
+    df["date"] = df["datetime"].dt.date.astype("datetime64")
+    
+    dates = sorted(set([str(d).split(" ")[0] for d in df["date"]]))
+
+    data_df = pd.DataFrame({
+        "date": sorted(np.random.choice(dates, size=n, replace=False)),
+        "instrument": [args.stock] * n,
+        "amount": np.random.randint(low=3, high=11, size=n) * 100.0,
+        "order_type": [0] * n,
+    }).set_index(["date", "instrument"])
+
+    os.makedirs(os.path.join("data", "training_order_split", group), exist_ok=True)
+    pickle.dump(data_df, open(os.path.join("data", "training_order_split", group, f"{args.stock}.pkl"), "wb"))

examples/rl/scripts/pickle_data_config.yml

@@ -0,0 +1,57 @@
+# start & end time for training/validation/test datasets
+start_time: !!str &start 2020-01-01
+end_time: !!str &end 2020-07-31
+train_end_time: !!str &tend 2020-03-31
+valid_start_time: !!str &vstart 2020-04-01
+valid_end_time: !!str &vend 2020-05-31
+test_start_time: !!str &tstart 2020-06-01
+# the instrument set
+instruments: &ins all
+# qlib related configuration
+qlib_conf:
+    provider_uri: ./data/bin # path to generated qlib bin
+    redis_port: 233
+feature_conf:
+    path: ./data/pickle/feature.pkl # output path of feature
+    class: DatasetH
+    module_path: qlib.data.dataset
+    kwargs:
+        handler:
+            class: HighFreqGeneralHandler
+            module_path: qlib.contrib.data.highfreq_handler
+            kwargs:
+                start_time: *start
+                end_time: *end
+                fit_start_time: *start
+                fit_end_time: *tend
+                instruments: *ins
+                day_length: 240 # how many minutes in one trading day
+                infer_processors:
+                - class: HighFreqNorm
+                  module_path: qlib.contrib.data.highfreq_processor
+                  kwargs:
+                    feature_save_dir: ./stat/  #  output path of statistics of features (for feature normalization)
+                    norm_groups: 
+                        price: 10
+                        volume: 2
+        segments:
+            train: !!python/tuple [*start, *tend]
+            valid: !!python/tuple [*vstart, *vend]
+            test: !!python/tuple [*tstart, *end]
+backtest_conf:
+    path: ./data/pickle/backtest.pkl # output path of backtest
+    class: DatasetH
+    module_path: qlib.data.dataset
+    kwargs:
+        handler:
+            class: HighFreqGeneralBacktestHandler
+            module_path: qlib.contrib.data.highfreq_handler
+            kwargs:
+                start_time: *start
+                end_time: *end
+                instruments: *ins
+                day_length: 240
+        segments:
+            train: !!python/tuple [*start, *tend]
+            valid: !!python/tuple [*vstart, *vend]
+            test: !!python/tuple [*tstart, *end]