pt: Fix multitask neighbor stat (#3367)

deepmd/pt/entrypoints/main.py

@@ -3,6 +3,9 @@
 import json
 import logging
 import os
+from copy import (
+    deepcopy,
+)
 from pathlib import (
     Path,
 )
@@ -75,9 +78,11 @@ def get_trainer(
     model_branch="",
     force_load=False,
     init_frz_model=None,
+    shared_links=None,
 ):
+    multi_task = "model_dict" in config.get("model", {})
     # argcheck
-    if "model_dict" not in config.get("model", {}):
+    if not multi_task:
         config = update_deepmd_input(config, warning=True, dump="input_v2_compat.json")
         config = normalize(config)
 
@@ -88,7 +93,6 @@ def get_trainer(
         assert dist.is_nccl_available()
         dist.init_process_group(backend="nccl")
 
-    multi_task = "model_dict" in config["model"]
     ckpt = init_model if init_model is not None else restart_model
     config["model"] = change_finetune_model_params(
         ckpt,
@@ -98,9 +102,6 @@ def get_trainer(
         model_branch=model_branch,
     )
     config["model"]["resuming"] = (finetune_model is not None) or (ckpt is not None)
-    shared_links = None
-    if multi_task:
-        config["model"], shared_links = preprocess_shared_params(config["model"])
 
     def prepare_trainer_input_single(
         model_params_single, data_dict_single, loss_dict_single, suffix=""
@@ -252,11 +253,33 @@ def train(FLAGS):
     SummaryPrinter()()
     with open(FLAGS.INPUT) as fin:
         config = json.load(fin)
+
+    # update multitask config
+    multi_task = "model_dict" in config["model"]
+    shared_links = None
+    if multi_task:
+        config["model"], shared_links = preprocess_shared_params(config["model"])
+
+    # do neighbor stat
     if not FLAGS.skip_neighbor_stat:
         log.info(
             "Calculate neighbor statistics... (add --skip-neighbor-stat to skip this step)"
         )
-        config["model"] = BaseModel.update_sel(config, config["model"])
+        if not multi_task:
+            config["model"] = BaseModel.update_sel(config, config["model"])
+        else:
+            training_jdata = deepcopy(config["training"])
+            training_jdata.pop("data_dict", {})
+            training_jdata.pop("model_prob", {})
+            for model_item in config["model"]["model_dict"]:
+                fake_global_jdata = {
+                    "model": deepcopy(config["model"]["model_dict"][model_item]),
+                    "training": deepcopy(config["training"]["data_dict"][model_item]),
+                }
+                fake_global_jdata["training"].update(training_jdata)
+                config["model"]["model_dict"][model_item] = BaseModel.update_sel(
+                    fake_global_jdata, config["model"]["model_dict"][model_item]
+                )
 
     trainer = get_trainer(
         config,
@@ -266,6 +289,7 @@ def train(FLAGS):
         FLAGS.model_branch,
         FLAGS.force_load,
         FLAGS.init_frz_model,
+        shared_links=shared_links,
     )
     trainer.run()