feat(pt/tf): init-(frz)-model use pretrain script (#3926)

Support `--use-pretrain-script` for pt&tf when doing init-(frz)-model.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **Tests**
- Enhanced and added new test cases for deep learning model
initialization and evaluation.
- Improved setup and cleanup processes for temporary files and
directories in tests to ensure a cleaner test environment.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

deepmd/main.py

@@ -258,7 +258,9 @@ def main_parser() -> argparse.ArgumentParser:
     parser_train.add_argument(
         "--use-pretrain-script",
         action="store_true",
-        help="Use model parameters from the script of the pretrained model instead of user input when doing finetuning. Note: This behavior is default and unchangeable in TensorFlow.",
+        help="When performing fine-tuning or init-model, "
+        "utilize the model parameters provided by the script of the pretrained model rather than relying on user input. "
+        "It is important to note that in TensorFlow, this behavior is the default and cannot be modified for fine-tuning. ",
     )
     parser_train.add_argument(
         "-o",

deepmd/pt/entrypoints/main.py

@@ -256,6 +256,21 @@ def train(FLAGS):
             model_branch=FLAGS.model_branch,
             change_model_params=FLAGS.use_pretrain_script,
         )
+    # update init_model or init_frz_model config if necessary
+    if (
+        FLAGS.init_model is not None or FLAGS.init_frz_model is not None
+    ) and FLAGS.use_pretrain_script:
+        if FLAGS.init_model is not None:
+            init_state_dict = torch.load(FLAGS.init_model, map_location=DEVICE)
+            if "model" in init_state_dict:
+                init_state_dict = init_state_dict["model"]
+            config["model"] = init_state_dict["_extra_state"]["model_params"]
+        else:
+            config["model"] = json.loads(
+                torch.jit.load(
+                    FLAGS.init_frz_model, map_location=DEVICE
+                ).get_model_def_script()
+            )
 
     # argcheck
     config = update_deepmd_input(config, warning=True, dump="input_v2_compat.json")

deepmd/tf/entrypoints/train.py

@@ -65,6 +65,7 @@ def train(
     is_compress: bool = False,
     skip_neighbor_stat: bool = False,
     finetune: Optional[str] = None,
+    use_pretrain_script: bool = False,
     **kwargs,
 ):
     """Run DeePMD model training.
@@ -93,6 +94,9 @@ def train(
         skip checking neighbor statistics
     finetune : Optional[str]
         path to pretrained model or None
+    use_pretrain_script : bool
+        Whether to use model script in pretrained model when doing init-model or init-frz-model.
+        Note that this option is true and unchangeable for fine-tuning.
     **kwargs
         additional arguments
 
@@ -123,6 +127,41 @@ def train(
             jdata, run_opt.finetune
         )
 
+    if (
+        run_opt.init_model is not None or run_opt.init_frz_model is not None
+    ) and use_pretrain_script:
+        from deepmd.tf.utils.errors import (
+            GraphWithoutTensorError,
+        )
+        from deepmd.tf.utils.graph import (
+            get_tensor_by_name,
+            get_tensor_by_name_from_graph,
+        )
+
+        err_msg = (
+            f"The input model: {run_opt.init_model if run_opt.init_model is not None else run_opt.init_frz_model} has no training script, "
+            f"Please use the model pretrained with v2.1.5 or higher version of DeePMD-kit."
+        )
+        if run_opt.init_model is not None:
+            with tf.Graph().as_default() as graph:
+                tf.train.import_meta_graph(
+                    f"{run_opt.init_model}.meta", clear_devices=True
+                )
+            try:
+                t_training_script = get_tensor_by_name_from_graph(
+                    graph, "train_attr/training_script"
+                )
+            except GraphWithoutTensorError as e:
+                raise RuntimeError(err_msg) from e
+        else:
+            try:
+                t_training_script = get_tensor_by_name(
+                    run_opt.init_frz_model, "train_attr/training_script"
+                )
+            except GraphWithoutTensorError as e:
+                raise RuntimeError(err_msg) from e
+        jdata["model"] = json.loads(t_training_script)["model"]
+
     jdata = update_deepmd_input(jdata, warning=True, dump="input_v2_compat.json")
 
     jdata = normalize(jdata)

source/tests/pt/test_init_frz_model.py

@@ -1,5 +1,8 @@
 # SPDX-License-Identifier: LGPL-3.0-or-later
 import json
+import os
+import shutil
+import tempfile
 import unittest
 from argparse import (
     Namespace,
@@ -21,12 +24,17 @@
     DeepPot,
 )
 
+from .common import (
+    run_dp,
+)
+
 
 class TestInitFrzModel(unittest.TestCase):
     def setUp(self):
         input_json = str(Path(__file__).parent / "water/se_atten.json")
         with open(input_json) as f:
             config = json.load(f)
+        config["model"]["descriptor"]["smooth_type_embedding"] = True
         config["training"]["numb_steps"] = 1
         config["training"]["save_freq"] = 1
         config["learning_rate"]["start_lr"] = 1.0
@@ -38,15 +46,30 @@ def setUp(self):
         ]
 
         self.models = []
-        for imodel in range(2):
-            if imodel == 1:
-                config["training"]["numb_steps"] = 0
-                trainer = get_trainer(deepcopy(config), init_frz_model=self.models[-1])
+        for imodel in range(3):
+            frozen_model = f"frozen_model{imodel}.pth"
+            if imodel == 0:
+                temp_config = deepcopy(config)
+                trainer = get_trainer(temp_config)
+            elif imodel == 1:
+                temp_config = deepcopy(config)
+                temp_config["training"]["numb_steps"] = 0
+                trainer = get_trainer(temp_config, init_frz_model=self.models[-1])
             else:
-                trainer = get_trainer(deepcopy(config))
-            trainer.run()
+                empty_config = deepcopy(config)
+                empty_config["model"]["descriptor"] = {}
+                empty_config["model"]["fitting_net"] = {}
+                empty_config["training"]["numb_steps"] = 0
+                tmp_input = tempfile.NamedTemporaryFile(delete=False, suffix=".json")
+                with open(tmp_input.name, "w") as f:
+                    json.dump(empty_config, f, indent=4)
+                run_dp(
+                    f"dp --pt train {tmp_input.name} --init-frz-model {self.models[-1]} --use-pretrain-script --skip-neighbor-stat"
+                )
+                trainer = None
 
-            frozen_model = f"frozen_model{imodel}.pth"
+            if imodel in [0, 1]:
+                trainer.run()
             ns = Namespace(
                 model="model.pt",
                 output=frozen_model,
@@ -58,6 +81,7 @@ def setUp(self):
     def test_dp_test(self):
         dp1 = DeepPot(str(self.models[0]))
         dp2 = DeepPot(str(self.models[1]))
+        dp3 = DeepPot(str(self.models[2]))
         cell = np.array(
             [
                 5.122106549439247480e00,
@@ -96,8 +120,26 @@ def test_dp_test(self):
         e1, f1, v1, ae1, av1 = ret1[0], ret1[1], ret1[2], ret1[3], ret1[4]
         ret2 = dp2.eval(coord, cell, atype, atomic=True)
         e2, f2, v2, ae2, av2 = ret2[0], ret2[1], ret2[2], ret2[3], ret2[4]
+        ret3 = dp3.eval(coord, cell, atype, atomic=True)
+        e3, f3, v3, ae3, av3 = ret3[0], ret3[1], ret3[2], ret3[3], ret3[4]
         np.testing.assert_allclose(e1, e2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(e1, e3, rtol=1e-10, atol=1e-10)
         np.testing.assert_allclose(f1, f2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(f1, f3, rtol=1e-10, atol=1e-10)
         np.testing.assert_allclose(v1, v2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(v1, v3, rtol=1e-10, atol=1e-10)
         np.testing.assert_allclose(ae1, ae2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(ae1, ae3, rtol=1e-10, atol=1e-10)
         np.testing.assert_allclose(av1, av2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(av1, av3, rtol=1e-10, atol=1e-10)
+
+    def tearDown(self):
+        for f in os.listdir("."):
+            if f.startswith("frozen_model") and f.endswith(".pth"):
+                os.remove(f)
+            if f.startswith("model") and f.endswith(".pt"):
+                os.remove(f)
+            if f in ["lcurve.out"]:
+                os.remove(f)
+            if f in ["stat_files"]:
+                shutil.rmtree(f)

source/tests/pt/test_init_model.py

@@ -0,0 +1,136 @@
+# SPDX-License-Identifier: LGPL-3.0-or-later
+import json
+import os
+import shutil
+import tempfile
+import unittest
+from copy import (
+    deepcopy,
+)
+from pathlib import (
+    Path,
+)
+
+import numpy as np
+
+from deepmd.pt.entrypoints.main import (
+    get_trainer,
+)
+from deepmd.pt.infer.deep_eval import (
+    DeepPot,
+)
+
+from .common import (
+    run_dp,
+)
+
+
+class TestInitModel(unittest.TestCase):
+    def setUp(self):
+        input_json = str(Path(__file__).parent / "water/se_atten.json")
+        with open(input_json) as f:
+            config = json.load(f)
+        config["model"]["descriptor"]["smooth_type_embedding"] = True
+        config["training"]["numb_steps"] = 1
+        config["training"]["save_freq"] = 1
+        config["learning_rate"]["start_lr"] = 1.0
+        config["training"]["training_data"]["systems"] = [
+            str(Path(__file__).parent / "water/data/single")
+        ]
+        config["training"]["validation_data"]["systems"] = [
+            str(Path(__file__).parent / "water/data/single")
+        ]
+
+        self.models = []
+        for imodel in range(3):
+            ckpt_model = f"model{imodel}.ckpt"
+            if imodel == 0:
+                temp_config = deepcopy(config)
+                temp_config["training"]["save_ckpt"] = ckpt_model
+                trainer = get_trainer(temp_config)
+            elif imodel == 1:
+                temp_config = deepcopy(config)
+                temp_config["training"]["numb_steps"] = 0
+                temp_config["training"]["save_ckpt"] = ckpt_model
+                trainer = get_trainer(temp_config, init_model=self.models[-1])
+            else:
+                empty_config = deepcopy(config)
+                empty_config["model"]["descriptor"] = {}
+                empty_config["model"]["fitting_net"] = {}
+                empty_config["training"]["numb_steps"] = 0
+                empty_config["training"]["save_ckpt"] = ckpt_model
+                tmp_input = tempfile.NamedTemporaryFile(delete=False, suffix=".json")
+                with open(tmp_input.name, "w") as f:
+                    json.dump(empty_config, f, indent=4)
+                run_dp(
+                    f"dp --pt train {tmp_input.name} --init-model {self.models[-1]} --use-pretrain-script --skip-neighbor-stat"
+                )
+                trainer = None
+
+            if imodel in [0, 1]:
+                trainer.run()
+            self.models.append(ckpt_model + ".pt")
+
+    def test_dp_test(self):
+        dp1 = DeepPot(str(self.models[0]))
+        dp2 = DeepPot(str(self.models[1]))
+        dp3 = DeepPot(str(self.models[2]))
+        cell = np.array(
+            [
+                5.122106549439247480e00,
+                4.016537340154059388e-01,
+                6.951654033828678081e-01,
+                4.016537340154059388e-01,
+                6.112136112297989143e00,
+                8.178091365465004481e-01,
+                6.951654033828678081e-01,
+                8.178091365465004481e-01,
+                6.159552512682983760e00,
+            ]
+        ).reshape(1, 3, 3)
+        coord = np.array(
+            [
+                2.978060152121375648e00,
+                3.588469695887098077e00,
+                2.792459820604495491e00,
+                3.895592322591093115e00,
+                2.712091020667753760e00,
+                1.366836847133650501e00,
+                9.955616170888935690e-01,
+                4.121324820711413039e00,
+                1.817239061889086571e00,
+                3.553661462345699906e00,
+                5.313046969500791583e00,
+                6.635182659098815883e00,
+                6.088601018589653080e00,
+                6.575011420004332585e00,
+                6.825240650611076099e00,
+            ]
+        ).reshape(1, -1, 3)
+        atype = np.array([0, 0, 0, 1, 1]).reshape(1, -1)
+
+        ret1 = dp1.eval(coord, cell, atype, atomic=True)
+        e1, f1, v1, ae1, av1 = ret1[0], ret1[1], ret1[2], ret1[3], ret1[4]
+        ret2 = dp2.eval(coord, cell, atype, atomic=True)
+        e2, f2, v2, ae2, av2 = ret2[0], ret2[1], ret2[2], ret2[3], ret2[4]
+        ret3 = dp3.eval(coord, cell, atype, atomic=True)
+        e3, f3, v3, ae3, av3 = ret3[0], ret3[1], ret3[2], ret3[3], ret3[4]
+        np.testing.assert_allclose(e1, e2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(e1, e3, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(f1, f2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(f1, f3, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(v1, v2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(v1, v3, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(ae1, ae2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(ae1, ae3, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(av1, av2, rtol=1e-10, atol=1e-10)
+        np.testing.assert_allclose(av1, av3, rtol=1e-10, atol=1e-10)
+
+    def tearDown(self):
+        for f in os.listdir("."):
+            if f.startswith("model") and f.endswith(".pt"):
+                os.remove(f)
+            if f in ["lcurve.out"]:
+                os.remove(f)
+            if f in ["stat_files"]:
+                shutil.rmtree(f)