Fix handling of optional [pretraining] block (#5954)

* Fix handling of optional [pretraining] block

* Remote pretraining from default config

* Fix test

* Add schema option for empty pretrain block

spacy/cli/debug_model.py

@@ -51,7 +51,7 @@ def debug_model_cli(
     with show_validation_error(config_path):
         config = util.load_config(config_path, overrides=config_overrides)
         nlp, config = util.load_model_from_config(config_path)
-    seed = config["pretraining"]["seed"]
+    seed = config["training"]["seed"]
     if seed is not None:
         msg.info(f"Fixing random seed: {seed}")
         fix_random_seed(seed)

spacy/cli/init_config.py

@@ -7,6 +7,7 @@ import srsly
 import re
 
 from .. import util
+from ..language import DEFAULT_CONFIG_PRETRAIN_PATH
 from ..schemas import RecommendationSchema
 from ._util import init_cli, Arg, Opt, show_validation_error, COMMAND
 
@@ -48,6 +49,7 @@ def init_fill_config_cli(
     # fmt: off
     base_path: Path = Arg(..., help="Base config to fill", exists=True, dir_okay=False),
     output_file: Path = Arg("-", help="File to save config.cfg to (or - for stdout)", allow_dash=True),
+    pretraining: bool = Opt(False, "--pretraining", "-p", help="Include config for pretraining (with 'spacy pretrain')"),
     diff: bool = Opt(False, "--diff", "-D", help="Print a visual diff highlighting the changes")
     # fmt: on
 ):
@@ -58,19 +60,24 @@ def init_fill_config_cli(
     can be used with a config generated via the training quickstart widget:
     https://nightly.spacy.io/usage/training#quickstart
     """
-    fill_config(output_file, base_path, diff=diff)
+    fill_config(output_file, base_path, pretraining=pretraining, diff=diff)
 
 
 def fill_config(
-    output_file: Path, base_path: Path, *, diff: bool = False
+    output_file: Path, base_path: Path, *, pretraining: bool = False, diff: bool = False
 ) -> Tuple[Config, Config]:
     is_stdout = str(output_file) == "-"
     msg = Printer(no_print=is_stdout)
     with show_validation_error(hint_fill=False):
         config = util.load_config(base_path)
         nlp, _ = util.load_model_from_config(config, auto_fill=True)
+    filled = nlp.config
+    if pretraining:
+        validate_config_for_pretrain(filled, msg)
+        pretrain_config = util.load_config(DEFAULT_CONFIG_PRETRAIN_PATH)
+        filled = pretrain_config.merge(filled)
     before = config.to_str()
-    after = nlp.config.to_str()
+    after = filled.to_str()
     if before == after:
         msg.warn("Nothing to auto-fill: base config is already complete")
     else:
@@ -84,7 +91,7 @@ def fill_config(
             print(diff_strings(before, after))
             msg.divider("END CONFIG DIFF")
             print("")
-    save_config(nlp.config, output_file, is_stdout=is_stdout)
+    save_config(filled, output_file, is_stdout=is_stdout)
 
 
 def init_config(
@@ -132,12 +139,9 @@ def init_config(
     msg.info("Generated template specific for your use case")
     for label, value in use_case.items():
         msg.text(f"- {label}: {value}")
-    use_transformer = bool(template_vars.use_transformer)
     with show_validation_error(hint_fill=False):
         config = util.load_config_from_str(base_template)
         nlp, _ = util.load_model_from_config(config, auto_fill=True)
-    if use_transformer:
-        nlp.config.pop("pretraining", {})  # TODO: solve this better
     msg.good("Auto-filled config with all values")
     save_config(nlp.config, output_file, is_stdout=is_stdout)
 
@@ -161,3 +165,15 @@ def has_spacy_transformers() -> bool:
         return True
     except ImportError:
         return False
+
+
+def validate_config_for_pretrain(config: Config, msg: Printer) -> None:
+    if "tok2vec" not in config["nlp"]["pipeline"]:
+        msg.warn(
+            "No tok2vec component found in the pipeline. If your tok2vec "
+            "component has a different name, you may need to adjust the "
+            "tok2vec_model reference in the [pretraining] block. If you don't "
+            "have a tok2vec component, make sure to add it to your [components] "
+            "and the pipeline specified in the [nlp] block, so you can pretrain "
+            "weights for it."
+        )

spacy/cli/pretrain.py

@@ -90,19 +90,20 @@ def pretrain(
     with show_validation_error(config_path):
         config = util.load_config(config_path, overrides=config_overrides)
         nlp, config = util.load_model_from_config(config)
-    # TODO: validate that [pretraining] block exists
+    pretrain_config = config["pretraining"]
+    if not pretrain_config:
+        # TODO: What's the solution here? How do we handle optional blocks?
+        msg.fail("The [pretraining] block in your config is empty", exits=1)
     if not output_dir.exists():
         output_dir.mkdir()
         msg.good(f"Created output directory: {output_dir}")
-    seed = config["pretraining"]["seed"]
+    seed = pretrain_config["seed"]
     if seed is not None:
         fix_random_seed(seed)
-    if use_gpu >= 0 and config["pretraining"]["use_pytorch_for_gpu_memory"]:
+    if use_gpu >= 0 and pretrain_config["use_pytorch_for_gpu_memory"]:
         use_pytorch_for_gpu_memory()
     config.to_disk(output_dir / "config.cfg")
     msg.good("Saved config file in the output directory")
-    pretrain_config = config["pretraining"]
-
     if texts_loc != "-":  # reading from a file
         with msg.loading("Loading input texts..."):
             texts = list(srsly.read_jsonl(texts_loc))

spacy/cli/train.py

@@ -117,7 +117,7 @@ def train(
 
     # Load a pretrained tok2vec model - cf. CLI command 'pretrain'
     if weights_data is not None:
-        tok2vec_path = config.get("pretraining", {}).get("tok2vec_model", None)
+        tok2vec_path = config["pretraining"].get("tok2vec_model", None)
         if tok2vec_path is None:
             msg.fail(
                 f"To use a pretrained tok2vec model, the config needs to specify which "

spacy/default_config.cfg

@@ -90,29 +90,3 @@ eps = 1e-8
 warmup_steps = 250
 total_steps = 20000
 initial_rate = 0.001
-
-[pretraining]
-max_epochs = 1000
-min_length = 5
-max_length = 500
-dropout = 0.2
-n_save_every = null
-batch_size = 3000
-seed = ${system.seed}
-use_pytorch_for_gpu_memory = ${system.use_pytorch_for_gpu_memory}
-tok2vec_model = "components.tok2vec.model"
-
-[pretraining.objective]
-type = "characters"
-n_characters = 4
-
-[pretraining.optimizer]
-@optimizers = "Adam.v1"
-beta1 = 0.9
-beta2 = 0.999
-L2_is_weight_decay = true
-L2 = 0.01
-grad_clip = 1.0
-use_averages = true
-eps = 1e-8
-learn_rate = 0.001

spacy/default_config_pretraining.cfg

@@ -0,0 +1,25 @@
+[pretraining]
+max_epochs = 1000
+min_length = 5
+max_length = 500
+dropout = 0.2
+n_save_every = null
+batch_size = 3000
+seed = ${system.seed}
+use_pytorch_for_gpu_memory = ${system.use_pytorch_for_gpu_memory}
+tok2vec_model = "components.tok2vec.model"
+
+[pretraining.objective]
+type = "characters"
+n_characters = 4
+
+[pretraining.optimizer]
+@optimizers = "Adam.v1"
+beta1 = 0.9
+beta2 = 0.999
+L2_is_weight_decay = true
+L2 = 0.01
+grad_clip = 1.0
+use_averages = true
+eps = 1e-8
+learn_rate = 0.001

spacy/language.py

@@ -37,6 +37,9 @@ from . import about
 # This is the base config will all settings (training etc.)
 DEFAULT_CONFIG_PATH = Path(__file__).parent / "default_config.cfg"
 DEFAULT_CONFIG = util.load_config(DEFAULT_CONFIG_PATH)
+# This is the base config for the [pretraining] block and currently not included
+# in the main config and only added via the 'init fill-config' command
+DEFAULT_CONFIG_PRETRAIN_PATH = Path(__file__).parent / "default_config_pretraining.cfg"
 
 
 class BaseDefaults:

spacy/schemas.py

@@ -233,6 +233,11 @@ class ConfigSchemaNlp(BaseModel):
         arbitrary_types_allowed = True
 
 
+class ConfigSchemaPretrainEmpty(BaseModel):
+    class Config:
+        extra = "forbid"
+
+
 class ConfigSchemaPretrain(BaseModel):
     # fmt: off
     max_epochs: StrictInt = Field(..., title="Maximum number of epochs to train for")
@@ -257,16 +262,17 @@ class ConfigSchemaPretrain(BaseModel):
 class ConfigSchema(BaseModel):
     training: ConfigSchemaTraining
     nlp: ConfigSchemaNlp
-    pretraining: Optional[ConfigSchemaPretrain]
+    pretraining: Union[ConfigSchemaPretrain, ConfigSchemaPretrainEmpty] = {}
     components: Dict[str, Dict[str, Any]]
 
     @root_validator
     def validate_config(cls, values):
         """Perform additional validation for settings with dependencies."""
         pt = values.get("pretraining")
-        if pt and pt.objective.get("type") == "vectors" and not values["nlp"].vectors:
-            err = "Need nlp.vectors if pretraining.objective.type is vectors"
-            raise ValueError(err)
+        if pt and not isinstance(pt, ConfigSchemaPretrainEmpty):
+            if pt.objective.get("type") == "vectors" and not values["nlp"].vectors:
+                err = "Need nlp.vectors if pretraining.objective.type is vectors"
+                raise ValueError(err)
         return values
 
     class Config:

spacy/tests/serialize/test_serialize_config.py

@@ -3,10 +3,11 @@ from thinc.config import Config, ConfigValidationError
 import spacy
 from spacy.lang.en import English
 from spacy.lang.de import German
-from spacy.language import Language
+from spacy.language import Language, DEFAULT_CONFIG
 from spacy.util import registry, load_model_from_config
 from spacy.ml.models import build_Tok2Vec_model, build_tb_parser_model
 from spacy.ml.models import MultiHashEmbed, MaxoutWindowEncoder
+from spacy.schemas import ConfigSchema
 
 from ..util import make_tempdir
 
@@ -299,3 +300,16 @@ def test_config_interpolation():
     nlp2 = English.from_config(interpolated)
     assert nlp2.config["training"]["train_corpus"]["path"] == ""
     assert nlp2.config["components"]["tagger"]["model"]["tok2vec"]["width"] == 342
+
+
+def test_config_optional_sections():
+    config = Config().from_str(nlp_config_string)
+    config = DEFAULT_CONFIG.merge(config)
+    assert "pretraining" not in config
+    filled = registry.fill_config(config, schema=ConfigSchema, validate=False)
+    # Make sure that optional "pretraining" block doesn't default to None,
+    # which would (rightly) cause error because it'd result in a top-level
+    # key that's not a section (dict). Note that the following roundtrip is
+    # also how Config.interpolate works under the hood.
+    new_config = Config().from_str(filled.to_str())
+    assert new_config["pretraining"] == {}