Merge pull request #6104 from svlandeg/fix/debug_model [ci skip]

spacy/cli/debug_model.py

@@ -1,5 +1,9 @@
-from typing import Dict, Any, Optional
+import warnings
+from typing import Dict, Any, Optional, Iterable
 from pathlib import Path
+
+from spacy.training import Example
+from spacy.util import dot_to_object
 from wasabi import msg
 from thinc.api import require_gpu, fix_random_seed, set_dropout_rate, Adam
 from thinc.api import Model, data_validation, set_gpu_allocator
@@ -59,23 +63,24 @@ def debug_model_cli(
         allocator = config["training"]["gpu_allocator"]
         if use_gpu >= 0 and allocator:
             set_gpu_allocator(allocator)
-        nlp, config = util.load_model_from_config(config_path)
+        nlp, config = util.load_model_from_config(config)
     seed = config["training"]["seed"]
     if seed is not None:
         msg.info(f"Fixing random seed: {seed}")
         fix_random_seed(seed)
     pipe = nlp.get_pipe(component)
-    if hasattr(pipe, "model"):
+    if not hasattr(pipe, "model"):
-        model = pipe.model
-    else:
         msg.fail(
             f"The component '{component}' does not specify an object that holds a Model.",
             exits=1,
         )
-    debug_model(model, print_settings=print_settings)
+    model = pipe.model
+    debug_model(config, nlp, model, print_settings=print_settings)
 
 
-def debug_model(model: Model, *, print_settings: Optional[Dict[str, Any]] = None):
+def debug_model(
+    config, nlp, model: Model, *, print_settings: Optional[Dict[str, Any]] = None
+):
     if not isinstance(model, Model):
         msg.fail(
             f"Requires a Thinc Model to be analysed, but found {type(model)} instead.",
@@ -92,10 +97,23 @@ def debug_model(model: Model, *, print_settings: Optional[Dict[str, Any]] = None
 
     # STEP 1: Initializing the model and printing again
     X = _get_docs()
-    Y = _get_output(model.ops.xp)
     # The output vector might differ from the official type of the output layer
     with data_validation(False):
-        model.initialize(X=X, Y=Y)
+        try:
+            train_corpus = dot_to_object(config, config["training"]["train_corpus"])
+            nlp.begin_training(lambda: train_corpus(nlp))
+            msg.info("Initialized the model with the training corpus.")
+        except ValueError:
+            try:
+                _set_output_dim(nO=7, model=model)
+                nlp.begin_training(lambda: [Example.from_dict(x, {}) for x in X])
+                msg.info("Initialized the model with dummy data.")
+            except:
+                msg.fail(
+                    "Could not initialize the model: you'll have to provide a valid train_corpus argument in the config file.",
+                    exits=1,
+                )
+
     if print_settings.get("print_after_init"):
         msg.divider(f"STEP 1 - after initialization")
         _print_model(model, print_settings)
@@ -103,9 +121,18 @@ def debug_model(model: Model, *, print_settings: Optional[Dict[str, Any]] = None
     # STEP 2: Updating the model and printing again
     optimizer = Adam(0.001)
     set_dropout_rate(model, 0.2)
+    # ugly hack to deal with Tok2Vec listeners
+    tok2vec = None
+    if model.has_ref("tok2vec") and model.get_ref("tok2vec").name == "tok2vec-listener":
+        tok2vec = nlp.get_pipe("tok2vec")
+    goldY = None
     for e in range(3):
-        Y, get_dX = model.begin_update(_get_docs())
+        if tok2vec:
-        dY = get_gradient(model, Y)
+            tok2vec.predict(X)
+        Y, get_dX = model.begin_update(X)
+        if goldY is None:
+            goldY = _simulate_gold(Y)
+        dY = get_gradient(goldY, Y, model.ops)
         get_dX(dY)
         model.finish_update(optimizer)
     if print_settings.get("print_after_training"):
@@ -113,15 +140,25 @@ def debug_model(model: Model, *, print_settings: Optional[Dict[str, Any]] = None
         _print_model(model, print_settings)
 
     # STEP 3: the final prediction
-    prediction = model.predict(_get_docs())
+    prediction = model.predict(X)
     if print_settings.get("print_prediction"):
         msg.divider(f"STEP 3 - prediction")
         msg.info(str(prediction))
 
+    msg.good(f"Succesfully ended analysis - model looks good.")
 
-def get_gradient(model, Y):
+
-    goldY = _get_output(model.ops.xp)
+def get_gradient(goldY, Y, ops):
-    return Y - goldY
+    return ops.asarray(Y) - ops.asarray(goldY)
+
+
+def _simulate_gold(element, counter=1):
+    if isinstance(element, Iterable):
+        for i in range(len(element)):
+            element[i] = _simulate_gold(element[i], counter + i)
+        return element
+    else:
+        return 1 / counter
 
 
 def _sentences():
@@ -138,8 +175,13 @@ def _get_docs(lang: str = "en"):
     return list(nlp.pipe(_sentences()))
 
 
-def _get_output(xp):
+def _set_output_dim(model, nO):
-    return xp.asarray([i + 10 for i, _ in enumerate(_get_docs())], dtype="float32")
+    # simulating dim inference by directly setting the nO argument of the model
+    if model.has_dim("nO") is None:
+        model.set_dim("nO", nO)
+    if model.has_ref("output_layer"):
+        if model.get_ref("output_layer").has_dim("nO") is None:
+            model.get_ref("output_layer").set_dim("nO", nO)
 
 
 def _print_model(model, print_settings):

spacy/errors.py

@@ -545,7 +545,8 @@ class Errors:
     E949 = ("Can only create an alignment when the texts are the same.")
     E952 = ("The section '{name}' is not a valid section in the provided config.")
     E953 = ("Mismatched IDs received by the Tok2Vec listener: {id1} vs. {id2}")
-    E954 = ("The Tok2Vec listener did not receive a valid input.")
+    E954 = ("The Tok2Vec listener did not receive any valid input from an upstream "
+            "component.")
     E955 = ("Can't find table(s) '{table}' for language '{lang}' in spacy-lookups-data.")
     E956 = ("Can't find component '{name}' in [components] block in the config. "
             "Available components: {opts}")