Fix bugs

Make startup faster, preinstall apt packages in docker
Revert SMT model zip from tar.gz to zip

dockerfile

@@ -40,7 +40,9 @@ RUN apt-get update && \
     apt-get install --no-install-recommends -y \
     curl \
     python$PYTHON_VERSION \
-    python$PYTHON_VERSION-distutils && \
+    python$PYTHON_VERSION-distutils \
+# these are needed for ClearML
+    git libsm6 libxext6 libxrender-dev libglib2.0-0 && \
     rm -rf /var/lib/apt/lists/* && \
     apt-get clean
 

dockerfile.cpu_only

@@ -26,6 +26,11 @@ RUN poetry export --with=gpu --without-hashes -f requirements.txt > requirements
 FROM python:$PYTHON_VERSION
 WORKDIR /root
 
+# these are needed for ClearML
+RUN apt-get update && \
+    apt-get install --no-install-recommends -y \
+    git libsm6 libxext6 libxrender-dev libglib2.0-0 && \
+
 COPY --from=builder /src/requirements.txt .
 RUN --mount=type=cache,target=/root/.cache \
     pip install --no-cache-dir -r requirements.txt && rm requirements.txt

machine/jobs/nmt_engine_build_job.py

@@ -70,7 +70,6 @@ def run(
             ) as model_trainer:
                 model_trainer.train(progress=phase_progress, check_canceled=check_canceled)
                 model_trainer.save()
-                train_corpus_size = model_trainer.stats.train_corpus_size
         else:
             logger.info("No matching entries in the source and target corpus - skipping training")
 
@@ -99,7 +98,7 @@ def run(
             logger.info("Saving model")
             model_path = self._nmt_model_factory.save_model()
             self._shared_file_service.save_model(model_path, self._config.save_model + "".join(model_path.suffixes))
-        return train_corpus_size
+        return parallel_corpus_size
 
 
 def _translate_batch(

machine/jobs/smt_engine_build_job.py

@@ -40,17 +40,11 @@ def run(
 
         with self._shared_file_service.get_source_pretranslations() as src_pretranslations:
             inference_step_count = sum(1 for _ in src_pretranslations)
-        if inference_step_count > 0:
-            phases = [
-                Phase(message="Training SMT model", percentage=0.85),
-                Phase(message="Training truecaser", percentage=0.05),
-                Phase(message="Pretranslating segments", percentage=0.1),
-            ]
-        else:
-            phases = [
-                Phase(message="Training SMT model", percentage=0.95),
-                Phase(message="Training truecaser", percentage=0.05),
-            ]
+        phases = [
+            Phase(message="Training SMT model", percentage=0.85),
+            Phase(message="Training truecaser", percentage=0.05),
+            Phase(message="Pretranslating segments", percentage=0.1),
+        ]
         progress_reporter = PhasedProgressReporter(progress, phases)
 
         if check_canceled is not None:
@@ -73,23 +67,22 @@ def run(
         if check_canceled is not None:
             check_canceled()
 
-        if inference_step_count > 0:
-            with ExitStack() as stack:
-                detokenizer = self._smt_model_factory.create_detokenizer()
-                truecaser = self._smt_model_factory.create_truecaser()
-                phase_progress = stack.enter_context(progress_reporter.start_next_phase())
-                engine = stack.enter_context(self._smt_model_factory.create_engine(tokenizer, detokenizer, truecaser))
-                src_pretranslations = stack.enter_context(self._shared_file_service.get_source_pretranslations())
-                writer = stack.enter_context(self._shared_file_service.open_target_pretranslation_writer())
-                current_inference_step = 0
+        with ExitStack() as stack:
+            detokenizer = self._smt_model_factory.create_detokenizer()
+            truecaser = self._smt_model_factory.create_truecaser()
+            phase_progress = stack.enter_context(progress_reporter.start_next_phase())
+            engine = stack.enter_context(self._smt_model_factory.create_engine(tokenizer, detokenizer, truecaser))
+            src_pretranslations = stack.enter_context(self._shared_file_service.get_source_pretranslations())
+            writer = stack.enter_context(self._shared_file_service.open_target_pretranslation_writer())
+            current_inference_step = 0
+            phase_progress(ProgressStatus.from_step(current_inference_step, inference_step_count))
+            batch_size = self._config["pretranslation_batch_size"]
+            for pi_batch in batch(src_pretranslations, batch_size):
+                if check_canceled is not None:
+                    check_canceled()
+                _translate_batch(engine, pi_batch, writer)
+                current_inference_step += len(pi_batch)
                 phase_progress(ProgressStatus.from_step(current_inference_step, inference_step_count))
-                batch_size = self._config["pretranslation_batch_size"]
-                for pi_batch in batch(src_pretranslations, batch_size):
-                    if check_canceled is not None:
-                        check_canceled()
-                    _translate_batch(engine, pi_batch, writer)
-                    current_inference_step += len(pi_batch)
-                    phase_progress(ProgressStatus.from_step(current_inference_step, inference_step_count))
 
         if "save_model" in self._config and self._config.save_model is not None:
             logger.info("Saving model")

machine/jobs/thot/thot_smt_model_factory.py

@@ -89,10 +89,10 @@ def create_truecaser(self) -> Truecaser:
         return UnigramTruecaser(model_path=self._model_dir / "unigram-casing-model.txt")
 
     def save_model(self) -> Path:
-        tar_file_path = Path(
+        tar_file_basename = Path(
             self._config.data_dir, self._config.shared_file_folder, "builds", self._config.build_id, "model"
         )
-        return Path(shutil.make_archive(str(tar_file_path), "gztar", self._model_dir))
+        return Path(shutil.make_archive(str(tar_file_basename), "zip", self._model_dir))
 
     @property
     def _model_dir(self) -> Path: