Merge branch 'main' into feature/damian/large_model_export

.github/workflows/integrations-check.yaml

@@ -65,11 +65,7 @@ jobs:
       - name: "Upgrade protobuf version"
         run: pip3 install --upgrade protobuf
       - name: "⚙️ Install dependencies"
-        run: pip3 install .[dev,torchvision,deepsparse,onnxruntime]
-      - name: "Install transformers integration"
-        run: sparseml.transformers.question_answering --help
-      - name: "Install yolov5 integration"
-        run: sparseml.yolov5.train --help
+        run: pip3 install .[dev,torchvision,deepsparse,onnxruntime,transformers,yolov5]
       - name: "🔬 Running integrations tests (cadence: pre-commit}})"
         run: make testinteg TARGETS=$TRANSFORMERS,$YOLOV5,$IMAGE_CLASSIFICATION
 

CODE_OF_CONDUCT.md

@@ -71,7 +71,7 @@ a project may be further defined and clarified by project maintainers.
 ## Enforcement
 
 Instances of abusive, harassing, or otherwise unacceptable behavior may be
-reported by contacting the project team at <community@neuralmagic.com>. All
+reported by contacting the project team using the [Neural Magic Contact Us Form](https://neuralmagic.com/contact). All
 complaints will be reviewed and investigated and will result in a response that
 is deemed necessary and appropriate to the circumstances. The project team is
 obligated to maintain confidentiality with regard to the reporter of an incident.

CONTRIBUTING.md

@@ -77,7 +77,7 @@ For documentation edits, include:
 
 ## Question or Problem
 
-Sign up or log in to our [**Deep Sparse Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). We are growing the community member by member and happy to see you there. Don’t forget to search through existing discussions to avoid duplication! Thanks!
+Sign up or log in to our [**Neural Magic Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). We are growing the community member by member and happy to see you there. Don’t forget to search through existing discussions to avoid duplication! Thanks!
 
 ## Developing SparseML
 

DEVELOPING.md

@@ -16,10 +16,10 @@ limitations under the License.
 
 # Developing SparseML
 
-SparseML is developed and tested using Python 3.7-3.9.
+SparseML is developed and tested using Python 3.7-3.10.
 To develop SparseML, you will also need the development dependencies and to follow the styling guidelines.
 
-Here's some details to get started.
+Here are some details to get started.
 
 ## Basic Commands
 

docker/Dockerfile

@@ -95,23 +95,19 @@ ARG MODE
 RUN \
     if [ -n "$BRANCH" ] ; then \
       echo Installing from BRANCH && \
-      $VENV/bin/pip install --no-cache-dir "./sparseml[onnxruntime,torchvision,ultralytics]"; \
+      $VENV/bin/pip install --no-cache-dir "./sparseml[onnxruntime,torchvision,transformers,yolov5,ultralytics]"; \
     elif [ "$MODE" = "nightly" ] ; then \
       if [ -z $VERSION] ; then \
-        $VENV/bin/pip install --no-cache-dir "sparseml-nightly[onnxruntime,torchvision,ultralytics]"; \
+        $VENV/bin/pip install --no-cache-dir "sparseml-nightly[onnxruntime,torchvision,transformers,yolov5,ultralytics]"; \
       else \
-        $VENV/bin/pip install --no-cache-dir "sparseml-nightly[onnxruntime,torchvision,ultralytics]==$VERSION"; \
+        $VENV/bin/pip install --no-cache-dir "sparseml-nightly[onnxruntime,torchvision,transformers,yolov5,ultralytics]==$VERSION"; \
       fi; \
     elif [ -z $VERSION] ; then \
-      $VENV/bin/pip install --no-cache-dir "sparseml[onnxruntime,torchvision,ultralytics]"; \
+      $VENV/bin/pip install --no-cache-dir "sparseml[onnxruntime,torchvision,transformers,yolov5,ultralytics]"; \
     else \
-      $VENV/bin/pip install --no-cache-dir "sparseml[onnxruntime,torchvision,ultralytics]==$VERSION"; \
+      $VENV/bin/pip install --no-cache-dir "sparseml[onnxruntime,torchvision,transformers,yolov5,ultralytics]==$VERSION"; \
     fi;
 
-RUN sparseml.transformers.question_answering --help \
-    && sparseml.yolov5.train --help \
-    && sparseml.ultralytics.train --help
-
 
 FROM cuda_builder AS container_branch_dev
 ARG VENV

docs/source/installation.md

@@ -16,7 +16,7 @@ limitations under the License.
 
 # Installation
 
-This repository is tested on Python 3.7-3.9, and Linux/Debian systems.
+This repository is tested on Python 3.7-3.10, and Linux/Debian systems.
 It is recommended to install in a [virtual environment](https://docs.python.org/3/library/venv.html) to keep your system in order.
 Currently supported ML Frameworks are the following: `torch>=1.1.0,<=1.8.0`, `tensorflow>=1.8.0,<=2.0.0`, `tensorflow.keras >= 2.2.0`.
 

research/mfac/README.md

@@ -102,4 +102,4 @@ and [gradual](https://github.com/neuralmagic/sparseml/blob/main/research/mfac/tu
 pruning with M-FAC.
 
 ## Need Help?
-For Neural Magic Support, sign up or log in to our [**Deep Sparse Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue.](https://github.com/neuralmagic/sparseml/issues)
+For Neural Magic Support, sign up or log in to our [**Neural Magic Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue.](https://github.com/neuralmagic/sparseml/issues)

research/mfac/tutorials/gradual_pruning_with_mfac.md

@@ -47,7 +47,7 @@ should be used.
 
 ## Need Help?
 
-For Neural Magic Support, sign up or log in to our [**Deep Sparse Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue.](https://github.com/neuralmagic/sparseml/issues)
+For Neural Magic Support, sign up or log in to our [**Neural Magic Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue.](https://github.com/neuralmagic/sparseml/issues)
 
 ## Setting Up
 
@@ -145,4 +145,4 @@ In this tutorial you applied both M-FAC and magnitude pruning with SparseML and
 their results. More information about M-FAC pruning and other tutorials can be found
 [here](https://github.com/neuralmagic/sparseml/blob/main/research/mfac).
 
-For Neural Magic Support, sign up or log in to our [**Deep Sparse Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue.](https://github.com/neuralmagic/sparseml/issues)
+For Neural Magic Support, sign up or log in to our [**Neural Magic Community Slack**](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue.](https://github.com/neuralmagic/sparseml/issues)

setup.py

@@ -66,11 +66,20 @@
     "gputils",
 ]
 _pytorch_all_deps = _pytorch_deps + [
-    "torchvision>=0.3.0,<=0.14",
+    "torchvision>=0.3.0,<0.15",
     "torchaudio<=0.13",
-    "torchvision>=0.3.0,<=0.14",
 ]
-_pytorch_vision_deps = _pytorch_deps + ["torchvision>=0.3.0,<=0.14"]
+_pytorch_vision_deps = _pytorch_deps + ["torchvision>=0.3.0,<0.15"]
+_transformers_deps = _pytorch_deps + [
+    f"{'nm-transformers' if is_release else 'nm-transformers-nightly'}"
+    f"~={version_nm_deps}",
+    "datasets<=1.18.4",
+    "scikit-learn",
+    "seqeval",
+]
+_yolov5_deps = _pytorch_vision_deps + [
+    f"{'nm-yolov5' if is_release else 'nm-yolov5-nightly'}~={version_nm_deps}"
+]
 _tensorflow_v1_deps = ["tensorflow<2.0.0", "tensorboard<2.0.0", "tf2onnx>=1.0.0,<1.6"]
 _tensorflow_v1_gpu_deps = [
     "tensorflow-gpu<2.0.0",
@@ -132,10 +141,12 @@ def _setup_extras() -> Dict:
         "torch": _pytorch_deps,
         "torch_all": _pytorch_all_deps,
         "torchvision": _pytorch_vision_deps,
+        "transformers": _transformers_deps,
         "tf_v1": _tensorflow_v1_deps,
         "tf_v1_gpu": _tensorflow_v1_gpu_deps,
         "tf_keras": _keras_deps,
         "ultralytics": _ultralytics_deps,
+        "yolov5": _yolov5_deps,
     }
 
 

src/sparseml/pytorch/image_classification/README_image_classification.md

@@ -158,7 +158,7 @@ To learn more, refer to the [appropriate documentation in the DeepSparse reposit
 
 ## Support
 
-For Neural Magic Support, sign up or log in to our [Deep Sparse Community Slack](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue](https://github.com/neuralmagic/sparseml/issues).
+For Neural Magic Support, sign up or log in to our [Neural Magic Community Slack](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ). Bugs, feature requests, or additional questions can also be posted to our [GitHub Issue Queue](https://github.com/neuralmagic/sparseml/issues).
 
 
 [torch]: https://pytorch.org/

src/sparseml/pytorch/sparsification/quantization/quantize_qat_export.py

@@ -1762,7 +1762,7 @@ def _propagate_mobilebert_embedding_quantization(model: ModelProto):
             continue
 
         embedding_array = numpy_helper.to_array(embedding_initializer)
-        if embedding_array.dtype != numpy.uint8:
+        if embedding_array.dtype not in [numpy.uint8, numpy.int8]:
             continue
 
         dequant_node = graph.get_node_single_child(gather_node)
@@ -1806,12 +1806,15 @@ def _propagate_mobilebert_embedding_quantization(model: ModelProto):
         # switch position of dequantize node
         for branch_node in graph.get_node_children(dequant_node):
             if branch_node.op_type == "Slice":
+                zero_point = graph.get_init_by_name(dequant_node.input[2])
+                zero_point_array = numpy_helper.to_array(zero_point)
                 branch_node.input[0] = gather_node.output[0]
                 pad_node = graph.get_node_single_child(branch_node)
                 pad_value = graph.get_init_by_name(pad_node.input[2])
                 pad_value_array = numpy_helper.to_array(pad_value)
-                pad_value_array = pad_value_array + 128
-                pad_value_array = pad_value_array.astype(numpy.uint8)
+                pad_value_array = (
+                    pad_value_array.astype(zero_point_array.dtype) + zero_point_array
+                )
                 model.graph.initializer.remove(pad_value)
                 pad_value = numpy_helper.from_array(
                     pad_value_array, name=pad_value.name

src/sparseml/pytorch/torchvision/train.py

@@ -267,6 +267,8 @@ def load_data(traindir, valdir, args):
             traindir,
             presets.ClassificationPresetTrain(
                 crop_size=train_crop_size,
+                mean=args.rgb_mean,
+                std=args.rgb_std,
                 interpolation=interpolation,
                 auto_augment_policy=auto_augment_policy,
                 random_erase_prob=random_erase_prob,
@@ -289,6 +291,8 @@ def load_data(traindir, valdir, args):
     else:
         preprocessing = presets.ClassificationPresetEval(
             crop_size=val_crop_size,
+            mean=args.rgb_mean,
+            std=args.rgb_std,
             resize_size=val_resize_size,
             interpolation=interpolation,
         )
@@ -1212,6 +1216,26 @@ def new_func(*args, **kwargs):
         "Note: Will be read from the checkpoint if not specified"
     ),
 )
+@click.option(
+    "--rgb-mean",
+    nargs=3,
+    default=(0.485, 0.456, 0.406),
+    type=float,
+    help=(
+        "RGB mean values used to shift input RGB values; "
+        "Note: Will use ImageNet values if not specified."
+    ),
+)
+@click.option(
+    "--rgb-std",
+    default=(0.229, 0.224, 0.225),
+    nargs=3,
+    type=float,
+    help=(
+        "RGB standard-deviation values used to normalize input RGB values; "
+        "Note: Will use ImageNet values if not specified."
+    ),
+)
 @click.pass_context
 def cli(ctx, **kwargs):
     """

src/sparseml/transformers/__init__.py

@@ -20,111 +20,20 @@
 
 import logging as _logging
 
-import pkg_resources
 from sparseml.analytics import sparseml_analytics as _analytics
 
 
 _analytics.send_event("python__transformers__init")
 
-_EXPECTED_VERSION = "4.23.1"
-
 
 _LOGGER = _logging.getLogger(__name__)
-_NM_TRANSFORMERS_TAR_TEMPLATE = (
-    "https://github.com/neuralmagic/transformers/releases/download/"
-    "{version}/transformers-4.23.1-py3-none-any.whl"
-)
-_NM_TRANSFORMERS_NIGHTLY = _NM_TRANSFORMERS_TAR_TEMPLATE.format(version="nightly")
-
-
-def _install_transformers_and_deps():
-
-    import subprocess as _subprocess
-    import sys as _sys
-
-    import sparseml as _sparseml
-
-    nm_transformers_release = (
-        "nightly" if not _sparseml.is_release else f"v{_sparseml.version_major_minor}"
-    )
-    transformers_requirement = _NM_TRANSFORMERS_TAR_TEMPLATE.format(
-        version=nm_transformers_release
-    )
-    try:
-        _subprocess.check_call(
-            [
-                _sys.executable,
-                "-m",
-                "pip",
-                "install",
-                transformers_requirement,
-                "datasets<=1.18.4",
-                "scikit-learn",
-                "seqeval",
-            ]
-        )
-
-        import transformers as _transformers
-
-        _LOGGER.info("sparseml-transformers and dependencies successfully installed")
-    except Exception:
-        raise ValueError(
-            "Unable to install and import sparseml-transformers dependencies check "
-            "that transformers is installed, if not, install via "
-            f"`pip install {_NM_TRANSFORMERS_NIGHTLY}`"
-        )
 
 
 def _check_transformers_install():
-    transformers_version = next(
-        (
-            pkg.version
-            for pkg in pkg_resources.working_set
-            if pkg.project_name.lower() == "transformers"
-        ),
-        None,
-    )
-
-    # Either no transformers install is found or wrong version installed
-    if transformers_version != _EXPECTED_VERSION:
-        import os
-
-        if os.getenv("NM_NO_AUTOINSTALL_TRANSFORMERS", False):
-            _LOGGER.warning(
-                "Unable to import, skipping auto installation "
-                "due to NM_NO_AUTOINSTALL_TRANSFORMERS"
-            )
-            # skip any further checks
-            return
-        else:
-            _LOGGER.warning(
-                f"sparseml-transformers v{_EXPECTED_VERSION} installation not "
-                f"detected. Installing  sparseml-transformers v{_EXPECTED_VERSION} "
-                "dependencies if transformers is already  installed in the "
-                "environment, it will be overwritten. Set  environment variable "
-                "NM_NO_AUTOINSTALL_TRANSFORMERS to disable"
-            )
-            _install_transformers_and_deps()
-
-    else:
-        import transformers as _transformers
-
-        # Edge case where user has expected version of transformers installed, but
-        # not the nm integrated one
-        if not _transformers.NM_INTEGRATED:
-            _install_transformers_and_deps()
-            raise RuntimeError(
-                "Installed transformers package has been overwritten with "
-                "sparseml-transformers. Stopping process as this is likely to cause "
-                "import issues. Please re-run command"
-            )
-
-    # re check import after potential install
-    try:
-        import transformers as _transformers
+    # check for NM integration in transformers version
+    import transformers as _transformers
 
-        assert _transformers.NM_INTEGRATED
-    except Exception:
+    if not _transformers.NM_INTEGRATED:
         _LOGGER.warning(
             "the neuralmagic fork of transformers may not be installed. it can be "
             "installed via "

src/sparseml/transformers/export.py

@@ -178,7 +178,7 @@ def load_task_dataset(
 
         data_training_args = DataTrainingArguments(**data_args)
         return get_tokenized_token_classification_dataset(
-            data_args=data_training_args, tokenizer=tokenizer, model=model
+            data_args=data_training_args, tokenizer=tokenizer, model=model or config
         )
 
     if (

src/sparseml/transformers/masked_language_modeling.py

@@ -51,7 +51,6 @@
     set_seed,
 )
 from transformers.trainer_utils import get_last_checkpoint
-from transformers.utils import check_min_version
 from transformers.utils.versions import require_version
 
 from sparseml.pytorch.utils.distributed import record
@@ -66,10 +65,6 @@
 ]
 
 
-# Will error if the minimal version of Transformers is not installed.
-# Remove at your own risks
-check_min_version("4.18.0.dev0")
-
 require_version(
     "datasets>=1.18.0",
     "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt",

src/sparseml/transformers/question_answering.py

@@ -45,7 +45,6 @@
     set_seed,
 )
 from transformers.trainer_utils import get_last_checkpoint
-from transformers.utils import check_min_version
 from transformers.utils.versions import require_version
 
 from sparseml.pytorch.utils.distributed import record
@@ -57,10 +56,6 @@
 from sparseml.transformers.utils import SparseAutoModel, get_shared_tokenizer_src
 
 
-# Will error if the minimal version of Transformers is not installed. Remove at your
-# own risks.
-check_min_version("4.18.0.dev0")
-
 require_version(
     "datasets>=1.18.0",
     "To fix: pip install -r examples/pytorch/question-answering/requirements.txt",