🚀 CI: Nightly Build

.github/PULL_REQUEST_TEMPLATE/pull_request_template.md → .github/pull_request_template.md

@@ -11,7 +11,7 @@
 - [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
 - [ ] This change requires a documentation update
 
-# Checklist:
+## Checklist
 
 - [ ] My code follows the [pre-commit style and check guidelines](https://openvinotoolkit.github.io/anomalib/guides/using_pre_commit.html#pre-commit-hooks) of this project.
 - [ ] I have performed a self-review of my code

.github/workflows/nightly.yml

@@ -0,0 +1,30 @@
+name: Nightly-regression Test
+
+on:
+  workflow_dispatch: # run on request (no need for PR)
+  schedule:
+    - cron: "0 0 * * *"
+
+jobs:
+  Tox:
+    runs-on: [self-hosted, linux, x64]
+    strategy:
+      max-parallel: 1
+    if: github.ref == 'refs/heads/development'
+    steps:
+      - name: Print GPU status
+        run: nvidia-smi
+      - name: CHECKOUT REPOSITORY
+        uses: actions/checkout@v2
+      - name: Install Tox
+        run: pip install tox
+      - name: Coverage
+        run: |
+          export ANOMALIB_DATASET_PATH=/media/data1/datasets/MVTec
+          export CUDA_VISIBLE_DEVICES=2
+          tox -e nightly
+      - name: Upload coverage result
+        uses: actions/upload-artifact@v2
+        with:
+          name: coverage
+          path: .tox/coverage.xml

.github/workflows/tox.yml → .github/workflows/pre_merge.yml

@@ -24,7 +24,7 @@ jobs:
         run: |
           export ANOMALIB_DATASET_PATH=/media/data1/datasets/MVTec
           export CUDA_VISIBLE_DEVICES=3
-          tox -e coverage
+          tox -e pre_merge
       - name: Upload coverage result
         uses: actions/upload-artifact@v2
         with:

anomalib/loggers/wandb.py

@@ -17,11 +17,12 @@
 from typing import Any, List, Optional, Union
 
 import numpy as np
-import wandb
 from matplotlib.figure import Figure
 from pytorch_lightning.loggers.wandb import WandbLogger
 from pytorch_lightning.utilities import rank_zero_only
 
+import wandb
+
 from .base import ImageLoggerBase
 
 

anomalib/utils/hpo/__init__.py

@@ -0,0 +1,15 @@
+"""Utils to help in HPO search."""
+
+# Copyright (C) 2020 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions
+# and limitations under the License.

anomalib/utils/hpo/config.py

@@ -0,0 +1,53 @@
+"""Utils to update configuration files."""
+
+# Copyright (C) 2020 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions
+# and limitations under the License.
+
+from typing import List
+
+from omegaconf import DictConfig
+
+
+def flatten_sweep_params(params_dict: DictConfig) -> DictConfig:
+    """Flatten the nested parameters section of the config object.
+
+    Args:
+        params_dict: DictConfig: The dictionary containing the hpo parameters in the original, nested, structure.
+
+    Returns:
+        flattened version of the parameter dictionary.
+    """
+
+    def process_params(nested_params: DictConfig, keys: List[str], flattened_params: DictConfig):
+        """Flatten nested dictionary.
+
+        Recursive helper function that traverses the nested config object and stores the leaf nodes in a flattened
+        dictionary.
+
+        Args:
+            nested_params: DictConfig: config object containing the original parameters.
+            keys: List[str]: list of keys leading to the current location in the config.
+            flattened_params: DictConfig: Dictionary in which the flattened parameters are stored.
+        """
+        for name, cfg in nested_params.items():
+            if isinstance(cfg, DictConfig):
+                process_params(cfg, keys + [str(name)], flattened_params)
+            else:
+                key = ".".join(keys + [str(name)])
+                flattened_params[key] = cfg
+
+    flattened_params_dict = DictConfig({})
+    process_params(params_dict, [], flattened_params_dict)
+
+    return flattened_params_dict

pyproject.toml

@@ -4,6 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [tool.isort]
 profile = "black"
+known_first_party = "wandb"
 sections = ['FUTURE', 'STDLIB', 'THIRDPARTY', 'FIRSTPARTY', 'LOCALFOLDER']
 
 [tool.black]

requirements/base.txt

@@ -7,7 +7,7 @@ kornia==0.5.6
 lxml==4.6.5
 matplotlib==3.4.3
 networkx~=2.5
-nncf==2.0.0
+nncf==2.1.0
 numpy~=1.19.5
 omegaconf==2.1.1
 pillow==9.0.0

tests/helpers/dataset.py

@@ -82,6 +82,7 @@ def __init__(
         test_shapes: List[str] = ["hexagon", "star"],
         path: Union[str, Path] = "./datasets/MVTec",
         use_mvtec: bool = False,
+        seed: int = 0,
     ) -> None:
         """Creates a context for Generating Dummy Dataset. Useful for wrapping test functions.
         NOTE: for MVTec dataset it does not return a category.
@@ -97,6 +98,7 @@ def __init__(
             test_shapes (List[str], optional): List of anomalous shapes. Defaults to ["triangle", "ellipse"].
             path (Union[str, Path], optional): Path to MVTec dataset. Defaults to "./datasets/MVTec".
             use_mvtec (bool, optional): Use MVTec dataset or dummy dataset. Defaults to False.
+            seed (int, optional): Fixes seed if any number greater than 0 is provided. 0 means no seed. Defaults to 0.
 
         Example:
             >>> @TestDataset
@@ -112,6 +114,7 @@ def __init__(
         self.test_shapes = test_shapes
         self.path = path
         self.use_mvtec = use_mvtec
+        self.seed = seed
 
     def __call__(self, func):
         @wraps(func)
@@ -129,6 +132,7 @@ def inner(*args, **kwds):
                     train_shapes=self.train_shapes,
                     test_shapes=self.test_shapes,
                     max_size=self.max_size,
+                    seed=self.seed,
                 ) as dataset_path:
                     kwds["category"] = "shapes"
                     return func(*args, path=dataset_path, **kwds)
@@ -150,6 +154,7 @@ class GeneratedDummyDataset(ContextDecorator):
             max_size (Optional[int], optional): Maximum size of the test shapes. Defaults to 10.
             train_shapes (List[str], optional): List of good shapes. Defaults to ["circle", "rectangle"].
             test_shapes (List[str], optional): List of anomalous shapes. Defaults to ["triangle", "ellipse"].
+            seed (int, optional): Fixes seed if any number greater than 0 is provided. 0 means no seed. Defaults to 0.
     """
 
     def __init__(
@@ -161,6 +166,7 @@ def __init__(
         max_size: Optional[int] = 10,
         train_shapes: List[str] = ["triangle", "rectangle"],
         test_shapes: List[str] = ["star", "hexagon"],
+        seed: int = 0,
     ) -> None:
         self.root_dir = mkdtemp()
         self.num_train = num_train
@@ -170,6 +176,7 @@ def __init__(
         self.image_height = img_height
         self.image_width = img_width
         self.max_size = max_size
+        self.seed = seed
 
     def _generate_dataset(self):
         """Generates dummy dataset in a temporary directory using the same
@@ -226,6 +233,8 @@ def _generate_dataset(self):
 
     def __enter__(self):
         """Creates the dataset in temp folder."""
+        if self.seed > 0:
+            np.random.seed(self.seed)
         self._generate_dataset()
         return self.root_dir
 

tests/helpers/model.py

@@ -0,0 +1,145 @@
+"""Common helpers for both nightly and pre-merge model tests."""
+
+# Copyright (C) 2020 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions
+# and limitations under the License.
+
+import os
+from typing import Dict, Tuple, Union
+
+import numpy as np
+from omegaconf import DictConfig, ListConfig
+from pytorch_lightning import LightningDataModule, Trainer
+from pytorch_lightning.callbacks import ModelCheckpoint
+
+from anomalib.config import get_configurable_parameters, update_nncf_config
+from anomalib.core.callbacks import get_callbacks
+from anomalib.core.callbacks.visualizer_callback import VisualizerCallback
+from anomalib.core.model.anomaly_module import AnomalyModule
+from anomalib.data import get_datamodule
+from anomalib.models import get_model
+
+
+def setup_model_train(
+    model_name: str,
+    dataset_path: str,
+    project_path: str,
+    nncf: bool,
+    category: str,
+    score_type: str = None,
+    weight_file: str = "weights/last.ckpt",
+    fast_run: bool = False,
+) -> Tuple[Union[DictConfig, ListConfig], LightningDataModule, AnomalyModule, Trainer]:
+    """Train the model based on the parameters passed.
+
+    Args:
+        model_name (str): Name of the model to train.
+        dataset_path (str): Location of the dataset.
+        project_path (str): Path to temporary project folder.
+        nncf (bool): Add nncf callback.
+        category (str): Category to train on.
+        score_type (str, optional): Only used for DFM. Defaults to None.
+        weight_file (str, optional): Path to weight file.
+        fast_run (bool, optional): If set to true, the model trains for only 1 epoch. We train for one epoch as
+            this ensures that both anomalous and non-anomalous images are present in the validation step.
+
+    Returns:
+        Tuple[DictConfig, LightningDataModule, AnomalyModule, Trainer]: config, datamodule, trained model, trainer
+    """
+    config = get_configurable_parameters(model_name=model_name)
+    if score_type is not None:
+        config.model.score_type = score_type
+    config.project.seed = 42
+    config.dataset.category = category
+    config.dataset.path = dataset_path
+    config.project.log_images_to = []
+
+    # If weight file is empty, remove the key from config
+    if "weight_file" in config.model.keys() and weight_file == "":
+        config.model.pop("weight_file")
+    else:
+        config.model.weight_file = weight_file
+
+    if nncf:
+        config.optimization.nncf.apply = True
+        config = update_nncf_config(config)
+        config.init_weights = None
+
+    # reassign project path as config is updated in `update_config_for_nncf`
+    config.project.path = project_path
+
+    datamodule = get_datamodule(config)
+    model = get_model(config)
+
+    callbacks = get_callbacks(config)
+
+    # Force model checkpoint to create checkpoint after first epoch
+    if fast_run == True:
+        for index, callback in enumerate(callbacks):
+            if isinstance(callback, ModelCheckpoint):
+                callbacks.pop(index)
+                break
+        model_checkpoint = ModelCheckpoint(
+            dirpath=os.path.join(config.project.path, "weights"),
+            filename="last",
+            monitor=None,
+            mode="max",
+            save_last=True,
+            auto_insert_metric_name=False,
+        )
+        callbacks.append(model_checkpoint)
+
+    for index, callback in enumerate(callbacks):
+        if isinstance(callback, VisualizerCallback):
+            callbacks.pop(index)
+            break
+
+    # Train the model.
+    if fast_run:
+        config.trainer.max_epochs = 1
+
+    trainer = Trainer(callbacks=callbacks, **config.trainer)
+    trainer.fit(model=model, datamodule=datamodule)
+    return config, datamodule, model, trainer
+
+
+def model_load_test(config: Union[DictConfig, ListConfig], datamodule: LightningDataModule, results: Dict):
+    """Create a new model based on the weights specified in config.
+
+    Args:
+        config ([Union[DictConfig, ListConfig]): Model config.
+        datamodule (LightningDataModule): Dataloader
+        results (Dict): Results from original model.
+
+    """
+    loaded_model = get_model(config)  # get new model
+
+    callbacks = get_callbacks(config)
+
+    for index, callback in enumerate(callbacks):
+        # Remove visualizer callback as saving results takes time
+        if isinstance(callback, VisualizerCallback):
+            callbacks.pop(index)
+            break
+
+    # create new trainer object with LoadModel callback (assumes it is present)
+    trainer = Trainer(callbacks=callbacks, **config.trainer)
+    # Assumes the new model has LoadModel callback and the old one had ModelCheckpoint callback
+    new_results = trainer.test(model=loaded_model, datamodule=datamodule)[0]
+    assert np.isclose(
+        results["image_AUROC"], new_results["image_AUROC"]
+    ), "Loaded model does not yield close performance results"
+    if config.dataset.task == "segmentation":
+        assert np.isclose(
+            results["pixel_AUROC"], new_results["pixel_AUROC"]
+        ), "Loaded model does not yield close performance results"