├── .gitignore ├── CODE_OF_CONDUCT.md ├── CONTRIBUTING.md ├── LICENSE ├── NOTICE ├── README.md ├── analyze_model_predictions.ipynb ├── code ├── custom_hook.py ├── pretrained_model.py ├── pretrained_model_with_debugger_hook.py └── requirements.txt ├── docker ├── Dockerfile └── evaluation.py ├── images ├── image.png └── screenshot.png ├── model └── model.pt └── utils.py /.gitignore: -------------------------------------------------------------------------------- 1 | #### Project Specific: 2 | 3 | advbox/ 4 | adversarial_examples/ 5 | GTSRB/ 6 | **.csv 7 | **.tar.gz 8 | **.zip 9 | 10 | #### Python Generic: 11 | 12 | # Byte-compiled / optimized / DLL files 13 | __pycache__/ 14 | *.py[cod] 15 | *$py.class 16 | 17 | # C extensions 18 | *.so 19 | 20 | # Distribution / packaging 21 | .Python 22 | build/ 23 | develop-eggs/ 24 | dist/ 25 | downloads/ 26 | eggs/ 27 | .eggs/ 28 | lib/ 29 | lib64/ 30 | parts/ 31 | sdist/ 32 | var/ 33 | wheels/ 34 | pip-wheel-metadata/ 35 | share/python-wheels/ 36 | *.egg-info/ 37 | .installed.cfg 38 | *.egg 39 | MANIFEST 40 | 41 | # PyInstaller 42 | # Usually these files are written by a python script from a template 43 | # before PyInstaller builds the exe, so as to inject date/other infos into it. 44 | *.manifest 45 | *.spec 46 | 47 | # Installer logs 48 | pip-log.txt 49 | pip-delete-this-directory.txt 50 | 51 | # Unit test / coverage reports 52 | htmlcov/ 53 | .tox/ 54 | .nox/ 55 | .coverage 56 | .coverage.* 57 | .cache 58 | nosetests.xml 59 | coverage.xml 60 | *.cover 61 | *.py,cover 62 | .hypothesis/ 63 | .pytest_cache/ 64 | 65 | # Translations 66 | *.mo 67 | *.pot 68 | 69 | # Django stuff: 70 | *.log 71 | local_settings.py 72 | db.sqlite3 73 | db.sqlite3-journal 74 | 75 | # Flask stuff: 76 | instance/ 77 | .webassets-cache 78 | 79 | # Scrapy stuff: 80 | .scrapy 81 | 82 | # Sphinx documentation 83 | docs/_build/ 84 | 85 | # PyBuilder 86 | target/ 87 | 88 | # Jupyter Notebook 89 | .ipynb_checkpoints 90 | 91 | # IPython 92 | profile_default/ 93 | ipython_config.py 94 | 95 | # pyenv 96 | .python-version 97 | 98 | # pipenv 99 | # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. 100 | # However, in case of collaboration, if having platform-specific dependencies or dependencies 101 | # having no cross-platform support, pipenv may install dependencies that don't work, or not 102 | # install all needed dependencies. 103 | #Pipfile.lock 104 | 105 | # PEP 582; used by e.g. github.com/David-OConnor/pyflow 106 | __pypackages__/ 107 | 108 | # Celery stuff 109 | celerybeat-schedule 110 | celerybeat.pid 111 | 112 | # SageMath parsed files 113 | *.sage.py 114 | 115 | # Environments 116 | .env 117 | .venv 118 | env/ 119 | venv/ 120 | ENV/ 121 | env.bak/ 122 | venv.bak/ 123 | 124 | # Spyder project settings 125 | .spyderproject 126 | .spyproject 127 | 128 | # Rope project settings 129 | .ropeproject 130 | 131 | # mkdocs documentation 132 | /site 133 | 134 | # mypy 135 | .mypy_cache/ 136 | .dmypy.json 137 | dmypy.json 138 | 139 | # Pyre type checker 140 | .pyre/ 141 | -------------------------------------------------------------------------------- /CODE_OF_CONDUCT.md: -------------------------------------------------------------------------------- 1 | ## Code of Conduct 2 | This project has adopted the [Amazon Open Source Code of Conduct](https://aws.github.io/code-of-conduct). 3 | For more information see the [Code of Conduct FAQ](https://aws.github.io/code-of-conduct-faq) or contact 4 | opensource-codeofconduct@amazon.com with any additional questions or comments. 5 | -------------------------------------------------------------------------------- /CONTRIBUTING.md: -------------------------------------------------------------------------------- 1 | # Contributing Guidelines 2 | 3 | Thank you for your interest in contributing to our project. Whether it's a bug report, new feature, correction, or additional 4 | documentation, we greatly value feedback and contributions from our community. 5 | 6 | Please read through this document before submitting any issues or pull requests to ensure we have all the necessary 7 | information to effectively respond to your bug report or contribution. 8 | 9 | 10 | ## Reporting Bugs/Feature Requests 11 | 12 | We welcome you to use the GitHub issue tracker to report bugs or suggest features. 13 | 14 | When filing an issue, please check existing open, or recently closed, issues to make sure somebody else hasn't already 15 | reported the issue. Please try to include as much information as you can. Details like these are incredibly useful: 16 | 17 | * A reproducible test case or series of steps 18 | * The version of our code being used 19 | * Any modifications you've made relevant to the bug 20 | * Anything unusual about your environment or deployment 21 | 22 | 23 | ## Contributing via Pull Requests 24 | Contributions via pull requests are much appreciated. Before sending us a pull request, please ensure that: 25 | 26 | 1. You are working against the latest source on the *master* branch. 27 | 2. You check existing open, and recently merged, pull requests to make sure someone else hasn't addressed the problem already. 28 | 3. You open an issue to discuss any significant work - we would hate for your time to be wasted. 29 | 30 | To send us a pull request, please: 31 | 32 | 1. Fork the repository. 33 | 2. Modify the source; please focus on the specific change you are contributing. If you also reformat all the code, it will be hard for us to focus on your change. 34 | 3. Ensure local tests pass. 35 | 4. Commit to your fork using clear commit messages. 36 | 5. Send us a pull request, answering any default questions in the pull request interface. 37 | 6. Pay attention to any automated CI failures reported in the pull request, and stay involved in the conversation. 38 | 39 | GitHub provides additional document on [forking a repository](https://help.github.com/articles/fork-a-repo/) and 40 | [creating a pull request](https://help.github.com/articles/creating-a-pull-request/). 41 | 42 | 43 | ## Finding contributions to work on 44 | Looking at the existing issues is a great way to find something to contribute on. As our projects, by default, use the default GitHub issue labels (enhancement/bug/duplicate/help wanted/invalid/question/wontfix), looking at any 'help wanted' issues is a great place to start. 45 | 46 | 47 | ## Code of Conduct 48 | This project has adopted the [Amazon Open Source Code of Conduct](https://aws.github.io/code-of-conduct). 49 | For more information see the [Code of Conduct FAQ](https://aws.github.io/code-of-conduct-faq) or contact 50 | opensource-codeofconduct@amazon.com with any additional questions or comments. 51 | 52 | 53 | ## Security issue notifications 54 | If you discover a potential security issue in this project we ask that you notify AWS/Amazon Security via our [vulnerability reporting page](http://aws.amazon.com/security/vulnerability-reporting/). Please do **not** create a public github issue. 55 | 56 | 57 | ## Licensing 58 | 59 | See the [LICENSE](LICENSE) file for our project's licensing. We will ask you to confirm the licensing of your contribution. 60 | 61 | We may ask you to sign a [Contributor License Agreement (CLA)](http://en.wikipedia.org/wiki/Contributor_License_Agreement) for larger changes. 62 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | 2 | Apache License 3 | Version 2.0, January 2004 4 | http://www.apache.org/licenses/ 5 | 6 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 7 | 8 | 1. Definitions. 9 | 10 | "License" shall mean the terms and conditions for use, reproduction, 11 | and distribution as defined by Sections 1 through 9 of this document. 12 | 13 | "Licensor" shall mean the copyright owner or entity authorized by 14 | the copyright owner that is granting the License. 15 | 16 | "Legal Entity" shall mean the union of the acting entity and all 17 | other entities that control, are controlled by, or are under common 18 | control with that entity. For the purposes of this definition, 19 | "control" means (i) the power, direct or indirect, to cause the 20 | direction or management of such entity, whether by contract or 21 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 22 | outstanding shares, or (iii) beneficial ownership of such entity. 23 | 24 | "You" (or "Your") shall mean an individual or Legal Entity 25 | exercising permissions granted by this License. 26 | 27 | "Source" form shall mean the preferred form for making modifications, 28 | including but not limited to software source code, documentation 29 | source, and configuration files. 30 | 31 | "Object" form shall mean any form resulting from mechanical 32 | transformation or translation of a Source form, including but 33 | not limited to compiled object code, generated documentation, 34 | and conversions to other media types. 35 | 36 | "Work" shall mean the work of authorship, whether in Source or 37 | Object form, made available under the License, as indicated by a 38 | copyright notice that is included in or attached to the work 39 | (an example is provided in the Appendix below). 40 | 41 | "Derivative Works" shall mean any work, whether in Source or Object 42 | form, that is based on (or derived from) the Work and for which the 43 | editorial revisions, annotations, elaborations, or other modifications 44 | represent, as a whole, an original work of authorship. For the purposes 45 | of this License, Derivative Works shall not include works that remain 46 | separable from, or merely link (or bind by name) to the interfaces of, 47 | the Work and Derivative Works thereof. 48 | 49 | "Contribution" shall mean any work of authorship, including 50 | the original version of the Work and any modifications or additions 51 | to that Work or Derivative Works thereof, that is intentionally 52 | submitted to Licensor for inclusion in the Work by the copyright owner 53 | or by an individual or Legal Entity authorized to submit on behalf of 54 | the copyright owner. For the purposes of this definition, "submitted" 55 | means any form of electronic, verbal, or written communication sent 56 | to the Licensor or its representatives, including but not limited to 57 | communication on electronic mailing lists, source code control systems, 58 | and issue tracking systems that are managed by, or on behalf of, the 59 | Licensor for the purpose of discussing and improving the Work, but 60 | excluding communication that is conspicuously marked or otherwise 61 | designated in writing by the copyright owner as "Not a Contribution." 62 | 63 | "Contributor" shall mean Licensor and any individual or Legal Entity 64 | on behalf of whom a Contribution has been received by Licensor and 65 | subsequently incorporated within the Work. 66 | 67 | 2. Grant of Copyright License. Subject to the terms and conditions of 68 | this License, each Contributor hereby grants to You a perpetual, 69 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 70 | copyright license to reproduce, prepare Derivative Works of, 71 | publicly display, publicly perform, sublicense, and distribute the 72 | Work and such Derivative Works in Source or Object form. 73 | 74 | 3. Grant of Patent License. Subject to the terms and conditions of 75 | this License, each Contributor hereby grants to You a perpetual, 76 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 77 | (except as stated in this section) patent license to make, have made, 78 | use, offer to sell, sell, import, and otherwise transfer the Work, 79 | where such license applies only to those patent claims licensable 80 | by such Contributor that are necessarily infringed by their 81 | Contribution(s) alone or by combination of their Contribution(s) 82 | with the Work to which such Contribution(s) was submitted. If You 83 | institute patent litigation against any entity (including a 84 | cross-claim or counterclaim in a lawsuit) alleging that the Work 85 | or a Contribution incorporated within the Work constitutes direct 86 | or contributory patent infringement, then any patent licenses 87 | granted to You under this License for that Work shall terminate 88 | as of the date such litigation is filed. 89 | 90 | 4. Redistribution. You may reproduce and distribute copies of the 91 | Work or Derivative Works thereof in any medium, with or without 92 | modifications, and in Source or Object form, provided that You 93 | meet the following conditions: 94 | 95 | (a) You must give any other recipients of the Work or 96 | Derivative Works a copy of this License; and 97 | 98 | (b) You must cause any modified files to carry prominent notices 99 | stating that You changed the files; and 100 | 101 | (c) You must retain, in the Source form of any Derivative Works 102 | that You distribute, all copyright, patent, trademark, and 103 | attribution notices from the Source form of the Work, 104 | excluding those notices that do not pertain to any part of 105 | the Derivative Works; and 106 | 107 | (d) If the Work includes a "NOTICE" text file as part of its 108 | distribution, then any Derivative Works that You distribute must 109 | include a readable copy of the attribution notices contained 110 | within such NOTICE file, excluding those notices that do not 111 | pertain to any part of the Derivative Works, in at least one 112 | of the following places: within a NOTICE text file distributed 113 | as part of the Derivative Works; within the Source form or 114 | documentation, if provided along with the Derivative Works; or, 115 | within a display generated by the Derivative Works, if and 116 | wherever such third-party notices normally appear. The contents 117 | of the NOTICE file are for informational purposes only and 118 | do not modify the License. You may add Your own attribution 119 | notices within Derivative Works that You distribute, alongside 120 | or as an addendum to the NOTICE text from the Work, provided 121 | that such additional attribution notices cannot be construed 122 | as modifying the License. 123 | 124 | You may add Your own copyright statement to Your modifications and 125 | may provide additional or different license terms and conditions 126 | for use, reproduction, or distribution of Your modifications, or 127 | for any such Derivative Works as a whole, provided Your use, 128 | reproduction, and distribution of the Work otherwise complies with 129 | the conditions stated in this License. 130 | 131 | 5. Submission of Contributions. Unless You explicitly state otherwise, 132 | any Contribution intentionally submitted for inclusion in the Work 133 | by You to the Licensor shall be under the terms and conditions of 134 | this License, without any additional terms or conditions. 135 | Notwithstanding the above, nothing herein shall supersede or modify 136 | the terms of any separate license agreement you may have executed 137 | with Licensor regarding such Contributions. 138 | 139 | 6. Trademarks. This License does not grant permission to use the trade 140 | names, trademarks, service marks, or product names of the Licensor, 141 | except as required for reasonable and customary use in describing the 142 | origin of the Work and reproducing the content of the NOTICE file. 143 | 144 | 7. Disclaimer of Warranty. Unless required by applicable law or 145 | agreed to in writing, Licensor provides the Work (and each 146 | Contributor provides its Contributions) on an "AS IS" BASIS, 147 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 148 | implied, including, without limitation, any warranties or conditions 149 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 150 | PARTICULAR PURPOSE. You are solely responsible for determining the 151 | appropriateness of using or redistributing the Work and assume any 152 | risks associated with Your exercise of permissions under this License. 153 | 154 | 8. Limitation of Liability. In no event and under no legal theory, 155 | whether in tort (including negligence), contract, or otherwise, 156 | unless required by applicable law (such as deliberate and grossly 157 | negligent acts) or agreed to in writing, shall any Contributor be 158 | liable to You for damages, including any direct, indirect, special, 159 | incidental, or consequential damages of any character arising as a 160 | result of this License or out of the use or inability to use the 161 | Work (including but not limited to damages for loss of goodwill, 162 | work stoppage, computer failure or malfunction, or any and all 163 | other commercial damages or losses), even if such Contributor 164 | has been advised of the possibility of such damages. 165 | 166 | 9. Accepting Warranty or Additional Liability. While redistributing 167 | the Work or Derivative Works thereof, You may choose to offer, 168 | and charge a fee for, acceptance of support, warranty, indemnity, 169 | or other liability obligations and/or rights consistent with this 170 | License. However, in accepting such obligations, You may act only 171 | on Your own behalf and on Your sole responsibility, not on behalf 172 | of any other Contributor, and only if You agree to indemnify, 173 | defend, and hold each Contributor harmless for any liability 174 | incurred by, or claims asserted against, such Contributor by reason 175 | of your accepting any such warranty or additional liability. 176 | -------------------------------------------------------------------------------- /NOTICE: -------------------------------------------------------------------------------- 1 | Copyright 2020 Amazon.com, Inc. or its affiliates. All Rights Reserved. 2 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ## Detecting and analyzing incorrect model predictions with Amazon SageMaker Model Monitor and Debugger 2 | 3 | This repository contains the notebook and scripts for the blogpost "Detecting and analyzing incorrect model predictions with Amazon SageMaker Model Monitor and Debugger" 4 | 5 | [Create a SageMaker notebook instance](https://docs.aws.amazon.com/sagemaker/latest/dg/howitworks-create-ws.html) and clone the repository: 6 | ``` 7 | git clone git@github.com:aws-samples/amazon-sagemaker-analyze-model-predictions.git 8 | ``` 9 | 10 | In the notebook [analyze_model_predictions.ipynb](analyze_model_predictions.ipynb) we first deploy a [ResNet18](https://arxiv.org/abs/1512.03385) model that has been trained to distinguish between 43 categories of traffic signs using the [German Traffic Sign dataset](https://ieeexplore.ieee.org/document/6033395). 11 | 12 | We will setup [SageMaker Model Monitor](https://aws.amazon.com/blogs/aws/amazon-sagemaker-model-monitor-fully-managed-automatic-monitoring-for-your-machine-learning-models/) to automatically capture inference requests and predictions. 13 | Afterwards we launch a [monitoring schedule](https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor-scheduling.html) that periodically kicks off a custom processing job to inspect collected data and to detect unexpected model behavior. 14 | 15 | We will then create adversarial images which lead to the model making incorrect predictions. Once Model Monitor detects this issue, we will use [SageMaker Debugger](https://aws.amazon.com/blogs/aws/amazon-sagemaker-debugger-debug-your-machine-learning-models/) to obtain visual explanations of the deployed model. This is done by updating the endpoint to emit tensors during inference and we will then use those tensors to compute saliency maps. 16 | 17 | The saliency map can be rendered as heat-map and reveals the parts of an image that were critical in the prediction. Below is an example (taken from the [German Traffic Sign dataset](http://benchmark.ini.rub.de/?section=gtsrb&subsection=dataset)): the image on the left is the input into the fine-tuned ResNet model, which predicted the image class 25 (‘Road work’). The right image shows the input image overlaid with a heat-map where red indicates the most relevant and blue the least relevant pixels for predicting the class 25. 18 | 19 |

20 | 21 |

22 | 23 | ## License 24 | 25 | This project is licensed under the Apache-2.0 License. 26 | 27 | -------------------------------------------------------------------------------- /code/custom_hook.py: -------------------------------------------------------------------------------- 1 | import smdebug.pytorch as smd 2 | import torch 3 | 4 | class CustomHook(smd.Hook): 5 | 6 | def image_gradients(self, image): 7 | """Register input image for backward pass, to get image gradients""" 8 | image.register_hook(self.backward_hook("image")) 9 | 10 | def forward_hook(self, module, inputs, outputs): 11 | module_name = self.module_maps[module] 12 | self._write_inputs(module_name, inputs) 13 | 14 | outputs.register_hook(self.backward_hook(module_name + "_output")) 15 | 16 | #record running mean and var of BatchNorm layers 17 | if isinstance(module, torch.nn.BatchNorm2d): 18 | self._write_outputs(module_name + ".running_mean", module.running_mean) 19 | self._write_outputs(module_name + ".running_var", module.running_var) 20 | 21 | self._write_outputs(module_name, outputs) 22 | self.last_saved_step = self.step 23 | -------------------------------------------------------------------------------- /code/pretrained_model.py: -------------------------------------------------------------------------------- 1 | '''SageMaker PyTorch inference container overrides to serve plain ResNet18 model''' 2 | 3 | # Python Built-Ins: 4 | import argparse 5 | from io import BytesIO 6 | import os 7 | 8 | # External Dependencies: 9 | import numpy as np 10 | from PIL import Image 11 | import torch 12 | import torch.nn as nn 13 | import torch.optim as optim 14 | from torchvision import models, transforms 15 | 16 | 17 | def model_fn(model_dir): 18 | #create model 19 | model = models.resnet18() 20 | 21 | #traffic sign dataset has 43 classes 22 | nfeatures = model.fc.in_features 23 | model.fc = nn.Linear(nfeatures, 43) 24 | 25 | #load model 26 | weights = torch.load(f'{model_dir}/model/model.pt', map_location=lambda storage, loc: storage) 27 | model.load_state_dict(weights) 28 | 29 | model.eval() 30 | model.cpu() 31 | 32 | return model 33 | 34 | 35 | def transform_fn(model, data, content_type, output_content_type): 36 | transform = transforms.Compose([ 37 | transforms.Resize((128,128)), 38 | transforms.ToTensor(), 39 | transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), 40 | ]) 41 | 42 | image = np.load(BytesIO(data)) 43 | image = Image.fromarray(image) 44 | image = transform(image) 45 | 46 | image = image.unsqueeze(0) 47 | 48 | #forward pass 49 | prediction = model(image) 50 | 51 | #get prediction 52 | predicted_class = prediction.data.max(1, keepdim=True)[1] 53 | 54 | response_body = np.array(predicted_class.cpu()).tolist() 55 | return response_body, output_content_type 56 | -------------------------------------------------------------------------------- /code/pretrained_model_with_debugger_hook.py: -------------------------------------------------------------------------------- 1 | '''SageMaker PyTorch inference container overrides to serve ResNet18 model with debug hook''' 2 | 3 | # Python Built-Ins: 4 | import argparse 5 | from io import BytesIO 6 | import logging 7 | import os 8 | from typing import Any, Union 9 | 10 | # External Dependencies: 11 | import numpy as np 12 | from PIL import Image 13 | import smdebug.pytorch as smd 14 | from smdebug import modes 15 | from smdebug.core.modes import ModeKeys 16 | import torch 17 | import torch.nn as nn 18 | import torch.optim as optim 19 | from torchvision import models, transforms 20 | 21 | # Local Dependencies: 22 | from custom_hook import CustomHook 23 | 24 | 25 | logger = logging.getLogger() 26 | 27 | 28 | class ModelWithDebugHook: 29 | def __init__(self, model: Any, hook: Union[smd.Hook, None]): 30 | '''Simple container to associate a 'model' with a SageMaker debug hook''' 31 | self.model = model 32 | self.hook = hook 33 | 34 | 35 | def model_fn(model_dir: str) -> ModelWithDebugHook: 36 | #create model 37 | model = models.resnet18() 38 | 39 | #traffic sign dataset has 43 classes 40 | nfeatures = model.fc.in_features 41 | model.fc = nn.Linear(nfeatures, 43) 42 | 43 | #load model 44 | weights = torch.load(f'{model_dir}/model/model.pt', map_location=lambda storage, loc: storage) 45 | model.load_state_dict(weights) 46 | 47 | model.eval() 48 | model.cpu() 49 | 50 | #hook configuration 51 | tensors_output_s3uri = os.environ.get('tensors_output') 52 | if tensors_output_s3uri is None: 53 | logger.warning( 54 | 'WARN: Skipping hook configuration as no tensors_output env var provided. ' 55 | 'Tensors will not be exported' 56 | ) 57 | hook = None 58 | else: 59 | save_config = smd.SaveConfig(mode_save_configs={ 60 | smd.modes.PREDICT: smd.SaveConfigMode(save_interval=1), 61 | }) 62 | 63 | hook = CustomHook( 64 | tensors_output_s3uri, 65 | save_config=save_config, 66 | include_regex='.*bn|.*bias|.*downsample|.*ResNet_input|.*image|.*fc_output', 67 | ) 68 | 69 | #register hook 70 | hook.register_module(model) 71 | 72 | #set mode 73 | hook.set_mode(modes.PREDICT) 74 | 75 | return ModelWithDebugHook(model, hook) 76 | 77 | 78 | def transform_fn(model_with_hook, data, content_type, output_content_type): 79 | model = model_with_hook.model 80 | hook = model_with_hook.hook 81 | 82 | val_transform = transforms.Compose([ 83 | transforms.Resize((128,128)), 84 | transforms.ToTensor(), 85 | transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), 86 | ]) 87 | 88 | image = np.load(BytesIO(data)) 89 | image = Image.fromarray(image) 90 | image = val_transform(image) 91 | 92 | image = image.unsqueeze(0) 93 | image = image.to('cpu').requires_grad_() 94 | if hook is not None: 95 | hook.image_gradients(image) 96 | 97 | #forward pass 98 | prediction = model(image) 99 | 100 | #get prediction 101 | predicted_class = prediction.data.max(1, keepdim=True)[1] 102 | output = prediction[0, predicted_class[0]] 103 | model.zero_grad() 104 | 105 | #compute gradients with respect to outputs 106 | output.backward() 107 | 108 | response_body = np.array(predicted_class.cpu()).tolist() 109 | return response_body, output_content_type 110 | -------------------------------------------------------------------------------- /code/requirements.txt: -------------------------------------------------------------------------------- 1 | Pillow 2 | smdebug==0.5.0 3 | -------------------------------------------------------------------------------- /docker/Dockerfile: -------------------------------------------------------------------------------- 1 | FROM python:3.7-slim-buster 2 | 3 | RUN pip3 install sagemaker 4 | ENV PYTHONUNBUFFERED=TRUE 5 | 6 | ADD evaluation.py / 7 | 8 | ENTRYPOINT ["python3", "/evaluation.py"] -------------------------------------------------------------------------------- /docker/evaluation.py: -------------------------------------------------------------------------------- 1 | """Custom Model Monitoring script for classification 2 | """ 3 | 4 | # Python Built-Ins: 5 | from collections import defaultdict 6 | import datetime 7 | import json 8 | import os 9 | import traceback 10 | from types import SimpleNamespace 11 | 12 | # External Dependencies: 13 | import numpy as np 14 | 15 | 16 | def get_environment(): 17 | """Load configuration variables for SM Model Monitoring job 18 | 19 | See https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor-byoc-contract-inputs.html 20 | """ 21 | try: 22 | with open("/opt/ml/config/processingjobconfig.json", "r") as conffile: 23 | defaults = json.loads(conffile.read())["Environment"] 24 | except Exception as e: 25 | traceback.print_exc() 26 | print("Unable to read environment vars from SM processing config file") 27 | defaults = {} 28 | 29 | return SimpleNamespace( 30 | dataset_format=os.environ.get("dataset_format", defaults.get("dataset_format")), 31 | dataset_source=os.environ.get( 32 | "dataset_source", 33 | defaults.get("dataset_source", "/opt/ml/processing/input/endpoint"), 34 | ), 35 | end_time=os.environ.get("end_time", defaults.get("end_time")), 36 | output_path=os.environ.get( 37 | "output_path", 38 | defaults.get("output_path", "/opt/ml/processing/resultdata"), 39 | ), 40 | publish_cloudwatch_metrics=os.environ.get( 41 | "publish_cloudwatch_metrics", 42 | defaults.get("publish_cloudwatch_metrics", "Enabled"), 43 | ), 44 | sagemaker_endpoint_name=os.environ.get( 45 | "sagemaker_endpoint_name", 46 | defaults.get("sagemaker_endpoint_name"), 47 | ), 48 | sagemaker_monitoring_schedule_name=os.environ.get( 49 | "sagemaker_monitoring_schedule_name", 50 | defaults.get("sagemaker_monitoring_schedule_name"), 51 | ), 52 | start_time=os.environ.get("start_time", defaults.get("start_time")), 53 | max_ratio_threshold=float(os.environ.get("THRESHOLD", defaults.get("THRESHOLD", "nan"))), 54 | ) 55 | 56 | 57 | if __name__=="__main__": 58 | env = get_environment() 59 | print(f"Starting evaluation with config:\n{env}") 60 | 61 | print("Analyzing collected data...") 62 | total_record_count = 0 # Including error predictions that we can't read the response for 63 | error_record_count = 0 64 | counts = defaultdict(int) # dict defaulting to 0 when unseen keys are requested 65 | for path, directories, filenames in os.walk(env.dataset_source): 66 | for filename in filter(lambda f: f.lower().endswith(".jsonl"), filenames): 67 | with open(os.path.join(path, filename), "r") as file: 68 | for entry in file: 69 | total_record_count += 1 70 | try: 71 | response = json.loads(json.loads(entry)["captureData"]["endpointOutput"]["data"]) 72 | except: 73 | error_record_count += 1 74 | continue 75 | 76 | # response will typically be a 1x1 array (single-request, single output field), but we 77 | # can handle batch inference too by looping over array: 78 | for record in response: 79 | prediction = record[0] 80 | counts[prediction] += 1 81 | print(f"Class prediction counts: {counts}") 82 | 83 | print("Calculating secondary summaries...") 84 | total_prediction_count = np.sum(list(counts.values())) 85 | max_count = float("-inf") 86 | max_class = None 87 | numeric_class_names = [] 88 | for class_name, count in counts.items(): 89 | try: 90 | numeric_class_names.append(class_name - 0) 91 | except: 92 | pass 93 | if count > max_count: 94 | max_count = count 95 | max_class = class_name 96 | max_class_ratio = max_count / total_prediction_count 97 | mean_numeric_label = np.average(numeric_class_names, weights=[counts[c] for c in numeric_class_names]) 98 | 99 | print("Checking for constraint violations...") 100 | violations = [] 101 | if max_class_ratio > env.max_ratio_threshold: 102 | violations.append({ 103 | "feature_name": "PredictedClass", 104 | "constraint_check_type": "baseline_drift_check", 105 | "description": "Class {} predicted {:.2f}% of the time: Exceeded {:.2f}% threshold".format( 106 | max_class, 107 | max_class_ratio * 100, 108 | env.max_ratio_threshold * 100, 109 | ), 110 | }) 111 | if error_record_count > 0: 112 | violations.append({ 113 | "feature_name": "PredictedClass", 114 | # TODO: Maybe this should be missing_column_check when error_record_count == total_record_count? 115 | "constraint_check_type": "completeness_check", 116 | "description": "Could not read predicted class for {} req/res pairs ({:.2f}% of total)".format( 117 | error_record_count, 118 | error_record_count * 100 / total_record_count, 119 | ), 120 | }) 121 | print(f"Violations: {violations if len(violations) else 'None'}") 122 | 123 | print("Writing violations file...") 124 | with open(os.path.join(env.output_path, "constraints_violations.json"), "w") as outfile: 125 | outfile.write(json.dumps( 126 | { "violations": violations }, 127 | indent=4, 128 | )) 129 | 130 | # You could also consider writing a statistics.json and constraints.json here, per the standard results: 131 | # https://docs.aws.amazon.com/sagemaker/latest/dg/model-monitor-interpreting-results.html 132 | 133 | print("Writing overall status output...") 134 | with open("/opt/ml/output/message", "w") as outfile: 135 | if len(violations): 136 | msg = f"CompletedWithViolations: {violations[0]['description']}" 137 | else: 138 | msg = "Completed: Job completed successfully with no violations." 139 | outfile.write(msg) 140 | print(msg) 141 | 142 | if env.publish_cloudwatch_metrics: 143 | print("Writing CloudWatch metrics...") 144 | with open("/opt/ml/output/metrics/cloudwatch/cloudwatch_metrics.jsonl", "a+") as outfile: 145 | # One metric per line (JSONLines list of dictionaries) 146 | # Remember these metrics are aggregated in graphs, so we report them as statistics on our dataset 147 | json.dump( 148 | { 149 | "MetricName": f"feature_data_PredictedClass", 150 | "Timestamp": datetime.datetime.now().strftime("%Y-%m-%dT%H:%M:%SZ"), 151 | "Dimensions": [ 152 | { "Name": "Endpoint", "Value": env.sagemaker_endpoint_name or "unknown" }, 153 | { 154 | "Name": "MonitoringSchedule", 155 | "Value": env.sagemaker_monitoring_schedule_name or "unknown", 156 | }, 157 | ], 158 | "StatisticValues": { 159 | "Maximum": np.max(numeric_class_names).item(), 160 | "Minimum": np.min(numeric_class_names).item(), 161 | "SampleCount": int(total_prediction_count), 162 | "Sum": np.sum( 163 | np.array(numeric_class_names) 164 | * np.array([counts[c] for c in numeric_class_names]) 165 | ).item(), 166 | }, 167 | }, 168 | outfile 169 | ) 170 | outfile.write("\n") 171 | pct_successful = (total_record_count - error_record_count) / total_record_count 172 | json.dump( 173 | { 174 | "MetricName": f"feature_non_null_PredictedClass", 175 | "Timestamp": datetime.datetime.now().strftime("%Y-%m-%dT%H:%M:%SZ"), 176 | "Dimensions": [ 177 | { "Name": "Endpoint", "Value": env.sagemaker_endpoint_name or "unknown" }, 178 | { 179 | "Name": "MonitoringSchedule", 180 | "Value": env.sagemaker_monitoring_schedule_name or "unknown", 181 | }, 182 | ], 183 | "StatisticValues": { 184 | "Maximum": pct_successful, 185 | "Minimum": pct_successful, 186 | "SampleCount": total_record_count, 187 | "Sum": pct_successful * total_record_count, 188 | }, 189 | }, 190 | outfile 191 | ) 192 | outfile.write("\n") 193 | # numpy types may not be JSON serializable: 194 | max_class_ratio = float(max_class_ratio) 195 | total_prediction_count = int(total_prediction_count) 196 | json.dump( 197 | { 198 | "MetricName": f"feature_baseline_drift_PredictedClass", 199 | "Timestamp": datetime.datetime.now().strftime("%Y-%m-%dT%H:%M:%SZ"), 200 | "Dimensions": [ 201 | { "Name": "Endpoint", "Value": env.sagemaker_endpoint_name or "unknown" }, 202 | { 203 | "Name": "MonitoringSchedule", 204 | "Value": env.sagemaker_monitoring_schedule_name or "unknown", 205 | }, 206 | ], 207 | "StatisticValues": { 208 | "Maximum": max_class_ratio, 209 | "Minimum": max_class_ratio, 210 | "SampleCount": total_prediction_count, 211 | "Sum": max_class_ratio * total_prediction_count, 212 | }, 213 | }, 214 | outfile 215 | ) 216 | outfile.write("\n") 217 | print("Done") 218 | -------------------------------------------------------------------------------- /images/image.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/aws-samples/amazon-sagemaker-analyze-model-predictions/c3db4a3e2bf9afff57c7b17b10b17066aba9316e/images/image.png -------------------------------------------------------------------------------- /images/screenshot.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/aws-samples/amazon-sagemaker-analyze-model-predictions/c3db4a3e2bf9afff57c7b17b10b17066aba9316e/images/screenshot.png -------------------------------------------------------------------------------- /model/model.pt: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/aws-samples/amazon-sagemaker-analyze-model-predictions/c3db4a3e2bf9afff57c7b17b10b17066aba9316e/model/model.pt -------------------------------------------------------------------------------- /utils.py: -------------------------------------------------------------------------------- 1 | # Python Built-Ins: 2 | import os 3 | 4 | # External Dependencies: 5 | from IPython.display import display, HTML 6 | import matplotlib.pyplot as plt 7 | import numpy as np 8 | import torch 9 | import torch.nn as nn 10 | from torchvision import datasets, models, transforms 11 | 12 | # Configuration: 13 | image_norm_mean = [0.485, 0.456, 0.406] 14 | image_norm_stddev = [0.229, 0.224, 0.225] 15 | 16 | 17 | def create_circular_mask(h, w, center=None, radius=None): 18 | """Create a boolean mask selecting a circular region (e.g. in an image) 19 | 20 | Sourced from the following StackOverflow post, with tweaks: 21 | https://stackoverflow.com/a/44874588/13352657 22 | """ 23 | if center is None: # use the middle of the image 24 | center = (int(w/2), int(h/2)) 25 | elif np.all(np.array(center) <= 1.): # Convert fractional to absolute 26 | center = (np.array((w, h)) * center).astype(int) 27 | if radius is None: # use the smallest distance between the center and image walls 28 | radius = min(center[0], center[1], w-center[0], h-center[1]) 29 | elif radius < 1.: # Convert fractional to absolute 30 | radius = min(w, h) * radius 31 | 32 | Y, X = np.ogrid[:h, :w] 33 | dist_from_center = np.sqrt((X - center[0])**2 + (Y-center[1])**2) 34 | 35 | mask = dist_from_center <= radius 36 | return mask 37 | 38 | 39 | def get_dataloader(): 40 | val_transform = transforms.Compose([ 41 | transforms.Resize((128, 128)), 42 | transforms.ToTensor(), 43 | transforms.Normalize(image_norm_mean, image_norm_stddev), 44 | ]) 45 | dataset = datasets.ImageFolder('GTSRB/Final_Test/', val_transform) 46 | val_dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True) 47 | 48 | return val_dataloader 49 | 50 | 51 | def tensor_to_imgarray(image, floating_point=False): 52 | """Convert a normalized tensor or matrix as used by the model into a standard image array 53 | 54 | Parameters 55 | ---------- 56 | image : Union[numpy.ndarray, torch.Tensor] 57 | A mean/std-normalized image tensor or matrix in inference format for the model 58 | floating_point : bool (Optional) 59 | Set True to skip conversion to 0-255 uint8 and return a 0-1.0 float ndarray instead 60 | """ 61 | if torch.is_tensor(image): 62 | image = image.cpu().numpy() 63 | if len(image.shape) > 3: 64 | # Leading batch dimension - take first el only 65 | image = image[tuple(0 if dim == 0 else slice(None) for dim in range(len(image.shape)))] 66 | 67 | image_shape = image.shape 68 | channeldim = image_shape.index(3) 69 | result = image 70 | 71 | # Move channel to correct (trailing) dim if not already: 72 | if channeldim < (len(image_shape) - 1): 73 | result = np.moveaxis(result, channeldim, -1) 74 | image_shape = result.shape 75 | channeldim = len(image_shape) - 1 76 | 77 | # Pad mean and stddev constants to image dimensions 78 | # TODO: Simplify this when we're consistent in what the image dimensions are! 79 | # We use a loop here in case some environments use numpy<1.18 when the functionality to accept a tuple of 80 | # axes was introduced: 81 | stddev = image_norm_stddev 82 | mean = image_norm_mean 83 | for _ in range(channeldim): 84 | stddev = np.expand_dims(stddev, 0) 85 | mean = np.expand_dims(mean, 0) 86 | for _ in range(channeldim + 1, len(image_shape)): 87 | stddev = np.expand_dims(stddev, -1) 88 | 89 | result = (result * stddev) + mean 90 | if floating_point: 91 | return np.clip(result, 0., 1.) 92 | else: 93 | return np.clip(result * 255.0, 0, 255).astype(np.uint8) 94 | 95 | 96 | def load_model(): 97 | #check if GPU is available and set context 98 | device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 99 | 100 | #load model 101 | model = models.resnet18() 102 | 103 | #traffic sign dataset has 43 classes 104 | nfeatures = model.fc.in_features 105 | model.fc = nn.Linear(nfeatures, 43) 106 | 107 | weights = torch.load('model/model.pt', map_location=lambda storage, loc: storage) 108 | model.load_state_dict(weights) 109 | 110 | for param in model.parameters(): 111 | param.requires_grad = False 112 | 113 | model.to(device).eval() 114 | return model 115 | 116 | 117 | def show_images_diff(image, adv_image, adv_label=None, class_names=None, cmap=None): 118 | adv_image = tensor_to_imgarray(adv_image, floating_point=True) 119 | image = tensor_to_imgarray(image, floating_point=True) 120 | 121 | fig, (ax0, ax1, ax2) = plt.subplots(ncols=3, figsize=(12, 4)) 122 | 123 | ax0.imshow(image) 124 | ax0.set_title('Original') 125 | ax0.set_axis_off() 126 | 127 | ax1.imshow(adv_image) 128 | if adv_label is None: 129 | ax1.set_title('Adversarial image') 130 | else: 131 | ax1.set_title(f'Model prediction: {class_names[adv_label] if class_names else adv_label}') 132 | ax1.set_axis_off() 133 | 134 | difference = adv_image - image 135 | 136 | # If colormapping, convert RGB to single lightness channel: 137 | if cmap is not None and 3 in difference.shape: 138 | channeldim = difference.shape.index(3) 139 | rgbindices = [ 140 | tuple(rgb if dim == channeldim else slice(None) for dim in range(len(difference.shape))) 141 | for rgb in range(3) 142 | ] 143 | # RGB->lightness function per PIL docs, but no need to import the lib just for this: 144 | # https://pillow.readthedocs.io/en/3.2.x/reference/Image.html#PIL.Image.Image.convert 145 | # L = R * 299/1000 + G * 587/1000 + B * 114/1000 146 | difference = ( 147 | difference[rgbindices[0]] * 0.299 148 | + difference[rgbindices[1]] * 0.587 149 | + difference[rgbindices[2]] * 0.114 150 | ) 151 | 152 | # Scale to a symmetric range around max absolute difference (which we print out), and map that to 0-1 153 | # for imshow. (When colormapping we could just use vmin/vmax, but this way we keep same path for both). 154 | maxdiff = abs(difference).max() 155 | difference = difference / (maxdiff * 2.0) + 0.5 156 | ax2.imshow(difference, cmap=cmap, vmin=0., vmax=1.) 157 | ax2.set_title(f'Diff ({-maxdiff:.4f} to {maxdiff:.4f})') 158 | ax2.set_axis_off() 159 | plt.tight_layout() 160 | plt.show() 161 | 162 | 163 | def plot_saliency_map( 164 | saliency_map, 165 | image, 166 | predicted_class=None, 167 | class_names=None, 168 | confidence=None, 169 | cmap=plt.cm.plasma, 170 | alpha=0.5, 171 | interest_center=(0.5, 0.5), 172 | interest_radius=0.4, 173 | max_bg_saliency_thresh=0.85, 174 | ): 175 | """Plot an image classification result with saliency map 176 | 177 | Parameters 178 | ---------- 179 | saliency_map : 180 | A *normalized* (range 0-1.0) importance/saliency map matching image height and width, but with no 181 | channel dimension. 182 | image : 183 | An image with leading channel dimension, normalized values (mean + std). 184 | TODO: Parameterize the normalization rather than hard-coding? 185 | predicted_class : Any (Optional) 186 | If supplied, the saliency overlay plot will be titled to indicate which class was detected. 187 | class_names : Mapping[Any, Any] (Optional) 188 | If supplied as well as predicted_class, the saliency overlay plot title will *also* be annotated with 189 | the "name" looked up from the raw predicted_class label. 190 | confidence : float (Optional) 191 | If supplied as well as predicted_class, the saliency overlay plot title will also be annotated with 192 | the confidence score. Should be in 0-1.0 range, will be displayed as percentage. 193 | cmap : matplotlib.pyplot.colors.ColorMap (Optional) 194 | A PyPlot colormap to apply for the saliency map. Defaults to plt.cm.plasma 195 | alpha : float (Optional) 196 | Opacity of the saliency heatmap to show in the overlay image. Defaults to 0.5 197 | interest_center : Tuple[float] (Optional) 198 | Relative (w, h) center of expected interest area in image (0.5,0.5 = middle by default). Used only 199 | when interest_radius is not None 200 | interest_radius : float (Optional) 201 | Relative radius of interest circle in image (0.4 for 80% diameter by default). When this parameter is 202 | not explicitly set to None, draw a 'circle of interest' on the plots and calculate the maximum and 203 | average saliency of points *outside* this region - to check for unexpected attention focus away from 204 | the subject of the image. 205 | max_bg_saliency_thresh : float (Optional) 206 | Display a warning box in the notebook when the maximum saliency *outside* the circle of interest is 207 | >= this value. 208 | """ 209 | # Revert image normalization 210 | image = tensor_to_imgarray(image, floating_point=True) 211 | 212 | # Given the saliency map has already been normalized to 0-1, we can apply pyplot colormap as below: 213 | # (Otherwise see mpl.colors.Normalize and plt.cm.ScalarMappable(norm=norm, cmap=cmap)) 214 | heatmap = cmap(saliency_map) 215 | heatmap = heatmap[:, :, :-1] # Trim off the alpha channel (always 1.0 anyway for typical cmaps) 216 | 217 | # Blend image with heatmap: 218 | combined_image = alpha * heatmap + (1-alpha) * image 219 | 220 | # Plot 221 | fig, (ax0, ax1, ax2) = plt.subplots(ncols=3, figsize=(12, 4)) 222 | ax0.imshow(image) 223 | ax0.set_axis_off() 224 | ax0.set_title("Input image") 225 | ax1.imshow(combined_image) 226 | ax1.set_axis_off() 227 | if predicted_class is None: 228 | ax1.set_title("Saliency overlay") 229 | else: 230 | ax1.set_title("Predicted '{}'{}{}".format( 231 | str(predicted_class), 232 | f" ({class_names[predicted_class]})" if class_names is not None else "", 233 | f", {confidence * 100:.1f}%" if confidence is not None else "", 234 | )) 235 | ax2.imshow(heatmap) 236 | ax2.set_axis_off() 237 | ax2.set_title("Saliency heatmap") 238 | plt.tight_layout() 239 | 240 | # If required, plot interest circles and calculate background saliency metrics: 241 | if interest_radius is not None: 242 | h = heatmap.shape[0] 243 | w = heatmap.shape[1] 244 | bg_mask = ~create_circular_mask(h, w, center=interest_center, radius=interest_radius) 245 | bg_saliency = bg_mask * saliency_map 246 | max_bg_saliency = np.max(bg_saliency) 247 | print(f"Max bg_saliency {max_bg_saliency}, average bg_saliency {np.mean(bg_saliency)}") 248 | 249 | wh = np.array((w, h)) 250 | # Unfortunately you can't re-use a mpl 'artist' between plots, and there's no copy method! Ugh 251 | plt_circle0 = plt.Circle( 252 | wh * interest_center, np.min(wh) * interest_radius, color='white', fill=False, 253 | ) 254 | ax0.add_artist(plt_circle0) 255 | plt_circle1 = plt.Circle( 256 | wh * interest_center, np.min(wh) * interest_radius, color='white', fill=False, 257 | ) 258 | ax1.add_artist(plt_circle1) 259 | plt_circle2 = plt.Circle( 260 | wh * interest_center, np.min(wh) * interest_radius, color='white', fill=False, 261 | ) 262 | ax2.add_artist(plt_circle2) 263 | 264 | plt.show() 265 | 266 | if interest_radius is not None and max_bg_saliency >= max_bg_saliency_thresh: 267 | display(HTML('\n'.join(( 268 | f'
', 269 | 'High saliency outside region of interest: prediction may be attending to unreliable background', 270 | 'context', 271 | '
', 272 | )))) 273 | --------------------------------------------------------------------------------