File size: 17,548 Bytes
d822266 |
1 |
{"head_branch": "master", "contributor": "keras-team", "sha_fail": "2989b6f8fc2aca8bf8d0ffa32e58ceeaa3ef35ef", "sha_success": "0484165fa7cba3c095d95da19b1c45945b47c60b", "language": "Python", "repo_owner": "keras-team", "repo_name": "keras", "workflow_name": "Tests", "workflow_filename": "actions.yml", "workflow_path": ".github/workflows/actions.yml", "workflow": "name: Tests\n\non:\n push:\n branches: [ master ]\n pull_request:\n release:\n types: [created]\n\npermissions:\n contents: read\n\njobs:\n build:\n strategy:\n fail-fast: false\n matrix:\n python-version: [3.9]\n backend: [tensorflow, jax, torch, numpy]\n name: Run tests\n runs-on: ubuntu-latest\n env:\n PYTHON: ${{ matrix.python-version }}\n KERAS_BACKEND: ${{ matrix.backend }}\n steps:\n - uses: actions/checkout@v3\n - name: Check for changes in keras/applications\n uses: dorny/paths-filter@v2\n id: filter\n with:\n filters: |\n applications:\n - 'keras/applications/**'\n - name: Set up Python\n uses: actions/setup-python@v4\n with:\n python-version: ${{ matrix.python-version }}\n - name: Get pip cache dir\n id: pip-cache\n run: |\n python -m pip install --upgrade pip setuptools\n echo \"dir=$(pip cache dir)\" >> $GITHUB_OUTPUT\n - name: pip cache\n uses: actions/cache@v3\n with:\n path: ${{ steps.pip-cache.outputs.dir }}\n key: ${{ runner.os }}-pip-${{ hashFiles('setup.py') }}-${{ hashFiles('requirements.txt') }}\n - name: Install dependencies\n run: |\n pip install -r requirements.txt --progress-bar off --upgrade\n pip install -e \".\" --progress-bar off --upgrade\n - name: Test applications with pytest\n if: ${{ steps.filter.outputs.applications == 'true' }}\n run: |\n pytest keras/applications --cov=keras.applications\n coverage xml --include='keras/applications/*' -o apps-coverage.xml\n - name: Codecov keras.applications\n if: ${{ steps.filter.outputs.applications == 'true' }}\n uses: codecov/codecov-action@v3\n with:\n env_vars: PYTHON,KERAS_BACKEND\n flags: keras.applications,keras.applications-${{ matrix.backend }}\n files: apps-coverage.xml\n fail_ci_if_error: true\n - name: Test with pytest\n run: |\n pytest keras --ignore keras/applications --cov=keras\n coverage xml --omit='keras/applications/*' -o core-coverage.xml\n - name: Codecov keras\n uses: codecov/codecov-action@v3\n with:\n env_vars: PYTHON,KERAS_BACKEND\n flags: keras,keras-${{ matrix.backend }}\n files: core-coverage.xml\n fail_ci_if_error: true\n\n format:\n name: Check the code format\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v3\n - name: Set up Python 3.9\n uses: actions/setup-python@v4\n with:\n python-version: '3.9'\n - name: Get pip cache dir\n id: pip-cache\n run: |\n python -m pip install --upgrade pip setuptools\n echo \"dir=$(pip cache dir)\" >> $GITHUB_OUTPUT\n - name: pip cache\n uses: actions/cache@v3\n with:\n path: ${{ steps.pip-cache.outputs.dir }}\n key: ${{ runner.os }}-pip-${{ hashFiles('setup.py') }}-${{ hashFiles('requirements.txt') }}\n - name: Install dependencies\n run: |\n pip install -r requirements.txt --progress-bar off --upgrade\n pip install -e \".\" --progress-bar off --upgrade\n - name: Lint\n run: bash shell/lint.sh\n", "logs": "keras/testing/test_utils.py 11 1 4 1 87%\nkeras/trainers/__init__.py 0 0 0 0 100%\nkeras/trainers/compile_utils.py 356 54 281 39 81%\nkeras/trainers/data_adapters/__init__.py 0 0 0 0 100%\nkeras/trainers/data_adapters/array_data_adapter.py 141 19 58 11 84%\nkeras/trainers/data_adapters/data_adapter.py 13 0 0 0 100%\nkeras/trainers/data_adapters/data_adapter_utils.py 92 21 61 13 73%\nkeras/trainers/data_adapters/generator_data_adapter.py 41 3 8 3 88%\nkeras/trainers/data_adapters/py_dataset_adapter.py 252 49 80 17 76%\nkeras/trainers/data_adapters/tf_dataset_adapter.py 47 0 20 0 100%\nkeras/trainers/data_adapters/torch_data_adapter.py 46 3 10 3 89%\nkeras/trainers/epoch_iterator.py 105 13 72 13 83%\nkeras/trainers/trainer.py 193 23 89 14 84%\nkeras/utils/__init__.py 24 0 0 0 100%\nkeras/utils/argument_validation.py 43 10 26 5 78%\nkeras/utils/audio_dataset_utils.py 87 16 52 4 83%\nkeras/utils/backend_utils.py 37 6 12 3 78%\nkeras/utils/code_stats.py 40 3 34 2 91%\nkeras/utils/dataset_utils.py 286 61 202 45 74%\nkeras/utils/dtype_utils.py 25 0 16 0 100%\nkeras/utils/file_utils.py 221 52 129 23 71%\nkeras/utils/image_dataset_utils.py 91 6 48 6 90%\nkeras/utils/image_utils.py 149 73 80 13 45%\nkeras/utils/io_utils.py 34 0 10 0 100%\nkeras/utils/jax_utils.py 7 3 4 1 45%\nkeras/utils/model_visualization.py 193 169 90 0 8%\nkeras/utils/module_utils.py 32 1 6 1 95%\nkeras/utils/naming.py 34 1 8 1 95%\nkeras/utils/nest.py 42 9 18 3 77%\nkeras/utils/numerical_utils.py 58 4 26 5 89%\nkeras/utils/progbar.py 132 25 60 9 78%\nkeras/utils/python_utils.py 62 5 28 4 90%\nkeras/utils/rng_utils.py 16 1 6 3 82%\nkeras/utils/sequence_utils.py 41 8 24 6 78%\nkeras/utils/shape_utils.py 15 1 19 1 94%\nkeras/utils/summary_utils.py 212 39 116 18 77%\nkeras/utils/text_dataset_utils.py 68 4 40 5 90%\nkeras/utils/tf_utils.py 68 32 38 6 53%\nkeras/utils/timeseries_dataset_utils.py 62 4 48 5 92%\nkeras/utils/torch_utils.py 23 2 6 1 90%\nkeras/utils/traceback_utils.py 107 79 48 1 19%\nkeras/utils/tracking.py 97 10 54 5 82%\nkeras/version.py 1 0 0 0 100%\n----------------------------------------------------------------------------------------------\nTOTAL 34210 12322 13279 1442 62%\n\n=========================== short test summary info ============================\nFAILED keras/callbacks/model_checkpoint_test.py::ModelCheckpointTest::test_model_checkpoint_loading - ValueError: Arguments `target` and `output` must have the same shape. Received: target.shape=torch.Size([5, 1]), output.shape=torch.Size([5, 2])\n=========== 1 failed, 2902 passed, 306 skipped, 1 xpassed in 54.27s ============\n##[error]Process completed with exit code 1.\n", "diff": "diff --git a/examples/keras_io/vision/knowledge_distillation.py b/examples/keras_io/vision/knowledge_distillation.py\nnew file mode 100644\nindex 0000000000..8b0f2796a8\n--- /dev/null\n+++ b/examples/keras_io/vision/knowledge_distillation.py\n@@ -0,0 +1,246 @@\n+\"\"\"\n+Title: Knowledge Distillation\n+Author: [Kenneth Borup](https://twitter.com/Kennethborup)\n+Converted to Keras 3 by: [Md Awsafur Rahman](https://awsaf49.github.io)\n+Date created: 2020/09/01\n+Last modified: 2020/09/01\n+Description: Implementation of classical Knowledge Distillation.\n+Accelerator: GPU\n+\"\"\"\n+\"\"\"\n+## Introduction to Knowledge Distillation\n+\n+Knowledge Distillation is a procedure for model\n+compression, in which a small (student) model is trained to match a large pre-trained\n+(teacher) model. Knowledge is transferred from the teacher model to the student\n+by minimizing a loss function, aimed at matching softened teacher logits as well as\n+ground-truth labels.\n+\n+The logits are softened by applying a \"temperature\" scaling function in the softmax,\n+effectively smoothing out the probability distribution and revealing\n+inter-class relationships learned by the teacher.\n+\n+**Reference:**\n+\n+- [Hinton et al. (2015)](https://arxiv.org/abs/1503.02531)\n+\"\"\"\n+\n+\"\"\"\n+## Setup\n+\"\"\"\n+\n+import os\n+\n+os.environ[\"KERAS_BACKEND\"] = \"tensorflow\"\n+\n+import keras\n+from keras import layers\n+from keras import ops\n+import tensorflow as tf\n+import numpy as np\n+\n+\"\"\"\n+## Construct `Distiller()` class\n+\n+The custom `Distiller()` class, overrides the `Model` methods `compile`, `compute_loss`,\n+and `call`. In order to use the distiller, we need:\n+\n+- A trained teacher model\n+- A student model to train\n+- A student loss function on the difference between student predictions and ground-truth\n+- A distillation loss function, along with a `temperature`, on the difference between the\n+soft student predictions and the soft teacher labels\n+- An `alpha` factor to weight the student and distillation loss\n+- An optimizer for the student and (optional) metrics to evaluate performance\n+\n+In the `compute_loss` method, we perform a forward pass of both the teacher and student,\n+calculate the loss with weighting of the `student_loss` and `distillation_loss` by `alpha`\n+and `1 - alpha`, respectively. Note: only the student weights are updated.\n+\"\"\"\n+\n+\n+class Distiller(keras.Model):\n+ def __init__(self, student, teacher):\n+ super().__init__()\n+ self.teacher = teacher\n+ self.student = student\n+\n+ def compile(\n+ self,\n+ optimizer,\n+ metrics,\n+ student_loss_fn,\n+ distillation_loss_fn,\n+ alpha=0.1,\n+ temperature=3,\n+ ):\n+ \"\"\"Configure the distiller.\n+\n+ Args:\n+ optimizer: Keras optimizer for the student weights\n+ metrics: Keras metrics for evaluation\n+ student_loss_fn: Loss function of difference between student\n+ predictions and ground-truth\n+ distillation_loss_fn: Loss function of difference between soft\n+ student predictions and soft teacher predictions\n+ alpha: weight to student_loss_fn and 1-alpha to distillation_loss_fn\n+ temperature: Temperature for softening probability distributions.\n+ Larger temperature gives softer distributions.\n+ \"\"\"\n+ super().compile(optimizer=optimizer, metrics=metrics)\n+ self.student_loss_fn = student_loss_fn\n+ self.distillation_loss_fn = distillation_loss_fn\n+ self.alpha = alpha\n+ self.temperature = temperature\n+\n+ def compute_loss(\n+ self, x=None, y=None, y_pred=None, sample_weight=None, allow_empty=False\n+ ):\n+ teacher_pred = self.teacher(x, training=False)\n+ student_loss = self.student_loss_fn(y, y_pred)\n+\n+ distillation_loss = self.distillation_loss_fn(\n+ ops.softmax(teacher_pred / self.temperature, axis=1),\n+ ops.softmax(y_pred / self.temperature, axis=1),\n+ ) * (self.temperature**2)\n+\n+ loss = self.alpha * student_loss + (1 - self.alpha) * distillation_loss\n+ return loss\n+\n+ def call(self, x):\n+ return self.student(x)\n+\n+\n+\"\"\"\n+## Create student and teacher models\n+\n+Initialy, we create a teacher model and a smaller student model. Both models are\n+convolutional neural networks and created using `Sequential()`,\n+but could be any Keras model.\n+\"\"\"\n+\n+# Create the teacher\n+teacher = keras.Sequential(\n+ [\n+ keras.Input(shape=(28, 28, 1)),\n+ layers.Conv2D(256, (3, 3), strides=(2, 2), padding=\"same\"),\n+ layers.LeakyReLU(negative_slope=0.2),\n+ layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding=\"same\"),\n+ layers.Conv2D(512, (3, 3), strides=(2, 2), padding=\"same\"),\n+ layers.Flatten(),\n+ layers.Dense(10),\n+ ],\n+ name=\"teacher\",\n+)\n+\n+# Create the student\n+student = keras.Sequential(\n+ [\n+ keras.Input(shape=(28, 28, 1)),\n+ layers.Conv2D(16, (3, 3), strides=(2, 2), padding=\"same\"),\n+ layers.LeakyReLU(negative_slope=0.2),\n+ layers.MaxPooling2D(pool_size=(2, 2), strides=(1, 1), padding=\"same\"),\n+ layers.Conv2D(32, (3, 3), strides=(2, 2), padding=\"same\"),\n+ layers.Flatten(),\n+ layers.Dense(10),\n+ ],\n+ name=\"student\",\n+)\n+\n+# Clone student for later comparison\n+student_scratch = keras.models.clone_model(student)\n+\n+\"\"\"\n+## Prepare the dataset\n+\n+The dataset used for training the teacher and distilling the teacher is\n+[MNIST](https://keras.io/api/datasets/mnist/), and the procedure would be equivalent for\n+any other\n+dataset, e.g. [CIFAR-10](https://keras.io/api/datasets/cifar10/), with a suitable choice\n+of models. Both the student and teacher are trained on the training set and evaluated on\n+the test set.\n+\"\"\"\n+\n+# Prepare the train and test dataset.\n+batch_size = 64\n+(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()\n+\n+# Normalize data\n+x_train = x_train.astype(\"float32\") / 255.0\n+x_train = np.reshape(x_train, (-1, 28, 28, 1))\n+\n+x_test = x_test.astype(\"float32\") / 255.0\n+x_test = np.reshape(x_test, (-1, 28, 28, 1))\n+\n+\n+\"\"\"\n+## Train the teacher\n+\n+In knowledge distillation we assume that the teacher is trained and fixed. Thus, we start\n+by training the teacher model on the training set in the usual way.\n+\"\"\"\n+\n+# Train teacher as usual\n+teacher.compile(\n+ optimizer=keras.optimizers.Adam(),\n+ loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),\n+ metrics=[keras.metrics.SparseCategoricalAccuracy()],\n+)\n+\n+# Train and evaluate teacher on data.\n+teacher.fit(x_train, y_train, epochs=5)\n+teacher.evaluate(x_test, y_test)\n+\n+\"\"\"\n+## Distill teacher to student\n+\n+We have already trained the teacher model, and we only need to initialize a\n+`Distiller(student, teacher)` instance, `compile()` it with the desired losses,\n+hyperparameters and optimizer, and distill the teacher to the student.\n+\"\"\"\n+\n+# Initialize and compile distiller\n+distiller = Distiller(student=student, teacher=teacher)\n+distiller.compile(\n+ optimizer=keras.optimizers.Adam(),\n+ metrics=[keras.metrics.SparseCategoricalAccuracy()],\n+ student_loss_fn=keras.losses.SparseCategoricalCrossentropy(\n+ from_logits=True\n+ ),\n+ distillation_loss_fn=keras.losses.KLDivergence(),\n+ alpha=0.1,\n+ temperature=10,\n+)\n+\n+# Distill teacher to student\n+distiller.fit(x_train, y_train, epochs=3)\n+\n+# Evaluate student on test dataset\n+distiller.evaluate(x_test, y_test)\n+\n+\"\"\"\n+## Train student from scratch for comparison\n+\n+We can also train an equivalent student model from scratch without the teacher, in order\n+to evaluate the performance gain obtained by knowledge distillation.\n+\"\"\"\n+\n+# Train student as doen usually\n+student_scratch.compile(\n+ optimizer=keras.optimizers.Adam(),\n+ loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),\n+ metrics=[keras.metrics.SparseCategoricalAccuracy()],\n+)\n+\n+# Train and evaluate student trained from scratch.\n+student_scratch.fit(x_train, y_train, epochs=3)\n+student_scratch.evaluate(x_test, y_test)\n+\n+\"\"\"\n+If the teacher is trained for 5 full epochs and the student is distilled on this teacher\n+for 3 full epochs, you should in this example experience a performance boost compared to\n+training the same student model from scratch, and even compared to the teacher itself.\n+You should expect the teacher to have accuracy around 97.6%, the student trained from\n+scratch should be around 97.6%, and the distilled student should be around 98.1%. Remove\n+or try out different seeds to use different weight initializations.\n+\"\"\"\n"} |