Added building a Docker image for the course

.github/workflows/docker.yml

@@ -0,0 +1,33 @@
+name: Build and publish Docker image for the project
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+    branches:
+      - main
+
+jobs:
+  Docker:
+    if: github.repository == 'antmicro/dl-in-iot-course'
+    runs-on: ubuntu-latest
+    steps:
+      - name: Cancel previous run
+        uses: styfle/[email protected]
+        with:
+          access_token: ${{ github.token }}
+      - name: Checkout sources
+        uses: actions/checkout@v4
+      - name: Build Docker image
+        run: docker build . -f environments/Dockerfile --tag ghcr.io/${{ github.repository }}
+      - name: Login to registry
+        if: github.ref == 'refs/heads/main' && github.event_name != 'pull_request'
+        uses: docker/login-action@v3
+        with:
+          registry: ghcr.io
+          username: ${{ github.actor }}
+          password: ${{ github.token }}
+      - name: Push image to registry
+        if: github.ref == 'refs/heads/main' && github.event_name != 'pull_request'
+        run: docker push ghcr.io/${{ github.repository }}

.github/workflows/pre-commit.yml

@@ -0,0 +1,20 @@
+name: pre-commit checks
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+jobs:
+  pre-commit:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout sources
+        uses: actions/checkout@v4
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: 3.11
+      - name: Run pre-commit
+        uses: pre-commit/[email protected]

.pre-commit-config.yaml

@@ -0,0 +1,12 @@
+repos:
+- repo: https://github.com/pre-commit/pre-commit-hooks
+  rev: v4.4.0
+  hooks:
+  - id: trailing-whitespace
+- repo: https://github.com/astral-sh/ruff-pre-commit
+  rev: v0.1.2
+  hooks:
+  - id: ruff
+    args: [--fix, --exit-non-zero-on-fix]
+  - id: ruff-format
+exclude: .+\.patch

README.md

@@ -1,8 +1,8 @@
 # Optimization of Deep Learning applications for IoT devices - Course tasks
 
-Copyright (c) 2021-2022 [Antmicro](https://www.antmicro.com)
+Copyright (c) 2021-2024 [Antmicro](https://www.antmicro.com)
 
-This repository contains tasks for laboratories for the "Optimization of Deep Learning applications for IoT devices" course.
+This repository contains tasks for laboratories for the "Optimization of Neural Network applications for IoT devices" course.
 
 ## Course classes
 
@@ -14,6 +14,7 @@ Please follow the links to go to the list of tasks:
 * [Lab 04 - Introduction to Apache TVM](dl_in_iot_course/l04_tvm)
 * [Lab 05 - Implementing a TensorFlow Lite delegate](dl_in_iot_course/l05_tflite_delegate)
 * [Lab 06 - Fine-tuning of model and operations in Apache TVM](dl_in_iot_course/l06_tvm_fine_tuning)
+* [Lab 07 - Accelerating ML models on FPGAs with TFLite Micro and CFU Playground](cfu-playground)
 
 ## Cloning the repository
 

dl_in_iot_course/__init__.py

@@ -7,4 +7,4 @@
 import os
 import sys
 
-sys.path.insert(0, os.path.abspath(__file__ + '../'))
+sys.path.insert(0, os.path.abspath(__file__ + "../"))

dl_in_iot_course/l02_quantization/README.md

@@ -50,7 +50,7 @@ It requires implementing methods for:
 * `[2pt]` Finish the `ImbalancedINT8Model` class:
 
     * Implement `optimize_model` method, where the `calibration_dataset_generator` will take all examples for objects with 5 class and use them for calibration:
-        
+
         * Use `self.dataset.dataX` and `self.dataset.dataY` to extract all inputs for a particular class.
         * Remember to use self.dataset.prepare_input_sample method.
 
@@ -64,7 +64,7 @@ It requires implementing methods for:
   ```
 
   In the `build/results` directory, the script will create:
-    
+
     * `<prefix>-metrics.md` file - contains basic metrics, such as accuracy, precision, sensitivity or G-Mean, along with inference time
     * `<prefix>-confusion-matrix.png` file - contains visualization of confusion matrix for the model evaluation.
   Those files will be created for:

dl_in_iot_course/l02_quantization/model_training.py

@@ -12,13 +12,13 @@ def __init__(self, modelpath: Path, dataset: PetDataset, from_file=True):
         self.from_file = from_file
         self.numclasses = dataset.numclasses
         self.mean, self.std = dataset.get_input_mean_std()
-        self.inputspec = tf.TensorSpec((1, 224, 224, 3), name='input_1')
+        self.inputspec = tf.TensorSpec((1, 224, 224, 3), name="input_1")
         self.dataset = dataset
         self.prepare()
 
     def load_model(self):
         tf.keras.backend.clear_session()
-        if hasattr(self, 'model') and self.model is not None:
+        if hasattr(self, "model") and self.model is not None:
             del self.model
         self.model = tf.keras.models.load_model(str(self.modelpath))
 
@@ -30,43 +30,23 @@ def prepare(self):
             self.load_model()
         else:
             self.base = tf.keras.applications.MobileNetV2(
-                input_shape=(224, 224, 3),
-                include_top=False,
-                weights='imagenet'
+                input_shape=(224, 224, 3), include_top=False, weights="imagenet"
             )
             self.base.trainable = False
-            avgpool = tf.keras.layers.GlobalAveragePooling2D()(
-                self.base.output
-            )
-            layer1 = tf.keras.layers.Dense(
-                1024,
-                activation='relu')(avgpool)
+            avgpool = tf.keras.layers.GlobalAveragePooling2D()(self.base.output)
+            layer1 = tf.keras.layers.Dense(1024, activation="relu")(avgpool)
             d1 = tf.keras.layers.Dropout(0.3)(layer1)
-            layer2 = tf.keras.layers.Dense(
-                512,
-                activation='relu')(d1)
+            layer2 = tf.keras.layers.Dense(512, activation="relu")(d1)
             d2 = tf.keras.layers.Dropout(0.3)(layer2)
-            layer3 = tf.keras.layers.Dense(
-                128,
-                activation='relu')(d2)
+            layer3 = tf.keras.layers.Dense(128, activation="relu")(d2)
             d3 = tf.keras.layers.Dropout(0.3)(layer3)
-            output = tf.keras.layers.Dense(
-                self.numclasses,
-                name='out_layer'
-            )(d3)
-            self.model = tf.keras.models.Model(
-                inputs=self.base.input,
-                outputs=output
-            )
+            output = tf.keras.layers.Dense(self.numclasses, name="out_layer")(d3)
+            self.model = tf.keras.models.Model(inputs=self.base.input, outputs=output)
         print(self.model.summary())
 
     def train_model(
-            self,
-            batch_size: int,
-            learning_rate: int,
-            epochs: int,
-            logdir: Path):
-
+        self, batch_size: int, learning_rate: int, epochs: int, logdir: Path
+    ):
         def preprocess_input(path, onehot):
             data = tf.io.read_file(path)
             img = tf.io.decode_jpeg(data, channels=3)
@@ -79,93 +59,69 @@ def preprocess_input(path, onehot):
             img = (img - self.mean) / self.std
             return img, tf.convert_to_tensor(onehot)
 
-        Xt, Xv, Yt, Yv = self.dataset.split_dataset(
-            0.25
-        )
+        Xt, Xv, Yt, Yv = self.dataset.split_dataset(0.25)
         Yt = list(self.dataset.onehotvectors[Yt])
         Yv = list(self.dataset.onehotvectors[Yv])
         traindataset = tf.data.Dataset.from_tensor_slices((Xt, Yt))
         traindataset = traindataset.map(
-            preprocess_input,
-            num_parallel_calls=tf.data.experimental.AUTOTUNE
+            preprocess_input, num_parallel_calls=tf.data.experimental.AUTOTUNE
         ).batch(batch_size)
         validdataset = tf.data.Dataset.from_tensor_slices((Xv, Yv))
         validdataset = validdataset.map(
-            preprocess_input,
-            num_parallel_calls=tf.data.experimental.AUTOTUNE
+            preprocess_input, num_parallel_calls=tf.data.experimental.AUTOTUNE
         ).batch(batch_size)
 
         tensorboard_callback = tf.keras.callbacks.TensorBoard(
-            str(logdir),
-            histogram_freq=1
+            str(logdir), histogram_freq=1
         )
 
         model_checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
-            filepath=str(logdir / 'weights.{epoch:02d}-{val_loss:.2f}.h5'),
-            monitor='val_categorical_accuracy',
-            mode='max',
-            save_best_only=True
+            filepath=str(logdir / "weights.{epoch:02d}-{val_loss:.2f}.h5"),
+            monitor="val_categorical_accuracy",
+            mode="max",
+            save_best_only=True,
         )
 
         self.model.compile(
             optimizer=tf.keras.optimizers.Adam(lr=learning_rate),
             loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True),
-            metrics=[
-                tf.keras.metrics.CategoricalAccuracy()
-            ]
+            metrics=[tf.keras.metrics.CategoricalAccuracy()],
         )
 
         self.model.fit(
             traindataset,
             epochs=epochs,
-            callbacks=[
-                tensorboard_callback,
-                model_checkpoint_callback
-            ],
-            validation_data=validdataset
+            callbacks=[tensorboard_callback, model_checkpoint_callback],
+            validation_data=validdataset,
         )
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     parser = argparse.ArgumentParser()
+    parser.add_argument("--modelpath", help="Path to the model file", type=Path)
+    parser.add_argument("--dataset-root", help="Path to the dataset file", type=Path)
     parser.add_argument(
-        '--modelpath',
-        help='Path to the model file',
-        type=Path
-    )
-    parser.add_argument(
-        '--dataset-root',
-        help='Path to the dataset file',
-        type=Path
-    )
-    parser.add_argument(
-        '--download-dataset',
-        help='Download the dataset before training',
-        action='store_true'
+        "--download-dataset",
+        help="Download the dataset before training",
+        action="store_true",
     )
     parser.add_argument(
-        '--batch-size',
-        help='Batch size for the training process',
+        "--batch-size",
+        help="Batch size for the training process",
         type=int,
-        default=128
+        default=128,
     )
     parser.add_argument(
-        '--learning-rate',
-        help='Starting learning rate for Adam optimizer',
+        "--learning-rate",
+        help="Starting learning rate for Adam optimizer",
         type=float,
-        default=0.0001
+        default=0.0001,
     )
     parser.add_argument(
-        '--num-epochs',
-        help='Number of training epochs',
-        type=int,
-        default=50
+        "--num-epochs", help="Number of training epochs", type=int, default=50
     )
     parser.add_argument(
-        '--logdir',
-        help='The path to the logging directory',
-        type=Path,
-        default='logs'
+        "--logdir", help="The path to the logging directory", type=Path, default="logs"
     )
 
     args = parser.parse_args()
@@ -175,10 +131,5 @@ def preprocess_input(path, onehot):
 
     args.logdir.mkdir(parents=True, exist_ok=True)
 
-    model.train_model(
-        args.batch_size,
-        args.learning_rate,
-        args.num_epochs,
-        args.logdir
-    )
+    model.train_model(args.batch_size, args.learning_rate, args.num_epochs, args.logdir)
     model.save_model()