Update upstream

README.md

@@ -15,7 +15,7 @@
 
 ![Discord](https://img.shields.io/discord/662098530411741184.svg?logo=discord&colorB=7289DA)
 
-Donkeycar is minimalist and modular self driving library for Python. It is developed for hobbyists and students with a focus on allowing fast experimentation and easy community contributions.  It is being actively used at the high school and university level for learning and research.  It offers a [rich graphical interface](https://docs.donkeycar.com/utility/ui/) and includes a [simulator](https://docs.donkeycar.com/guide/deep_learning/simulator/) so you can experiment with self-driving even before you build a robot.
+Donkeycar is a minimalist and modular self driving library for Python. It is developed for hobbyists and students with a focus on allowing fast experimentation and easy community contributions.  It is being actively used at the high school and university level for learning and research.  It offers a [rich graphical interface](https://docs.donkeycar.com/utility/ui/) and includes a [simulator](https://docs.donkeycar.com/guide/deep_learning/simulator/) so you can experiment with self-driving even before you build a robot.
 
 #### Quick Links
 * [Donkeycar Updates & Examples](http://donkeycar.com)

donkeycar/__init__.py

@@ -3,7 +3,7 @@
 from pyfiglet import Figlet
 import logging
 
-__version__ = '5.2.dev2'
+__version__ = '5.3.dev1'
 
 logging.basicConfig(level=os.environ.get('LOGLEVEL', 'INFO').upper())
 

donkeycar/parts/camera.py

@@ -35,7 +35,7 @@ def __init__(self, image_w=160, image_h=120, image_d=3,
         self.camera = Picamera2()
         config = self.camera.create_preview_configuration(
             config_dict, transform=transform)
-        self.camera.align_configuration(config)
+        # self.camera.align_configuration(config)   # this created issues with the libcamera2 library on the Pi5, which automatically changes the resolution to 128x120 in an attempt to align with native sensor resolution
         self.camera.configure(config)
         # try min / max frame rate as 0.1 / 1 ms (it will be slower though)
         self.camera.set_controls({"FrameDurationLimits": (100, 1000)})

donkeycar/parts/interpreter.py

@@ -4,14 +4,38 @@
 import numpy as np
 from typing import Union, Sequence, List
 
-import tensorflow as tf
-from tensorflow import keras
-from tensorflow.python.saved_model import tag_constants, signature_constants
-from tensorflow.python.compiler.tensorrt import trt_convert as trt
+try:
+    import tensorflow as tf
+    from tensorflow import keras
+    from tensorflow.python.saved_model import tag_constants, signature_constants
+    from tensorflow.python.compiler.tensorrt import trt_convert as trt
+except ImportError:
+    tf = None
+    keras = None
+    tag_constants = None
+    signature_constants = None
+    trt = None
 
 logger = logging.getLogger(__name__)
 
 
+def get_tflite_interpreter():
+    """Get TFLite Interpreter from tflite-runtime or full TensorFlow."""
+    try:
+        from tflite_runtime.interpreter import Interpreter
+        return Interpreter
+    except ImportError:
+        pass
+    try:
+        from ai_edge_litert.interpreter import Interpreter
+        return Interpreter
+    except ImportError:
+        pass
+    if tf is not None:
+        return tf.lite.Interpreter
+    raise ImportError("No TFLite runtime found. Install tflite-runtime or tensorflow.")
+
+
 def has_trt_support():
     try:
         converter = trt.TrtGraphConverterV2()
@@ -91,14 +115,14 @@ def set_model(self, pilot: 'KerasPilot') -> None:
         """ Some interpreters will need the model"""
         pass
 
-    def set_optimizer(self, optimizer: tf.keras.optimizers.Optimizer) -> None:
+    def set_optimizer(self, optimizer) -> None:
         pass
 
     def compile(self, **kwargs):
         raise NotImplementedError('Requires implementation')
 
     @abstractmethod
-    def get_input_shape(self, input_name) -> tf.TensorShape:
+    def get_input_shape(self, input_name):
         pass
 
     def predict(self, img_arr: np.ndarray, *other_arr: np.ndarray) \
@@ -127,7 +151,7 @@ class KerasInterpreter(Interpreter):
 
     def __init__(self):
         super().__init__()
-        self.model: tf.keras.Model = None
+        self.model = None
 
     def set_model(self, pilot: 'KerasPilot') -> None:
         self.model = pilot.create_model()
@@ -146,10 +170,10 @@ def set_model(self, pilot: 'KerasPilot') -> None:
         self.shapes = (dict(zip(self.input_keys, input_shape)),
                        dict(zip(self.output_keys, output_shape)))
 
-    def set_optimizer(self, optimizer: tf.keras.optimizers.Optimizer) -> None:
+    def set_optimizer(self, optimizer) -> None:
         self.model.optimizer = optimizer
 
-    def get_input_shape(self, input_name) -> tf.TensorShape:
+    def get_input_shape(self, input_name):
         assert self.model, 'Model not set'
         return self.shapes[0][input_name]
 
@@ -174,6 +198,18 @@ def predict_from_dict(self, input_dict):
     def load(self, model_path: str) -> None:
         logger.info(f'Loading model {model_path}')
         self.model = keras.models.load_model(model_path, compile=False)
+        # Set input_keys and output_keys after loading (same as set_model)
+        input_shape = self.model.input_shape
+        if type(input_shape) is not list:
+            input_shape = [input_shape]
+        output_shape = self.model.output_shape
+        if type(output_shape) is not list:
+            output_shape = [output_shape]
+
+        self.input_keys = self.model.input_names
+        self.output_keys = self.model.output_names
+        self.shapes = (dict(zip(self.input_keys, input_shape)),
+                       dict(zip(self.output_keys, output_shape)))
 
     def load_weights(self, model_path: str, by_name: bool = True) -> \
             None:
@@ -263,7 +299,8 @@ def load(self, model_path):
             'TFlitePilot should load only .tflite files'
         logger.info(f'Loading model {model_path}')
         # Load TFLite model and extract input and output keys
-        self.interpreter = tf.lite.Interpreter(model_path=model_path)
+        Interpreter = get_tflite_interpreter()
+        self.interpreter = Interpreter(model_path=model_path)
         self.signatures = self.interpreter.get_signature_list()
         self.runner = self.interpreter.get_signature_runner()
         self.input_keys = self.signatures['serving_default']['inputs']
@@ -312,7 +349,7 @@ def set_model(self, pilot: 'KerasPilot') -> None:
         # state as the trt model hasn't been loaded yet
         self.pilot = pilot
 
-    def get_input_shape(self, input_name) -> tf.TensorShape:
+    def get_input_shape(self, input_name):
         assert self.graph_func, "Requires loadin the tensorrt model first"
         return self.graph_func.structured_input_signature[1][input_name].shape
 

donkeycar/parts/keras.py

@@ -15,23 +15,27 @@
 from typing import Dict, Tuple, Optional, Union, List, Sequence, Callable, Any
 from logging import getLogger
 
-from tensorflow.python.data.ops.dataset_ops import DatasetV1, DatasetV2
-
 import donkeycar as dk
 from donkeycar.utils import normalize_image, linear_bin
 from donkeycar.pipeline.types import TubRecord
 from donkeycar.parts.interpreter import Interpreter, KerasInterpreter
 
-import tensorflow as tf
-from tensorflow import keras
-from tensorflow.keras.layers import (Dense, Input,Convolution2D,
-    MaxPooling2D, Activation, Dropout, Flatten, LSTM, BatchNormalization,
-    Conv3D, MaxPooling3D, Conv2DTranspose)
-
-from tensorflow.keras.layers import TimeDistributed as TD
-from tensorflow.keras.backend import concatenate
-from tensorflow.keras.models import Model
-from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
+try:
+    import tensorflow as tf
+    from tensorflow import keras
+    from tensorflow.python.data.ops.dataset_ops import DatasetV1, DatasetV2
+    from tensorflow.keras.layers import (Dense, Input, Convolution2D,
+        MaxPooling2D, Activation, Dropout, Flatten, LSTM, BatchNormalization,
+        Conv3D, MaxPooling3D, Conv2DTranspose)
+    from tensorflow.keras.layers import TimeDistributed as TD
+    from tensorflow.keras.backend import concatenate
+    from tensorflow.keras.models import Model
+    from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint
+except ImportError:
+    tf = None
+    keras = None
+    DatasetV1 = None
+    DatasetV2 = None
 
 ONE_BYTE_SCALE = 1.0 / 255.0
 
@@ -86,7 +90,7 @@ def set_optimizer(self, optimizer_type: str,
             raise Exception(f"Unknown optimizer type: {optimizer_type}")
         self.interpreter.set_optimizer(optimizer)
 
-    def get_input_shape(self, input_name) -> tf.TensorShape:
+    def get_input_shape(self, input_name):
         return self.interpreter.get_input_shape(input_name)
 
     def seq_size(self) -> int:
@@ -110,9 +114,8 @@ def run(self, img_arr: np.ndarray, *other_arr: List[float]) \
         # self.output_shape() first dictionary keys, because that's how we
         # set up the model
         values = (norm_img_arr, ) + np_other_array
-        # note output_shapes() returns a 2-tuple of dicts for input shapes
-        # and output shapes(), so we need the first tuple here
-        input_dict = dict(zip(self.output_shapes()[0].keys(), values))
+        # use interpreter's input_keys directly (works with TFLite and Keras)
+        input_dict = dict(zip(self.interpreter.input_keys, values))
         return self.inference_from_dict(input_dict)
 
     def inference_from_dict(self, input_dict: Dict[str, np.ndarray]) \
@@ -146,7 +149,7 @@ def train(self,
               verbose: int = 1,
               min_delta: float = .0005,
               patience: int = 5,
-              show_plot: bool = False) -> tf.keras.callbacks.History:
+              show_plot: bool = False):
         """
         trains the model
         """
@@ -242,7 +245,7 @@ def output_types(self) -> Tuple[Dict[str, np.typename], ...]:
         types = tuple({k: tf.float64 for k in d} for d in shapes)
         return types
 
-    def output_shapes(self) -> Dict[str, tf.TensorShape]:
+    def output_shapes(self):
         return {}
 
     def __str__(self) -> str: