🔧 Add Apple Silicon (M1) Compatibility — Dependency & API Updates

common/utils/models_utils.py

@@ -9,7 +9,6 @@
 
 import os
 from tabulate import tabulate
-from keras.utils.layer_utils import count_params
 from typing import Dict, Optional, Tuple, List
 import tensorflow as tf
 from tensorflow.keras.models import Model
@@ -20,6 +19,7 @@
 import onnxruntime
 import mlflow
 from common.utils import log_to_file
+from tensorflow.keras import backend as K
 
 
 def ai_interp_input_quant(ai_interp, data: np.array, data_scale: float, data_offset: float, file_extension: str):
@@ -286,7 +286,11 @@ def model_summary(model):
         if layer_type == "InputLayer":
             layer_shape = model.input.shape
         else:
-            layer_shape = layer.output_shape
+            if hasattr(layer, 'output') and hasattr(layer.output, 'shape'):
+                layer_shape = layer.output.shape
+            else:
+                layer_shape = 'Not available'
+
         is_trainable = True if layer.trainable else False
         num_params = layer.count_params()
         if layer.trainable:
@@ -300,8 +304,8 @@ def model_summary(model):
     print(tabulate(table, headers=["Layer index", "Trainable", "Name", "Type", "Params#", "Output shape"]))
     print(108 * '-')
     print("Total params:", model.count_params())
-    print("Trainable params: ", count_params(model.trainable_weights))
-    print("Non-trainable params: ", count_params(model.non_trainable_weights))
+    print("Trainable params: ", sum([K.count_params(w) for w in model.trainable_weights]))
+    print("Non-trainable params: ", sum([K.count_params(w) for w in model.non_trainable_weights]))
     print(108 * '-')
     print("Total layers:", num_layers)
     print("Trainable layers:", trainable_layers)
@@ -443,4 +447,4 @@ def compute_confusion_matrix(test_set: tf.data.Dataset = None, model: Model = No
 
     # Calculate the confusion matrix.
     cm = sklearn.metrics.confusion_matrix(labels, logits)
-    return cm, test_accuracy
+    return cm, test_accuracy

image_classification/src/data_augmentation/data_augmentation_layer.py

@@ -10,21 +10,18 @@
 import tensorflow as tf
 from tensorflow import keras
 from keras import backend
-from keras.engine import base_layer
-from keras.engine import base_preprocessing_layer
-from keras.utils import control_flow_util
+from keras.layers import Layer
 from src.data_augmentation import data_augmentation
 
 
-class DataAugmentationLayer(base_layer.Layer):
+class DataAugmentationLayer(Layer):
 
     def __init__(self,
                 data_augmentation_fn,
                 config=None,
                 pixels_range=None,
                 batches_per_epoch=None,
                 **kwargs):
-        base_preprocessing_layer.keras_kpl_gauge.get_cell('DataAugmentationLayer').set(True)
         super(DataAugmentationLayer, self).__init__(**kwargs)
         self.data_augmentation_fn = data_augmentation_fn
         self.config_dict = config
@@ -70,7 +67,11 @@ def transform_input_data():
 
             return outputs
 
-        return control_flow_util.smart_cond(training, transform_input_data, lambda: inputs)
+        if training:
+            return transform_input_data()
+        else:
+            return inputs
+
 
 
     def get_config(self):

image_classification/src/models/mobilenetv2.py

@@ -10,8 +10,26 @@
 
 import tensorflow as tf
 from tensorflow import keras
-from keras.applications import imagenet_utils
 from tensorflow.keras import layers
+from tensorflow.keras import backend as K
+
+def correct_pad(inputs, kernel_size):
+    """Returns a tuple for zero-padding for 2D conv with 'same' padding."""
+    img_dim = 1 if K.image_data_format() == 'channels_first' else 2
+    input_size = K.int_shape(inputs)[img_dim:img_dim + 2]
+    if isinstance(kernel_size, int):
+        kernel_size = (kernel_size, kernel_size)
+    adjust = (1, 1)
+    correct = ((kernel_size[0] // 2, kernel_size[0] // 2),
+               (kernel_size[1] // 2, kernel_size[1] // 2))
+    if input_size[0] is not None and input_size[0] % 2 == 0:
+        correct = ((kernel_size[0] // 2 - 1, kernel_size[0] // 2),
+                   correct[1])
+    if input_size[1] is not None and input_size[1] % 2 == 0:
+        correct = (correct[0],
+                   (kernel_size[1] // 2 - 1, kernel_size[1] // 2))
+    return correct
+
 
 
 def _make_divisible(v, divisor, min_value=None):
@@ -60,7 +78,7 @@ def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id):
 
     """
     # Get the shape of the input tensor.
-    shape = inputs.get_shape().as_list()
+    shape = inputs.shape
     # Get the number of channels in the input tensor.
     in_channels = shape[-1]
     # Calculate the number of filters for the pointwise 1x1 convolution.
@@ -80,7 +98,7 @@ def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id):
     # Depthwise 3x3 convolution.
     if stride == 2:
         # If the stride is 2, pad the input tensor to maintain the same output size.
-        x = layers.ZeroPadding2D(padding=imagenet_utils.correct_pad(x, 3))(x)
+        x = layers.ZeroPadding2D(padding=correct_pad(x, 3))(x)
     # Perform the depthwise 3x3 convolution.
     x = layers.DepthwiseConv2D(kernel_size=3, strides=stride, use_bias=False,
                                padding='same' if stride == 1 else 'valid')(x)

requirements.txt

@@ -1,31 +1,34 @@
-tensorflow==2.8.3
-mlflow==2.3.0
+tensorflow==2.16.2; sys_platform == "linux"
+tensorflow-macos==2.16.2; sys_platform == "darwin" and platform_machine == "arm64"
+tensorflow-metal==1.2.0; sys_platform == "darwin" and platform_machine == "arm64"
+mlflow==2.22.0
 hydra-core==1.3.2
 omegaconf==2.3.0
-numpy==1.23.4
-pandas==1.5.3
+numpy==1.26.4
+pandas==2.2.3
 wget==3.2
-munch==2.5.0
-seaborn==0.12.2
-scikit-learn==1.2.2
-tqdm==4.65.0
-skl2onnx==1.14.0
-onnx==1.12.0
+munch==4.0.0
+seaborn==0.13.2
+scikit-learn==1.6.1
+tqdm==4.67.1
+skl2onnx==1.18.0
+onnx==1.17.0
 onnxconverter-common==1.13.0
-onnxruntime==1.15.1
-optuna==3.1.1
+onnxruntime==1.22.0
+optuna==4.3.0
 imgaug==0.4.0
-matplotlib==3.6.2
+matplotlib==3.10.3
 opencv_python==4.6.0.66
-joblib==1.2.0
-librosa==0.10.0.post2
-mako==1.2.4
+joblib==1.5.0
+librosa==0.11.0
+mako==1.3.10
 pyserial==3.5
-colorama==0.4.5
-Requests==2.28.2
-urllib3==1.26.13
+colorama==0.4.6
+requests==2.32.3
+urllib3==2.4.0
 xmlrunner==1.7.7
-marshmallow==3.20.1
-tf2onnx==1.14.0
-clearml==1.16.5
-pip-system-certs==4.0
+marshmallow==4.0.0
+tf2onnx==1.16.1
+clearml==1.18.0
+pip-system-certs==4.0
+tabulate==0.9.0