Convert the Inception v4 model into coreML model

Author: DevOpsThinh

.gitignore

@@ -1,2 +1,4 @@
+inception_v4_dogscats.mlmodel
+inceptionv4_dogscats_weights.h5
 us-shareprices-daily.csv
 /.idea

ml/deep_learning/keras_functional_api/converting_keras_inceptionv4_model_coreml.py

@@ -0,0 +1,31 @@
+# Author/ Learner: Nguyen Truong Thinh
+# Contact me: nguyentruongthinhvn2020@gmail.com || +84393280504
+#
+# Use case: Create a Keras Inception v4 model that can be used to convert into the Core ML Format
+#                 via CoreML Tool .
+#                  The model will be trained & converted on the Kaggle Dogs vs Cats dataset via Keras functional api.
+
+import coremltools as ct
+
+from ml.deep_learning.keras_functional_api.inception_v4_network import model
+
+weights_path = 'inceptionv4_dogscats_weights.h5'
+model.load_weights(weights_path)
+
+coreml_model = ct.convert(model, source="tensorflow", minimum_deployment_target=ct.target.iOS13)
+
+coreml_model.author = 'Nguyen Truong Thinh'
+coreml_model.short_description = 'An Inception v4 model trained on the Kaggle Dogs vs Cats dataset via Keras ' \
+                                 'functional api.'
+coreml_model.save("inception_v4_dogscats.mlmodel")
+
+"""
+Running TensorFlow Graph Passes: 100%|██████████| 6/6 [00:00<00:00,  8.02 passes/s]
+Converting TF Frontend ==> MIL Ops: 100%|██████████| 1258/1258 [00:00<00:00, 1981.15 ops/s]
+Running MIL frontend_tensorflow2 pipeline: 100%|██████████| 7/7 [00:00<00:00, 155.55 passes/s]
+Running MIL default pipeline: 100%|██████████| 56/56 [00:04<00:00, 12.48 passes/s]
+Running MIL backend_neuralnetwork pipeline: 100%|██████████| 8/8 [00:00<00:00, 256.00 passes/s]
+Translating MIL ==> NeuralNetwork Ops: 100%|██████████| 1538/1538 [00:02<00:00, 647.43 ops/s]
+
+Process finished with exit code 0
+"""

ml/deep_learning/keras_functional_api/helper_function.py

@@ -175,10 +175,10 @@ def build_inception_b_block(input_tensor):
     """
     # (384 1x1 convolutions) - This is the first branch for the left
     branch_a = conv2d_batch_norm_relu(input_tensor, 384, 1, 1)
-    # This is the second branch for the left
+    # This is the second branch from the left
     branch_b = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(input_tensor)
     branch_b = conv2d_batch_norm_relu(branch_b, 128, 1, 1)
-    # This  is the third branch from the left
+    # This is the third branch from the left
     branch_c = conv2d_batch_norm_relu(input_tensor, 192, 1, 1)
     branch_c = conv2d_batch_norm_relu(branch_c, 224, 1, 7)
     branch_c = conv2d_batch_norm_relu(branch_c, 256, 7, 1)

ml/deep_learning/keras_functional_api/inception_v4_network.py

@@ -39,7 +39,7 @@
 # Use the Adam optimizer and compile the model.
 adam_opt = Adam(lr=INITIAL_LEARNING_RATE)
 model = Model(INPUT_TENSOR, network_output, name='InceptionV4')
-model.compile(optimizer=adam_opt, loss='binary_crossentropy', metrics=["accuracy"])
+compiled_model = model.compile(optimizer=adam_opt, loss='binary_crossentropy', metrics=["accuracy"])
 # Display a summary of the layers of the model.
 model.summary()
 # ===========================================

ml/deep_learning/keras_functional_api/training_the_inception_v4_model.py

@@ -15,16 +15,18 @@
 from sklearn.model_selection import train_test_split
 
 from inception_v4_network import model
+from ml.deep_learning.keras_functional_api.helper_function import plot_training_stats
 
 # ----------------- A computer without an NVIDIA GPU,
-# then uncomment the first two statements
-# weights_path = './inceptionv4.dogscats_weights.h5'
+# Download it at: https://drive.google.com/file/d/1NqMs2js-uOyOLL9MB0nzX0AoHx-hy8ty/view?usp=sharing
+# Place it on the following path: "ml/deep_learning/keras_functional_api/inceptionv4.dogscats_weights.h5"
+# then uncomment the first two statements:
+
+# weights_path = 'inceptionv4.dogscats_weights.h5'
 # model.load_weights(weights_path, by_name=True)
 
 # ----------------- A computer have an NVIDIA GPU
 # Adjust these to match the dimensions of our input image.
-from ml.deep_learning.keras_functional_api.helper_function import plot_training_stats
-
 IMAGE_HEIGHT = 299
 IMAGE_WIDTH = 299
 # Reduce this if this model does not fit on your GPU.
@@ -114,3 +116,10 @@
 
 end_time = datetime.datetime.now()
 print(f'training completed in: {end_time - start_time}')
+
+"""
+=> Attention! (Note):
+The Inception-v4 model was trained on a Windows 10 computer workstation with 64GB of RAM 
+and an NVIDIA RTX 2080Ti 11Gb graphics card. It took about three hours to train 30 epochs, 
+with the best model having an accuracy of 87.9 percent.
+"""