Surface Crack Detection And Localization - Transfer learning

Naveen / Surface Crack Detection And Localization Public

Target: Cortex-M4F 80MHz Clone this project

Neural Network settings

Training settings

Number of training cycles

Please provide a valid number of training cycles (numeric only)

Use learned optimizer

Learning rate

Please provide a valid number for the learning rate (between 0 and 1)

Training processor

Please provide a valid training processor option

Data augmentation

Advanced training settings

Validation set size

Please provide a valid number for the train/validate split (between 0 and 1)

Split train/validation set on metadata key

Batch size

Auto-weight classes

Profile int8 model

Neural network architecture

Save

import math from pathlib import Path import tensorflow as tf from tensorflow.keras import Model from tensorflow.keras.layers import Dense, GlobalAveragePooling2D, UpSampling2D, Permute, Reshape, Softmax, Add from tensorflow.keras.optimizers import Adam, Adadelta from tensorflow.keras.losses import categorical_crossentropy sys.path.append('./resources/libraries') import ei_tensorflow.training WEIGHTS_PATH = './transfer-learning-weights/keras/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_1.0_160.h5' INPUT_SHAPE = (160, 160, 3) base_model = tf.keras.applications.MobileNetV2( input_shape = INPUT_SHAPE, alpha=1, weights = WEIGHTS_PATH ) last_layer = base_model.layers[-2].output dense_layer = Dense(classes) output_pred = Softmax(name="prediction")(dense_layer(last_layer)) conv_layer = base_model.layers[-4].output reshape_layer = Reshape((conv_layer.shape[1] * conv_layer.shape[2] , -1))(conv_layer) dot_output = dense_layer(reshape_layer) transpose = Permute((2, 1))(dot_output) reshape_2_layer = Reshape((-1, conv_layer.shape[1] , conv_layer.shape[2]))(transpose) SIZE = (int(INPUT_SHAPE[1] / conv_layer.shape[2]), int(INPUT_SHAPE[0] / conv_layer.shape[1])) output_act_map = UpSampling2D(size=SIZE, interpolation="bilinear", data_format="channels_first", name="activation_map")(reshape_2_layer) model = Model(inputs=base_model.inputs, outputs=[output_pred, output_act_map]) TRAINABLE_START_IDX = -12 for layer in model.layers[:TRAINABLE_START_IDX]: layer.trainable = False print(model.summary()) print("Debug:", model.layers[TRAINABLE_START_IDX].name, model.layers[TRAINABLE_START_IDX].trainable) model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005), loss={'prediction': 'categorical_crossentropy', 'activation_map': None }, metrics={'prediction': ['accuracy'], 'activation_map': [None] }) BATCH_SIZE = 32 EPOCHS=30 train_dataset = train_dataset.batch(BATCH_SIZE, drop_remainder=False) validation_dataset = validation_dataset.batch(BATCH_SIZE, drop_remainder=False) callbacks.append(BatchLoggerCallback(BATCH_SIZE, train_sample_count, epochs=EPOCHS)) model.fit(train_dataset, validation_data=validation_dataset, epochs=EPOCHS, verbose=2, callbacks=callbacks) print('') print('Training done.', flush=True)

Input layer (76,800 features)

MobileNetV2 96x96 0.35 (final layer: 16 neurons, 0.1 dropout)

Output layer (3 classes)

Model

Did you know? You can customize your model through the Expert view (click on to switch),
or can even bring your own model (in PyTorch, Keras or scikit-learn).

Model	Author
MobileNetV1 96x96 0.25 Officially supported A pre-trained multi-layer convolutional network designed to efficiently classify images. Uses around 105.9K RAM and 301.6K ROM with default settings and optimizations.	Edge Impulse
MobileNetV1 96x96 0.2 Officially supported Uses around 83.1K RAM and 218.3K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV1 96x96 0.1 Officially supported Uses around 53.2K RAM and 101K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 96x96 0.35 Officially supported Uses around 296.8K RAM and 575.2K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 96x96 0.1 Officially supported Uses around 270.2K RAM and 212.3K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 96x96 0.05 Officially supported Uses around 265.3K RAM and 162.4K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 160x160 1.0 Officially supported Uses around 1.3M RAM and 2.6M ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
MobileNetV2 160x160 0.75 Officially supported Uses around 1.3M RAM and 1.7M ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
MobileNetV2 160x160 0.5 Officially supported Uses around 700.7K RAM and 982.4K ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
MobileNetV2 160x160 0.35 Officially supported Uses around 683.3K RAM and 658.4K ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
EfficientNet B0 Community Transfer learning model based on efficientnetb0_notop.h5 weights. This is a much larger model than MobileNet for Linux devices and accelerators.	Community blocks
NVIDIA TAO Image Classification Professional Enterprise Image classification model for general purpose use. Configurable backbones optimized for targets from MCU to GPU. Only supports RGB images. Pre-trained weights only support 224x224 resolution. Image width and height must be greater than 32. Training requires GPU.	Edge Impulse Inc.

The backbone is the foundational element of a neural network model, responsible for extracting meaningful features from input data, enabling subsequent parts to perform anomaly scoring or classification.

Description	Author
EfficientNet V2B0 Uses around 755K RAM based on your input size, and between 240-1675K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.35 Uses around 636K RAM based on your input size, and between 69-134K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.50 Uses around 643K RAM based on your input size, and between 78-168K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.75 Uses around 1263K RAM based on your input size, and between 109-266K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 1.0 Uses around 1273K RAM based on your input size, and between 126-372K ROM depending on the number of layers, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 0.1 Uses around 622K RAM based on your input size, and 58K ROM, with default compiler settings. Supports both RGB and grayscale.	Edge Impulse

The scoring function operates on the features extracted by the backbone, and returns anomaly scores for each segment of the image.

Description	Author	Recommended
MobileNetV1 96x96 0.25 A pre-trained multi-layer convolutional network designed to efficiently classify images. Uses around 105.9K RAM and 301.6K ROM with default settings and optimizations.	Edge Impulse
MobileNetV1 96x96 0.2 Uses around 83.1K RAM and 218.3K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV1 96x96 0.1 Uses around 53.2K RAM and 101K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 96x96 0.35 Uses around 296.8K RAM and 575.2K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 96x96 0.1 Uses around 270.2K RAM and 212.3K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 96x96 0.05 Uses around 265.3K RAM and 162.4K ROM with default settings and optimizations. Works best with 96x96 input size. Supports both RGB and grayscale.	Edge Impulse
MobileNetV2 160x160 1.0 Uses around 1.3M RAM and 2.6M ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
MobileNetV2 160x160 0.75 Uses around 1.3M RAM and 1.7M ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
MobileNetV2 160x160 0.5 Uses around 700.7K RAM and 982.4K ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
MobileNetV2 160x160 0.35 Uses around 683.3K RAM and 658.4K ROM with default settings and optimizations. Works best with 160x160 input size. Supports RGB only.	Edge Impulse
EfficientNet B0 Transfer learning model based on efficientnetb0_notop.h5 weights. This is a much larger model than MobileNet for Linux devices and accelerators.	Edge Impulse
NVIDIA TAO Image Classification Image classification model for general purpose use. Configurable backbones optimized for targets from MCU to GPU. Only supports RGB images. Pre-trained weights only support 224x224 resolution. Image width and height must be greater than 32. Training requires GPU.	Edge Impulse

These users will be notified over email when this job finishes.

	Name
	Naveen

Neural Network settings

Training settings

Augmentation settings

Advanced training settings

Neural network architecture

Model