This repo contain weights of MobileNetV1, MobileNetV2 models trained on 96x96 greyscale images of ImageNet dataset. suitable for a transfer learning model for image classification, object detection tasks.

These are Lightweight, energy efficient and memory efficient models that can be deployed on Edge devices such as Microcontrollers.

Transfer learning with MobileNetV1:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import (
    Dense,
    Dropout,
    Input
)

from tensorflow.keras.models import Model
from keras.applications.mobilenet import MobileNet
from tensorflow.keras.optimizers import Adam

epochs = 50
optimizer = Adam(learning_rate=0.0005)   # use any learning rate

input_tensor = Input(shape=(96, 96, 1))
mobilenet_model = MobileNet(
    input_shape=(96, 96, 1),
    input_tensor=input_tensor, 
    pooling="avg", 
    alpha=0.25,   # 0.25, 0.2, 0.1 
    weights="mobilenetV1_0.25_96x96_greyscale_weights.h5", # 0.25, 0.2, 0.1
    include_top=False
    )

mobilenet_model.trainable = False

mobilenet_output = mobilenet_model.output

# Dense layer
dense_layer = Dense(256, activation="relu")(mobilenet_output)

# Dropout layer
dropout_layer = Dropout(0.1)(dense_layer)

# classification layer
classification_layer = Dense(num_classes, activation='softmax')(dropout_layer)

model = Model(inputs=mobilenet_model.input, outputs=classification_layer)

print("Compiling model...")
model.compile(loss="categorical_crossentropy",
                optimizer=optimizer,
                metrics=["accuracy"])

model.summary()

Transfer learning with MobileNetV2:

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import (
    Dense,
    Dropout,
    Input
)

from tensorflow.keras.models import Model
from keras.applications.mobilenet_v2 import MobileNetV2
from tensorflow.keras.optimizers import Adam

epochs = 50
optimizer = Adam(learning_rate=0.0005)   # use any learning rate

input_tensor = Input(shape=(96, 96, 1))
mobilenet_model = MobileNetV2(
    input_shape=(96, 96, 1),
    input_tensor=input_tensor, 
    pooling="avg", 
    alpha=0.35,   # 0.35, 0.1, 0.05  
    weights="mobilenetV2_0.35_96x96_greyscale_weights.h5", # 0.35, 0.1, 0.05
    include_top=False
    )

mobilenet_model.trainable = False

mobilenet_output = mobilenet_model.output

# Dense layer
dense_layer = Dense(256, activation="relu")(mobilenet_output)

# Dropout layer
dropout_layer = Dropout(0.1)(dense_layer)

# classification layer
classification_layer = Dense(num_classes, activation='softmax')(dropout_layer)

model = Model(inputs=mobilenet_model.input, outputs=classification_layer)

print("Compiling model...")
model.compile(loss="categorical_crossentropy",
                optimizer=optimizer,
                metrics=["accuracy"])

model.summary()

If you want to Fine-tune the MobileNet models on a specific task rather than transfer learning then set trainable parameter to 'True'. This will retrain base model and finetune the weights for required task.

mobilenet_model.trainable = True  # unfreeze the base model