| import tensorflow as tf | |
| policy = tf.keras.mixed_precision.Policy("mixed_float16") | |
| tf.keras.mixed_precision.set_global_policy(policy) | |
| from tensorflow import keras | |
| from tensorflow.keras import layers | |
| from keras_cv_attention_models import efficientnet | |
| input_shape = (image_frames, None, None, 3) | |
| image_frames = 60 | |
| image_size = 384 | |
| backbone_path = 'efficientnetv2-s-21k-ft1k.h5' | |
| backbone = efficientnet.EfficientNetV2S(pretrained=backbone_path,dropout=1e-6, num_classes=0, include_preprocessing = True) | |
| backbone.summary() | |
| backbone.trainable = False | |
| inputs = keras.Input(shape=input_shape) | |
| backbone_inputs = keras.Input(shape=(None, None, 3)) | |
| y = backbone(backbone_inputs) | |
| y = layers.Flatten()(y) | |
| y = layers.Dense(32, activation="relu")(y) | |
| y = layers.Dropout(0.1)(y) | |
| x = layers.TimeDistributed(keras.Model(backbone_inputs, y))(inputs) | |
| x = layers.Dropout(0.1)(x) | |
| x = layers.LSTM(128, return_sequences=True)(x) | |
| x = layers.Dropout(0.1)(x) | |
| x = layers.LSTM(128, return_sequences=False)(x) | |
| x = layers.Dropout(0.1)(x) | |
| x = layers.Dense(128, activation="relu")(x) | |
| x = layers.Dropout(0.1)(x) | |
| x = layers.Dense(64, activation="relu")(x) | |
| x = layers.Dropout(0.1)(x) | |
| x = layers.Dense(48, activation="relu")(x) | |
| x = layers.Dropout(0.1)(x) | |
| x = layers.Dense(32, activation="relu")(x) | |
| x = layers.Dropout(0.1)(x) | |
| outputs = layers.Dense(9, activation="relu")(x) | |
| model = keras.Model(inputs, outputs) | |
| model.compile( | |
| optimizer=keras.optimizers.Adam(1e-3), | |
| loss="mean_squared_error", | |
| metrics=["mean_squared_error", "mean_absolute_error"] | |
| ) |