我正在训练有关视频帧序列的模型。我将视频存储为tfrecords文件（每个视频是一个单独的tfrecords文件），每个帧都编码为一个tf.train.Example。在训练期间，我读取了几批帧，并通过模型运行它们。但是，我看到来自不同视频的帧有可能以同一批结束。对于我正在训练的模型，这没有任何意义，实际上在损害性能。有没有办法确保一批帧只包含来自同一视频的帧？
答案 0 :(得分：0)
我最终使用了APP_PATH="${TARGET_BUILD_DIR}/${WRAPPER_NAME}"
# This script loops through the frameworks embedded in the application and
# removes unused architectures.
find "$APP_PATH" -name '*.framework' -type d | while read -r FRAMEWORK
do
FRAMEWORK_EXECUTABLE_NAME=$(defaults read "$FRAMEWORK/Info.plist" CFBundleExecutable)
FRAMEWORK_EXECUTABLE_PATH="$FRAMEWORK/$FRAMEWORK_EXECUTABLE_NAME"
echo "Executable is $FRAMEWORK_EXECUTABLE_PATH"
EXTRACTED_ARCHS=()
for ARCH in $ARCHS
do
echo "Extracting $ARCH from $FRAMEWORK_EXECUTABLE_NAME"
lipo -extract "$ARCH" "$FRAMEWORK_EXECUTABLE_PATH" -o "$FRAMEWORK_EXECUTABLE_PATH-$ARCH"
EXTRACTED_ARCHS+=("$FRAMEWORK_EXECUTABLE_PATH-$ARCH")
done
echo "Merging extracted architectures: ${ARCHS}"
lipo -o "$FRAMEWORK_EXECUTABLE_PATH-merged" -create "${EXTRACTED_ARCHS[@]}"
rm "${EXTRACTED_ARCHS[@]}"
echo "Replacing original executable with thinned version"
rm "$FRAMEWORK_EXECUTABLE_PATH"
mv "$FRAMEWORK_EXECUTABLE_PATH-merged" "$FRAMEWORK_EXECUTABLE_PATH"
done
方法。谓词比较第一帧和最后一帧的视频ID。如果它们不同，则将整个批次删除：
tf.data.Dataset.filter