* Add fused conv support * Set all saved_model values to non trainable * Fix TFLite fp16 model export * Fix int8 TFLite conversion