Ultrafast electron migration in molecules is the progenitor of all chemical reactions and biological functions after light-matter interaction [1–4]. Following this ultrafast dynamics, however, has been an enduring endeavor [5, 6]. Recently, it has been shown that high-harmonic spectroscopy (HHS) is able to probe dynamics with attosecond temporal and sub-angstrom spatial resolution [7–10]. Still, real-time visualization of single-molecule dynamics continues to be a great challenge because experimental harmonic spectra are due to the coherent averages of light emission from individual molecules of different alignments. Here, we show that from high harmonics generated with single-color and two-color probe lasers in a pump-probe experiment, the complex amplitude and phase of harmonics from a single fixed-in-space molecule can be reconstructed using modern machine learning (ML) algorithm. From the complex single-molecule dipoles for different harmonics, we construct a series of film clips of hole density distributions of the cation at time steps of 50 attoseconds (1 as=10^{−18} s) to make a classical “movie” of electron migration after tunnel ionization of the molecule. Moreover, the angular dependence of molecular charge migration is fully resolved. By examining these clips, we observed that holes do not just “migrate” along the laser direction, but they may “swirl” around the atom centers. The ML-based HHS proposed here establishes a general reconstruction scheme for studying ultrafast charge migration in molecules, paving a way for further advance in tracing and controlling photochemical reactions by femtosecond lasers.