Résumé : Several works in the past decade have used the ratio between total (rest 8-1000$\mu$m) infrared and radio (rest 1.4 GHz) luminosity in star-forming galaxies (q$_{TIR}$), often referred to as the "infrared-radio correlation" (IRRC), to calibrate radio emission as a star formation rate (SFR) indicator. Previous studies constrained the evolution of q$_{TIR}$ with redshift, finding a mild but significant decline, that is yet to be understood. For the first time, we re-calibrate q$_{TIR}$ as a function of both stellar mass (M$_{*}$) and redshift, starting from an M$_{*}$-selected sample of >400,000 star-forming galaxies in the COSMOS field, identified via (NUV-r)/(r-J) colours, at redshifts 0.1<z<4.5. Within each (M$_{*}$,z) bin, we stack the deepest available infrared/sub-mm and radio images. We fit the stacked IR spectral energy distributions with typical star-forming galaxy and IR-AGN templates, and carefully remove radio AGN candidates via a recursive approach. We find that the IRRC evolves primarily with M$_{*}$, with more massive galaxies displaying systematically lower q$_{TIR}$. A secondary, weaker dependence on redshift is also observed. The best-fit analytical expression is the following: q$_{TIR}$(M$_{*}$,z)=(2.646$\pm$0.024)$\times$(1+z)$^{(-0.023\pm0.008)}$-(0.148$\pm$0.013)$\times$($\log~M_{*}$/M$_{\odot}$-10). Adding the UV dust-uncorrected contribution to the IR as a proxy for the total SFR, would further steepen the q$_{TIR}$ dependence on M$_{*}$. The lower IR/radio ratio in more massive galaxies could be possibly linked to higher SFR surface density, which induces larger cosmic-ray scale heights. Our findings highlight that using radio emission as a proxy for SFR requires novel M$_{*}$-dependent recipes, that will enable us to convert detections from future ultra deep radio surveys into accurate SFR measurements down to low-SFR, low-M$_{*}$ galaxies.