Understanding how bulges grow in galaxies is a critical step towards unveiling the link between galaxy morphology and star-formation. To do so, it is necessary to decompose large sample of galaxies at different epochs into their main components (bulges and discs). This is particularly challenging, especially at high redshifts, where galaxies are poorly resolved. This work presents a catalog of bulge-disc decompositions of the surface brightness profiles of ∼17.600 H-band-selected galaxies in the CANDELS fields (F160W<23, 0<z<2) in 4 to 7 filters covering a spectral range of 430-1600 nm. This is the largest available catalog of this kind up to z=2. By using a novel approach based on deep learning to select the best model to fit, we manage to control systematics arising from wrong model selection and obtain less-contaminated samples than previous works. We show that the derived strucaroctural properties are within ∼10--20% of random uncertainties. We then fit stellar population models to the decomposed spectral energy distributions of bulges and discs and derive stellar masses (and stellar mass bulge-to-total ratios) as well as rest-frame colors (U,V,J) for bulges and discs separately. All data products are publicly released with this paper and through the web page https://lerma.obspm.fr/huertas/form_CANDELS and will be used for scientific analysis in forthcoming works.