This paper reports the results and post-challenge analyses of ChaLearn's\nAutoDL challenge series, which helped sorting out a profusion of AutoML\nsolutions for Deep Learning (DL) that had been introduced in a variety of\nsettings, but lacked fair comparisons. All input data modalities (time series,\nimages, videos, text, tabular) were formatted as tensors and all tasks were\nmulti-label classification problems. Code submissions were executed on hidden\ntasks, with limited time and computational resources, pushing solutions that\nget results quickly. In this setting, DL methods dominated, though popular\nNeural Architecture Search (NAS) was impractical. Solutions relied on\nfine-tuned pre-trained networks, with architectures matching data modality.\nPost-challenge tests did not reveal improvements beyond the imposed time limit.\nWhile no component is particularly original or novel, a high level modular\norganization emerged featuring a "meta-learner", "data ingestor", "model\nselector", "model/learner", and "evaluator". This modularity enabled ablation\nstudies, which revealed the importance of (off-platform) meta-learning,\nensembling, and efficient data management. Experiments on heterogeneous module\ncombinations further confirm the (local) optimality of the winning solutions.\nOur challenge legacy includes an ever-lasting benchmark\n(http://autodl.chalearn.org), the open-sourced code of the winners, and a free\n"AutoDL self-service".\n