Training in Feed Forward Deep Neural Networks is a memory-intensive operation which is usually performed on GPUs with limited memory capacities. This may force data scientists to limit the depth of the models or the resolution of the input data if data does not fit in the GPU memory. The re-materialization technique, whose idea comes from the checkpointing strategies developed in the Automatic Differentiation literature, allows data scientists to limit the memory requirements related to the storage of intermediate data (activations), at the cost of an increase in the computational cost. This paper introduces a new strategy of re-materialization of activations that significantly reduces memory usage. It consists in selecting which activations are saved and which activations are deleted during the forward phase and then recomputing the deleted activations when they are needed during the backward phase. We propose an original computation model that combines two types of activation savings: either only storing the layer inputs or recording the complete history of operations that produced the outputs. This paper focuses on the fully heterogeneous case, where the computation time and the memory requirement of each layer is different. We prove that finding the optimal solution is NP-hard and that classical techniques from Automatic Differentiation literature do not apply. Moreover, the classical assumption of memory persistence of materialized activations, used to simplify the search of optimal solutions, does not hold anymore. Thus, we propose a weak memory persistence property and provide a dynamic program to compute the optimal sequence of computations. This algorithm is made available through the Rotor software, a PyTorch plug-in dealing with any network consisting of a sequence of layers, each of them having an arbitrarily complex structure. Through extensive experiments, we show that our implementation consistently outperforms existing re-materialization approaches for a large class of networks, image sizes, and batch sizes.

Modern Deep Neural Networks (DNNs) require significant memory to store weight, activations, and other intermediate tensors during training. Hence, many models don’t fit one GPU device or can be trained using only a small per-GPU batch size. This survey provides a systematic overview of the approaches that enable more efficient DNNs training. We analyze techniques that save memory and make good use of computation and communication resources on architectures with a single or several GPUs. We summarize the main categories of strategies and compare strategies within and across categories. Along with approaches proposed in the literature, we discuss available implementations.

Rematerialization and offloading are two well known strategies to save memory during the training phase of deep neural networks, allowing data scientists to consider larger models, batch sizes or higher resolution data. Rematerialization trades memory for computation time, whereas Offloading trades memory for data movements. As these two resources are independent, it is appealing to consider the simultaneous combination of both strategies to save even more memory. We precisely model the costs and constraints corresponding to Deep Learning frameworks such as PyTorch or Tensorflow, we propose optimal algorithms to find a valid sequence of memory-constrained operations and finally, we evaluate the performance of proposed algorithms on realistic networks and computation platforms. Our experiments show that the possibility to offload can remove one third of the overhead of rematerialization, and that together they can reduce the memory used for activations by a factor 4 to 6, with an overhead below 20%.