… UPSCALE1 to prune models with any pruning pattern. By removing constraints from existing
pruners, we improve ImageNet accuracy for post-training pruned … channels, UPSCALE …

… network, using Unconstrained Blueprint Separable Convolutions, in … pruning and learning
rate, we noted that halving the learning rate at each step of Relevant Channel Pruning …

… pruning scheme, we can choose channel pruning [60], pattern-based pruning [45], or block-based
pruning … 14] could take several seconds for them to upscale only one image on a large …

… In summary, the unconstrained channel pruning offers a very promising solution for the
soybean recognition task by balancing the model size and the inference accuracy. With the …

… by integrating unstructured pruning with channel pruning, we … , termed channel-adaptive
magnitude-based pruning (CAMP). … an unconstrained channel pruning, mitigating inter-channel …

… [50] proposed UPSCALE, which addresses the redundancy issue by reordering channels and
… However, it requires the careful design of a corresponding unconstrained channel pruning …

… In a recent study [26], unconstrained masking generally outperforms post-training accuracy
of pruned models by an average of 2.1% on ImageNet-1K. Compared to unconstrained head …

… Our method is also related to the unconstrained channel pruning for CNNs [39]. However,
our method explores this idea from the perspective of breaking structural dependence and …

… Instead of chopping down the spatially bigger feature map, we upscale the … of channels
to prune. Once the algorithm prunes less than 5 channels of the network per step, the pruning …

… As a general trend, we observe that higher upscale factor does not affect the inference time
significantly. This is due to the fact that the upscale factor only affects the number of channels …