深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma

本文主要实验文献文献《Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding》算法,在tiny-yolo coco上的压缩效果,在darknet基础上,编写该算法进行压缩实验,结果如下:

原始模型大小64M:mAP=0.224 

深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma_第1张图片


训练500次,模型大小54M:mAP=0.203

深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma_第2张图片

训练5000次,模型大小49M:mAP=0.214


深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma_第3张图片

训练50000次,模型大小39M:mAP=0.221

深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma_第4张图片

训练100000次,模型大小30M:mAP=0.231

深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma_第5张图片

训练180000次,模型大小23M:mAP=0.232

深度学习(六十九)darknet 实现实验 Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffma_第6张图片

懒得训练了,目前每迭代1500次,下降0.1M,要训练到底,应该是可以压缩到18M左右

你可能感兴趣的:(深度学习,深度学习)