search

lukemelas / EfficientNet-PyTorch

A PyTorch implementation of EfficientNet
Apache License 2.0
7.87k stars 1.53k forks source link

The speed of training is slow? #10

Open HITerStudy opened 5 years ago

HITerStudy commented 5 years ago

Sorry, I found a problem, that was, when I used the pretrained model 'efficientNet-B0' to fine-tune on my own dataset by gpu device, I found the speed of training was slow, thus, is there something wrong? Please give me an answer if having some solutions? Thanks a lot!

BowieHsu commented 5 years ago

maybe it's because the group convolution in pytorch is slow? i am new to pytorch, but i found when i train the mobilenetv3 using pytorch, it's slower than resnet50 and resnet34.

lukemelas commented 5 years ago

I've posted this on a couple of other issues, but there's an open PyTorch issue that may be relevant. It notes that grouped convs with FP32 are slow (and will be faster soon): pytorch/pytorch#18631

ChenXiao61 commented 5 years ago

@HITerStudy , Hello, I also want to fine-tune my own limited data using pretrained model 'efficientNet-B0', would you mind telling me how did you do that? Best regards, Xiao Chen

HITerStudy commented 5 years ago

@HITerStudy , Hello, I also want to fine-tune my own limited data using pretrained model 'efficientNet-B0', would you mind telling me how did you do that? Best regards, Xiao Chen

Of course, I also want to discuss with friends! And, I modified some layers of pretrained model 'efficientNet-B0' to adapt the needs of my own datasets, and other unchanged layers can be trained with the pretrained parameters. Thus, the model can be employed to fine-tune on other datasets.

ChenXiao61 commented 5 years ago

Well，how many layers did you modify and how about the results of fune-tuning the model?

发自我的iPhone

------------------ Original ------------------ From: HITerStudy notifications@github.com Date: Sat,Jun 15,2019 3:57 PM To: lukemelas/EfficientNet-PyTorch EfficientNet-PyTorch@noreply.github.com Cc: Chen Xiao mylovechenxiao@qq.com, Comment comment@noreply.github.com Subject: Re: [lukemelas/EfficientNet-PyTorch] The speed of training is slow? (#10)

@HITerStudy , Hello, I also want to fine-tune my own limited data using pretrained model 'efficientNet-B0', would you mind telling me how did you do that? Best regards, Xiao Chen

Of course, I also want to discuss with friends! And, I modified some layers of pretrained model 'efficientNet-B0' to adapt the needs of my own datasets, and other unchanged layers can be trained with the pretrained parameters. Thus, the model can be employed to fine-tune on other datasets.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

HITerStudy commented 5 years ago

Two layers, those are, first layer for conv as the channel of image is different, and the last layer for softmax. The performance was ordinary, and the results could be better if some optimization strategies are applied, maybe!

---Original--- From: "Chen Xiao"notifications@github.com Date: Fri, Jun 14, 2019 17:39 PM To: "lukemelas/EfficientNet-PyTorch"EfficientNet-PyTorch@noreply.github.com; Cc: "Mention"mention@noreply.github.com;"HITerStudy"827380444@qq.com; Subject: Re: [lukemelas/EfficientNet-PyTorch] The speed of training is slow? (#10)

@HITerStudy , Hello, I also want to fine-tune my own limited data using pretrained model 'efficientNet-B0', would you mind telling me how did you do that? Best regards, Xiao Chen

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.