search

shuangshuangguo / caffe2pytorch-tsn

Transform the caffe model to pytorch model for Temporal Segment Network
73 stars 16 forks source link

MaxPooling with floor mode not consistent with original implementation #1

Open scenarios opened 6 years ago

scenarios commented 6 years ago

Hi, I've noticed you are using the maxpooling without ceil mode in 'InceptionModule2', which is not consistent with the caffe implementation. But you can still reproduce the results with the pretrained caffemodel? BTW, I made the tsn caffemodel available in tensorflow, however, with every layer keeping identical with the TSN caffe implementation, there's still a 10% drop.

shuangshuangguo commented 6 years ago

@scenarios My work just transfers caffe-model to pytorch-model by reading layer parameters. When I get parameters, I use tsn-pytorch code given by authors to test. How does max pooling perform doesn't matter result. About the second question, I think maybe you should check your input, such as all of transformation.