the dropout layer in train.prototxt was cancelled ,but i found overfitting when trying to test my caffemodel by using eval_detection.py:
however,the caffemodel provided in Thumos's README.md has beautiful performance. Does train.prototxt really nead dropout layer when you train Thumos? thank you for your reply
...
}
#layer {
name: "drop6"
type: "Dropout"
bottom: "fc6"
top: "fc6"
dropout_param {
dropout_ratio: 0.5
}
#}
...
the dropout layer in train.prototxt was cancelled ,but i found overfitting when trying to test my caffemodel by using eval_detection.py:
however,the caffemodel provided in Thumos's README.md has beautiful performance. Does train.prototxt really nead dropout layer when you train Thumos? thank you for your reply
...
}
#layer {
name: "drop6"
type: "Dropout"
bottom: "fc6"
top: "fc6"
dropout_param {
dropout_ratio: 0.5
}
#}
...