Extracting features from trained model

[vj-1988]

@anvaysrivastava : I found script to extract features but any instructions on how to convert train.prototxt to deploy.prototxt would be appreciated.

[anvaysrivastava]

Hey, Sorry but I had not contributed in this project for a long time and the question is a out of context for me.

[vj-1988]

@ananya-h-a : can you kindly share me the vgg16_triplet_shallow_deploy.prototxt that is mentioned in extract features.py file?

[ClareQi]

@vj-1988 I have the same problem as you, have you solved it?

[sdy1106]

@ClareQi @vj-1988 Do you have solved this issue? If solved, I'll be appreciate for sharing the the file for me.

[vj-1988]

@sdy1106 & @ClareQi : No. I stopped working with this repo.

[markeloff]

@sdy1106 hi，哥们，你要的deploy.prototxt name: "Visnet" layer { name: "data_q" type: "Input" top: "data_q" input_param { shape: { dim: 1 dim: 3 dim: 224 dim: 224 } } } layer { name: "data_q2" type: "Pooling" bottom: "data_q" top: "data_q2" pooling_param { pool: MAX kernel_size: 4 stride: 4 } } layer { name: "data_q3" type: "Pooling" bottom: "data_q" top: "data_q3" pooling_param { pool: MAX kernel_size: 8 stride: 8 } }

#Q layer { bottom: "data_q" top: "conv1_1_q" name: "conv1_1" type: "Convolution" convolution_param { num_output: 64 pad: 1 kernel_size: 3 } param { name: "conv1_1_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv1_1_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv1_1_q" top: "conv1_1_q" name: "relu1_1" type: "ReLU" } layer { bottom: "conv1_1_q" top: "conv1_2_q" name: "conv1_2" type: "Convolution" convolution_param { num_output: 64 pad: 1 kernel_size: 3 } param { name: "conv1_2_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv1_2_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv1_2_q" top: "conv1_2_q" name: "relu1_2" type: "ReLU" } layer { bottom: "conv1_2_q" top: "pool1_q" name: "pool1" type: "Pooling" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { bottom: "pool1_q" top: "conv2_1_q" name: "conv2_1" type: "Convolution" convolution_param { num_output: 128 pad: 1 kernel_size: 3 } param { name: "conv2_1_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv2_1_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv2_1_q" top: "conv2_1_q" name: "relu2_1" type: "ReLU" } layer { bottom: "conv2_1_q" top: "conv2_2_q" name: "conv2_2" type: "Convolution" convolution_param { num_output: 128 pad: 1 kernel_size: 3 } param { name: "conv2_2_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv2_2_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv2_2_q" top: "conv2_2_q" name: "relu2_2" type: "ReLU" } layer { bottom: "conv2_2_q" top: "pool2_q" name: "pool2" type: "Pooling" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { bottom: "pool2_q" top: "conv3_1_q" name: "conv3_1" type: "Convolution" convolution_param { num_output: 256 pad: 1 kernel_size: 3 } param { name: "conv3_1_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv3_1_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv3_1_q" top: "conv3_1_q" name: "relu3_1" type: "ReLU" } layer { bottom: "conv3_1_q" top: "conv3_2_q" name: "conv3_2" type: "Convolution" convolution_param { num_output: 256 pad: 1 kernel_size: 3 } param { name: "conv3_2_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv3_2_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv3_2_q" top: "conv3_2_q" name: "relu3_2" type: "ReLU" } layer { bottom: "conv3_2_q" top: "conv3_3_q" name: "conv3_3" type: "Convolution" convolution_param { num_output: 256 pad: 1 kernel_size: 3 } param { name: "conv3_3_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv3_3_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv3_3_q" top: "conv3_3_q" name: "relu3_3" type: "ReLU" } layer { bottom: "conv3_3_q" top: "pool3_q" name: "pool3" type: "Pooling" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { bottom: "pool3_q" top: "conv4_1_q" name: "conv4_1" type: "Convolution" convolution_param { num_output: 512 pad: 1 kernel_size: 3 } param { name: "conv4_1_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv4_1_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv4_1_q" top: "conv4_1_q" name: "relu4_1" type: "ReLU" } layer { bottom: "conv4_1_q" top: "conv4_2_q" name: "conv4_2" type: "Convolution" convolution_param { num_output: 512 pad: 1 kernel_size: 3 } param { name: "conv4_2_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv4_2_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv4_2_q" top: "conv4_2_q" name: "relu4_2" type: "ReLU" } layer { bottom: "conv4_2_q" top: "conv4_3_q" name: "conv4_3" type: "Convolution" convolution_param { num_output: 512 pad: 1 kernel_size: 3 } param { name: "conv4_3_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv4_3_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv4_3_q" top: "conv4_3_q" name: "relu4_3" type: "ReLU" } layer { bottom: "conv4_3_q" top: "pool4_q" name: "pool4" type: "Pooling" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { bottom: "pool4_q" top: "conv5_1_q" name: "conv5_1" type: "Convolution" convolution_param { num_output: 512 pad: 1 kernel_size: 3 } param { name: "conv5_1_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv5_1_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv5_1_q" top: "conv5_1_q" name: "relu5_1" type: "ReLU" } layer { bottom: "conv5_1_q" top: "conv5_2_q" name: "conv5_2" type: "Convolution" convolution_param { num_output: 512 pad: 1 kernel_size: 3 } param { name: "conv5_2_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv5_2_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv5_2_q" top: "conv5_2_q" name: "relu5_2" type: "ReLU" } layer { bottom: "conv5_2_q" top: "conv5_3_q" name: "conv5_3" type: "Convolution" convolution_param { num_output: 512 pad: 1 kernel_size: 3 } param { name: "conv5_3_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv5_3_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "conv5_3_q" top: "conv5_3_q" name: "relu5_3" type: "ReLU" } layer { bottom: "conv5_3_q" top: "pool5_q" name: "pool5" type: "Pooling" pooling_param { pool: MAX kernel_size: 2 stride: 2 } } layer { bottom: "pool5_q" top: "fc6_q" name: "fc6" type: "InnerProduct" inner_product_param { num_output: 4096 } param { name: "fc6_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "fc6_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "fc6_q" top: "fc6_q" name: "relu6" type: "ReLU" } layer { bottom: "fc6_q" top: "fc6_q" name: "drop6" type: "Dropout" dropout_param { dropout_ratio: 0.4 } } layer { bottom: "fc6_q" top: "fc7_q" name: "fc7" type: "InnerProduct" inner_product_param { num_output: 4096 } param { name: "fc7_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "fc7_b" lr_mult: 2.0 decay_mult: 0.0 } } layer { bottom: "fc7_q" top: "fc7_q" name: "relu7" type: "ReLU" } layer { bottom: "fc7_q" top: "fc7_q" name: "drop7" type: "Dropout" dropout_param { dropout_ratio: 0.4 } } layer { bottom: "fc7_q" top: "fc8_q" name: "fc8_q" type: "InnerProduct" param { name: "fc8_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "fc8_b" lr_mult: 2.0 decay_mult: 0.0 } inner_product_param { num_output: 4096 } }

layer { name: "reshape_q" type: "Reshape" bottom: "fc8_q" top: "fc8_q_r" reshape_param { shape { dim: 0 dim: 0 dim: 1 dim: 1 } } } layer { name: "fc8_q_norm" type: "LRN" bottom: "fc8_q_r" top: "fc8_q_norm" lrn_param { local_size: 8191 alpha: 8191 beta: 0.5 } }

#Q shallow

layer { name: "conv1_q2" type: "Convolution" bottom: "data_q2" top: "conv1_q2" param { name: "conv1_shallow_w2" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv1_shallow_b2" lr_mult: 2.0 decay_mult: 0.0 } convolution_param { num_output: 96 kernel_size: 8 stride: 4 pad_h: 4 pad_w: 4 } } layer { name: "relu1_q2" type: "ReLU" bottom: "conv1_q2" top: "conv1_q2" }

layer { name: "norm1_q2" type: "LRN" bottom: "conv1_q2" top: "norm1_q2" lrn_param { local_size: 5 alpha: 9.99999974738e-05 beta: 0.75 } } layer { name: "pool1_q2" type: "Pooling" bottom: "norm1_q2" top: "pool1_q2" pooling_param { pool: MAX kernel_size: 7 stride: 4 pad_h: 3 pad_w: 3 } }

layer { name: "conv1_q3" type: "Convolution" bottom: "data_q3" top: "conv1_q3" param { name: "conv1_shallow_w3" lr_mult: 1.0 decay_mult: 1.0 } param { name: "conv1_shallow_b3" lr_mult: 2.0 decay_mult: 0.0 } convolution_param { num_output: 96 kernel_size: 8 stride: 4 pad_h: 4 pad_w: 4 } } layer { name: "relu1_q3" type: "ReLU" bottom: "conv1_q3" top: "conv1_q3" }

layer { name: "norm1_q3" type: "LRN" bottom: "conv1_q3" top: "norm1_q3" lrn_param { local_size: 5 alpha: 9.99999974738e-05 beta: 0.75 } } layer { name: "pool1_q3" type: "Pooling" bottom: "norm1_q3" top: "pool1_q3" pooling_param { pool: MAX kernel_size: 3 stride: 2 pad_h: 1 pad_w: 1 } }

layer { name: "pool1_q2_flat" type: "Flatten" bottom: "pool1_q2" top: "pool1_q2_flat" } layer { name: "pool1_q3_flat" type: "Flatten" bottom: "pool1_q3" top: "pool1_q3_flat" } layer { name: "shallow_q" type: "Concat" bottom: "pool1_q2_flat" bottom: "pool1_q3_flat" top: "shallow_q" } layer { name: "reshape_shallow_q" type: "Reshape" bottom: "shallow_q" top: "shallow_q_r" reshape_param { shape { dim: 0 dim: 0 dim: 1 dim: 1 } } } layer { name: "shallow_q_norm" type: "LRN" bottom: "shallow_q_r" top: "shallow_q_norm" lrn_param { local_size: 8191 alpha: 8191.0 beta: 0.5 } }

layer { name: "q_concat" type: "Concat" bottom: "shallow_q_norm" bottom: "fc8_q_norm" top: "q_concat" } layer { name: "linear_embedding_q" type: "InnerProduct" bottom: "q_concat" top: "linear_embedding_q" param { name: "linear_embedding_w" lr_mult: 1.0 decay_mult: 1.0 } param { name: "linear_embedding_b" lr_mult: 2.0 decay_mult: 0.0 } inner_product_param { num_output: 4096 } }

layer { name: "reshape_leq" type: "Reshape" bottom: "linear_embedding_q" top:"linear_embedding_q_re" reshape_param { shape { dim: 0 dim: 0 dim: 1 dim: 1 } } }

layer { name: "linear_embedding_q_norm" type: "LRN" bottom: "linear_embedding_q_re" top: "linear_embedding_q_norm" lrn_param { local_size: 8191 alpha: 8191.0 beta: 0.5 } }

[ClareQi]

@markeloff 你好，后来我自己写出来了，不过还是谢谢你！我的和你完全一样，但是在最后多了一层： layer { name: "prob" type: "Softmax" bottom: "linear_embedding_q_norm" top: "prob" } 不知道这个需要不需要。

[markeloff]

@ClareQi 你好，看你需求，是想得到分类结果还是获取特征就行了，如果只是特征的话softmax不需要的

[ClareQi]

@markeloff 噢噢，好的，谢谢。另外想与您交流下，您是怎么做数据集的？还是用源码中提供的？我目前做的数据集是在淘宝上爬取一些商品橱窗图，然后人工选择QP,再用脚本生成同一类别下的不同产品作为N。

[markeloff]

@ClareQi qq聊 814048640

[aiaini66]

@markeloff 你好，这个工程是不是本来就不完整的，整么都没有找到caffe文件夹，readme里说的改了caffe里的两个cpp文件，怎么连caffe编译的文件都没有？望你能给我解答下疑惑。感激不尽！

[aiaini66]

@ClareQi @markeloff 你好，这个工程是不是本来就不完整的，整么都没有找到caffe文件夹，readme里说的改了caffe里的两个cpp文件，怎么连caffe编译的文件都没有？望你能给我解答下疑惑。感激不尽！