nigroup/nideep

collection of utilities to use with deep learning libraries (e.g. caffe)

nideep's Issues

doubt in net surgery

I am trying to train pascal context 2010 dataset for segmentation using FCNN.I have two questions.

1. My doubt is Initially when i perform net surgery on the VGG16.prototxt to get VGG16_Full_conv.prototxt,the input data is in the form of images as below,

name: "VGG_ILSVRC_16_layers_FCNN"
input: "data"
input_dim: 10
input_dim: 3
input_dim: 224
input_dim: 224
....
}

After net surgery when I copy the VGG16_FCNN layers to train_val.prototxt which is the prorotxt file to train the FCNN 32 strides layers,the network is skipping the input layer values.It is skipping because train_val.prototxt input datatype is in lmdb format as below,

name: "FCN"
layer {
name: "data"
type: "Data"
top: "data"
include {
phase: TRAIN
}
transform_param {
mean_value: 104.00699
mean_value: 116.66877
mean_value: 122.67892
}
data_param {
source: "VOCdevkit/VOC2010/context_imgs_train_lmdb"
batch_size: 1
backend: LMDB
}
}

So should I convert input type "LMDB" into type "data" and give input as images like the VGG16.protoxt and VGG16_Fully_conv.prototxt.

In short,Should not the weights between the input layer and the next layer be also included?

2).When I perform net surgery should I also copy the fc8 layer weights,because VGG 16 outputs 1000 classes whereas FCNN is for 60 classes?

At present I am not including input layer and fc8 layer weights and the initial loss is 812751.In google caffe group you have posted your loss around 767455?I am using the same train/val slpit as you posted in the google groups.

Apologies for the long post

fileSystemUtils used in pascal_context_to_lmdb.py but not provided

The file pascal_context_to_lmdb.py contains statement import fileSystemUtils as fs, but it is not provided in the repository not is it publicly available elsewhere.

Include tensorflow dependent modules in Travis CI testing

import tensorflow in travis CI currently causes segmentation faults. Not sure where this is coming from. A temporary solution was to exclude these modules from CI.

TwoEars NumSrcs incomplete

class introduced in #69

problem with data preparation

I want to prepare the data & label layer for training pixel labeling network, for example using pascal context. I found nideep generate lmdbs for images and labels separately and all images and labels are in original size. However it's hard to use these generated lmdbs in caffe's prototxt because it requires equal size input (224*224) for both images and labels. My question is whether it's possible to put images and labels into one lmdb, how to do it with nideep? I think put them into one lmdb could let me use crop operation in caffe's data layer. Thanks!