Caffe: wrong output from net.blobs['fc6'].data Caffenet.

[midnitekoder]

I want go generate the output of 'fc6' layer of Caffenet for my image and use it as feature vector. But, when I compare the output of two different images, I get exactly the same output. Please, tell me what is the issue. I have attached for your reference.
I am savingnet.blobs['fc6'].data[0]

import numpy as np
import matplotlib.pyplot as plt
import time
from array import array

import sys

caffe_root = '/home/iplab_apu/Desktop/caffe-master/'

sys.path.insert(0,caffe_root+'python')
import caffe


caffe.set_mode_cpu()

model_def = caffe_root + 'models/bvlc_reference_caffenet/deploy.prototxt'
model_weights = caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'
net = caffe.Net(model_def,      # defines the structure of the model
                model_weights,  # contains the trained weights
                caffe.TEST)     # use test mode (e.g., don't perform dropout)
                                   # load the mean ImageNet image (as distributed with Caffe) for subtraction
mu = np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy')
print mu
mu = mu.mean(1).mean(1)  # average over pixels to obtain the mean (BGR) pixel values
print 'mean-subtracted values:', zip('BGR', mu)
net.blobs['data'].reshape(1,        # batch size
                          3,         # 3-channel (BGR) images
                          227, 227)  # image size is 227x227

                         # create transformer for the input called 'data'

transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})

transformer.set_transpose('data', (2,0,1))  # move image channels to outermost dimension
transformer.set_mean('data', mu)            # subtract the dataset-mean value in each channel
transformer.set_raw_scale('data', 255)      # rescale from [0, 1] to [0, 255]
transformer.set_channel_swap('data', (2,1,0))  # swap channels from RGB to BGR
# set the size of the input (we can skip this if we're happy
#  with the default; we can also change it later, e.g., for different batch sizes)



image = caffe.io.load_image('/home/iplab_apu/Desktop/ankit/drive-download-20160823T093738Z/ISBI2016_ISIC_Part1_Training_Data/ISIC_0000001.jpg')

image = caffe.io.resize(image, (227,227))
transformed_image=[]
transformed_image = transformer.preprocess('data', image)


net.blobs['data'].data[...] = transformed_image
t=time.clock()
output = net.forward()
t2=time.clock()
output_prob = output['prob'][0]

np.save('feature_vec_arr_fc6_1.npy',net.blobs['fc6'].data[0])

[abhimanyudubey]

When you call
output = net.forward()
it replaces _all_ layers by the input, hence your data gets overwritten, basically making your call to
net.blobs['data'].data[...] = transformed_image
redundant.

You should replace those lines with:
output = net.forward(data=np.asarray(transformed_image))

This should work. Please refer to the caffe-users group for further queries.

[midnitekoder]

It gave the following error:

Traceback (most recent call last):
  File "compare.py", line 52, in <module>
    output = net.forward(data=np.asarray(transformed_image))
  File "/home/iplab_apu/Desktop/caffe-master/python/caffe/pycaffe.py", line 118, in _Net_forward
    raise Exception('Input is not batch sized')
Exception: Input is not batch sized

[abhimanyudubey]

Please set the batch size to 1 in deploy.prototxt.

[midnitekoder]

Still the same error

[abhimanyudubey]

Apologies, replace the line mentioned earlier with:
output = net.forward(data=np.asarray([transformed_image]))

[midnitekoder]

Thats okay. It runs but generates the same output of 'fc6' layer for two different images.

[abhimanyudubey]

Replace your line:
np.save('feature_vec_arr_fc6_1.npy',net.blobs['fc6'].data[0])
with:
np.save('feature_vec_arr_fc6_1.npy',output['fc6'].data[0])

And the call to forward() with forward_all(). For further queries, please use the Google group.

[midnitekoder]

Yes, I would use Google group but just to update, output does not have 'fc6' key as it only has 'prob' key in its dictionary.

[abhimanyudubey]

Check this: https://github.com/AlfredXiangWu/python_misc/blob/master/caffe/caffe_ftr.py

[midnitekoder]

Thanks :) I'll go through the code.

[midnitekoder]

But, why is it thatnet.blobs['fc6'].data[0] gives the same value for two different images?
I think it is relevant to this forum because data blobs are supposed to give the output of layer which is supposed to be different for different images.

[shelhamer]

Please ask usage questions on the mailing list.

From https://github.com/BVLC/caffe/blob/master/CONTRIBUTING.md:

_Please do not post usage, installation, or modeling questions, or other requests for help to Issues._
Use the caffe-users list instead. This helps developers maintain a clear, uncluttered, and efficient view of the state of Caffe.