Does LMDB data layer automatically do mean subtraction?

[zic...@ualberta.ca]

Hi,

I'm training a conv net on grayscale images, which were simply copied to 3 channels and in range [0,255.] before saving to lmdb.

I verified that the image looks normal when reading from lmdb with the following code:

lmdb_file = "/usr/work/data/train_lmdb" 

lmdb_env = lmdb.open(lmdb_file, readonly=True)

lmdb_txn = lmdb_env.begin()

lmdb_cursor = lmdb_txn.cursor()

datum = caffe.proto.caffe_pb2.Datum()

  

i = 0

print "Starting reading"

for key, value in lmdb_cursor:

  if i < 1:

    print 'iteration: ', i

    datum.ParseFromString(value)

    label = datum.label

    flat_x = np.fromstring(datum.data, dtype=np.float32)

    x = flat_x.reshape(datum.channels, datum.height, datum.width)

    print 'label: ', label

    i += 1

  else:

    break

lmdb_env.close()
plt.imshow(x.transpose(1,2,0)/255, cmap=plt.cm.gray)
plt.show()

Now the training does not seem to work from the beginning. I checked the data to make sure it's correct. But the image looks strange to me. 

I wonder if caffe does mean subtraction even if it's not defined in the net model. Below is the model of the data layer. And I was just doing this to visualize:

plt.imshow(solver.test_nets[0].blobs['data'].data[0,...].transpose(1,2,0)/255, cmap=plt.cm.gray)
plt.show()
plt.imshow(solver.net.blobs['data'].data[0,...].transpose(1,2,0)/255, cmap=plt.cm.gray)
plt.show()

layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  data_param {
    source: "/usr/work/data/train_lmdb"
    backend: LMDB
    batch_size: 50
  }
  include {
    phase: TRAIN
  }
}
layer {
  name: "data"
  type: "Data"
  top: "data"
  top: "label"
  data_param {
    source: "/usr/work/data/test_lmdb"
    backend: LMDB
    batch_size: 50
  }
  include {
    phase: TEST
  }
}

Does anyone spot any possible issue?

Thanks,

Vincent  

[zic...@ualberta.ca]

I'm thinking maybe the way I created lmdb wasn't what caffe was expecting?

Below is a simplified version of how I saved the data to lmdb:

# Resized image
X = np.zeros((100, 3, RESIZE_HEIGHT, RESIZE_WIDTH), dtype=np.float32)

# this is the label

y = np.array(img_labels, dtype=np.int32)
map_size = X.nbytes * 50
# Read the images into X
X = ...

...

# write to lmdb

      env = lmdb.open(test_lmdb, map_size=map_size)

      with env.begin(write=True) as txn:

        for i in xrange(100):

          print "Image id: ", i

          datum = caffe.proto.caffe_pb2.Datum()

          datum.channels = X.shape[1]

          datum.height = X.shape[2]

          datum.width = X.shape[3]

          datum.float_data = X[i].tobytes()

          datum.label = int(y[i])

          str_id = '{:08}'.format(i)

          txn.put(str_id.encode('ascii'), datum.SerializeToString())

[zic...@ualberta.ca]

Sorry there was a typo:

datum.float_data = X[i].tobytes()

should be:


datum.data = X[i].tobytes()

I've also corrected it inline below.

I wanted to save the image(type:float32) as float number into lmdb. I was suspecting that there's something wrong with using datum.data instead of datum.float_data, so I was just doing some testing to see. Not sure yet.

On Thursday, April 28, 2016 at 2:15:47 PM UTC-6, zic...@ualberta.ca wrote:

I'm thinking maybe the way I created lmdb wasn't what caffe was expecting?

Below is a simplified version of how I saved the data to lmdb:

# Resized image
X = np.zeros((100, 3, RESIZE_HEIGHT, RESIZE_WIDTH), dtype=np.float32)

# this is the label

y = np.array(img_labels, dtype=np.int32)
map_size = X.nbytes * 50
# Read the images into X
X = ...

...

# write to lmdb

      env = lmdb.open(test_lmdb, map_size=map_size)

      with env.begin(write=True) as txn:

        for i in xrange(100):

          print "Image id: ", i

          datum = caffe.proto.caffe_pb2.Datum()

          datum.channels = X.shape[1]

          datum.height = X.shape[2]

          datum.width = X.shape[3]

          datum.data = X[i].tobytes()

[zic...@ualberta.ca]

Just figured it out. In case anyone comes across similar issues, here's the solution:

The issue was that when I saved the float number image data into lmdb, I incorrectly converted them into bytes and put it into datum.data instead of float_data. When caffe reads from it, I think it interprets it as integer so the input data to the net is totally messed up. 

The correct way is to use the filed float_data in Datum message:

datum.float_data.extend(X[i]) when saving image X[i] into datum.