[
https://issues.apache.org/jira/browse/SYSTEMML-1583?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16000968#comment-16000968
]
Frederick Reiss commented on SYSTEMML-1583:
-------------------------------------------
I presume that the conversion from a .caffemodel file to our binary block
format is bottlenecked on the I/O system, is it not? Back-of-the-envelope, even
large models of hundreds of megabytes should take less than a second to convert.
I recommend skipping the proprietary binary coefficient format and disk layout.
If you need to generate the files in the background as a short-term expedient,
that's ok for now, but at least don't force the user to create and manage them.
And let's use the actual Caffe model zoo at
[https://github.com/BVLC/caffe/wiki/Model-Zoo], not create an out-of-date
partial copy on Github.
> Implement converter in Python to convert caffemodel in SystemML format
> ----------------------------------------------------------------------
>
> Key: SYSTEMML-1583
> URL: https://issues.apache.org/jira/browse/SYSTEMML-1583
> Project: SystemML
> Issue Type: Sub-task
> Reporter: Niketan Pansare
> Assignee: Arvind Surve
>
> Ideally, this converter shouldnot require the caffe to be installed. Please
> see
> http://stackoverflow.com/questions/37572948/extracting-weights-from-caffemodel-without-caffe-installed-in-python
> An example code to convert a caffe model to csv if caffe is installed:
> {code}
> import caffe
> import numpy as np
> #net =
> caffe.Net('/home/biuser/nike/barista/VGG_ILSVRC_19_layers_train_val.prototxt',
> caffe.TEST)
> net =
> caffe.Net('/home/biuser/VGG_trained_models/VGG_ILSVRC_19_layers_deploy.prototxt',
> '/home/biuser/VGG_trained_models/VGG_ILSVRC_19_layers.caffemodel',
> caffe.TEST)
> #surgery.transplant(net, base_net)
> for l in [ "conv1_1", "conv1_2", "conv2_1", "conv2_2", "conv3_1", "conv3_2",
> "conv3_3", "conv3_4", "conv4_1", "conv4_2", "conv4_3", "conv4_4", "conv5_1",
> "conv5_2", "conv5_3", "conv5_4", "fc6", "fc7", "fc8" ]:
> w = net.params[l][0].data
> w = w.reshape(w.shape[0], -1)
> b = net.params[l][1].data
> b = b.reshape(b.shape[0], -1)
> # You may have to reshape it for fc layers
> np.savetxt("VGG_trained_models/" + l + "_weight.csv", w,
> delimiter=",")
> np.savetxt("VGG_trained_models/" + l + "_bias.csv", b, delimiter=",")
> {code}
> Here is an example pyspark script to test this JIRA:
> {code}
> from systemml.mllearn import Caffe2DML
> from pyspark.sql import SQLContext
> import numpy as np
> import urllib, os, scipy.ndimage
> from PIL import Image
> import systemml as sml
> # ImageNet specific parameters
> img_shape = (3, 224, 224)
> # Downloads a jpg image, resizes it to 224 and return as numpy array in N X
> CHW format
> url =
> 'https://upload.wikimedia.org/wikipedia/commons/thumb/5/58/MountainLion.jpg/312px-MountainLion.jpg'
> outFile = 'test.jpg'
> urllib.urlretrieve(url, outFile)
> input_image = sml.convertImageToNumPyArr(Image.open(outFile),
> img_shape=img_shape)
> # Download the ResNet network
> import urllib
> urllib.urlretrieve('https://raw.githubusercontent.com/niketanpansare/model_zoo/master/caffe/vision/resnet/ilsvrc12/ResNet_50_network.proto',
> 'ResNet_50_network.proto')
> urllib.urlretrieve('https://raw.githubusercontent.com/niketanpansare/model_zoo/master/caffe/vision/resnet/ilsvrc12/ResNet_50_solver.proto',
> 'ResNet_50_solver.proto')
> home_dir = os.path.expanduser('~')
> # let's assume that this function is implemented as
> saveAsBinaryBlock(inputCaffeModel, outputDir)
> resnet_pretrained_weight_dir = os.path.join(home_dir, 'model_zoo', 'caffe',
> 'vision', 'resnet', 'ilsvrc12', 'ResNet_50_pretrained_weights')
> urllib.urlretrieve('https://deepdetect.com/models/resnet/ResNet-50-model.caffemodel',
> 'ResNet-50-model.caffemodel')
> #######################################################################
> # To be implemented as part of this JIRA
> sml.saveAsBinaryBlock('ResNet-50-model.caffemodel',
> resnet_pretrained_weight_dir)
> #######################################################################
> resnet = Caffe2DML(sqlCtx, solver='ResNet_50_solver.proto',
> weights=resnet_pretrained_weight_dir, input_shape=img_shape)
> resnet.predict(input_image)
> # This should return array(['cougar, puma, catamount, mountain lion, painter,
> panther, Felis '], dtype='|S64')
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
