def print_net_info(model_path, phase, net_params):
# disable caffe logging
net = Net(model_path, phase=phase)
# Blobs = Data, net_params.layer = Layers <=> wont be equal
nb_layers = len(net_params.layer)
print "Net architecture : ", model_path
print " - layer name (layer type) : (batch size, feature dim, width, height) | number of filters, (filter dim, filter width, filter height), stride, padding"
counter=0
for k, v in net.blobs.items():
params = []
if k in net.params:
params = net.params[k][0].data.shape
padding = None
stride = None
layer_type = None
if nb_layers > 0:
layer = net_params.layer[counter]
layer_type = layer.type
if layer_type == 'Convolution' or layer_type == 'Deconvolution':
stride = layer.convolution_param.stride[0] if len(layer.convolution_param.stride._values) else 1
padding = layer.convolution_param.pad[0] if len(layer.convolution_param.pad._values) else 0
elif layer_type == 'Pooling':
stride = layer.pooling_param.stride
padding = layer.pooling_param.pad
if padding and stride and layer_type is not None:
print "%s (%s) : %s | %s, %s, stride=%s, padding=%s" %(k, layer_type, v.data.shape, params[0], params[1:], stride, padding)
else:
print "%s : %s | %s, %s" %(k, v.data.shape, params[0], params[1:])
else:
print "%s : %s" %(k, v.data.shape)
counter += 1
def build(self, network_filename='inception.prototxt'):
"""main method."""
network = self._build_network()
print network
with open(network_filename, 'w') as network_file:
network_file.write(text_format.MessageToString(network))
return Net(network_filename)
def build(self):
"""main method."""
network = self._build_network()
print network
network_filename = 'inception.prototxt'
with open(network_filename, 'w') as network_file:
network_file.write(text_format.MessageToString(network))
net = Net(network_filename)
net.set_phase_test()
net.set_mode_gpu()
print net.forward()
def build_network(self, netname, batchsize=1, numstep=24):
"""main method."""
if netname == 'rnn':
network = self._build_rnn_network(batchsize=batchsize, numstep=numstep)
else:
print('unknown netname: %s' % netname)
return
network_filename = '%s_t%d_finetuning.prototxt' % (netname,numstep)
print network
with open(network_filename, 'w') as network_file:
network_file.write(text_format.MessageToString(network))
return Net(network_filename)
def classify(
image,
model=_static_file("deploy.prototxt"),
weights=_static_file("resnet_50_1by2_nsfw.caffemodel"),
):
"""
Determine the probability that an image is SFW or NSFW.
Parameters
----------
image : a PIL ImageFile object
Keyword Arguments
-----------------
These arguments default to Yahoo's open_nsfw defaults. If you have
your own trained models, you may pass the paths to the `.prototxt` and
`.caffemodel` files.
model : a string path to a Caffe model file (e.g. deploy.prototxt)
weights : a string path to a Caffe weights file (e.g. caffenet.caffemodel)
Returns
-------
(sfw, nsfw) : a tuple with SFW and NSFW probabilities
"""
net = Net(model, 1, weights=weights)
transformer = Transformer({
"data": net.blobs["data"].data.shape
})
# Move image channels to outermost
transformer.set_transpose("data", (2, 0, 1))
# Subtract the dataset-mean value in each channel
transformer.set_mean("data", numpy.array([104, 117, 123]))
# Rescale from [0, 1] to [0, 255]
transformer.set_raw_scale("data", 255)
# Swap channels from RGB to BGR
transformer.set_channel_swap("data", (2, 1, 0))
sfw, nsfw = _process(_resize(image), net=net, transformer=transformer)
return (sfw, nsfw)
X = np.rollaxis(bottom[0].data, 1, 4)
batch_size, h, w, _ = X.shape
num_p = h * w
BY = bottom[1].data
BY[:, 1:5] /= self.scale
obj_score, Y, gt_class = process(X, BY)
top[0].data[...] = obj_score
top[1].data[...] = Y
top[2].data[...] = gt_class
def backward(self, top, propagate_down, bottom):
pass
if __name__ == '__main__':
from caffe import Net, TRAIN
import cv2
net = Net('od2.pt', 'od_train_iter_62692.caffemodel', TRAIN)
for count in range(100):
net.forward()
Bx, score = reconstruct(net.blobs['bbox'].data,
net.blobs['obj_score'].data,
net.blobs['label'].data)
img = visualize('data/pvtdb/VOC2007/JPEGImages/%06d.jpg' % count, \
Bx, net.blobs['data'].data, 32)
cv2.imshow('%06d.jpg' % count, img)
key = cv2.waitKey(0)
cv2.destroyAllWindows()
if key == 27:
break
By = gt_boxes[gt_boxes[:,0] == n, 1:5]
IOU = iou(Bx, By)
print IOU.max(axis=0)
return (Bx, score)
def visualize(filename, Bx, data, scale):
img = data[0].copy()
img[0] += 103
img[1] += 116
img[2] += 123
img = np.array(np.rollaxis(img, 0, 3), \
dtype=np.uint8).copy()
for bx in Bx * scale:
cv2.rectangle(img, (bx[0], bx[1]), (bx[2], bx[3]), \
(0, 0, 255), 2)
return img
if __name__ == '__main__':
from caffe import Net, TRAIN
net = Net('caffe-materials/practice6/net-test.pt', 'caffe-materials/practice6/net_trained.cm', TRAIN)
for count in range(100):
net.forward()
Bx, score = reconstruct(net.blobs['bbox'].data, net.blobs['obj_score'].data, net.blobs['label'].data)
img = visualize('data/pvtdb/VOC2007/JPEGImages/%06d.jpg' % count, \
Bx, net.blobs['data'].data, 32)
cv2.imshow('%06d.jpg' % count, img)
key = cv2.waitKey(0)
cv2.destroyAllWindows()
if key == 27:
break
#! python
#coding=utf-8
import os, sys
caffe_root = '/data/darwin-caffe-base/' # caffe的根目录
os.chdir(caffe_root) #os.chdir()用于改变当前工作目录到指定的路径
sys.path.insert(0, caffe_root + 'python')
from caffe import Net, TEST
import caffe
import numpy as np
net = Net('/data/mgn_caffe/resnet50/ResNet-50-deploy.prototxt',
'/data/mgn_caffe/resnet50/ResNet-50-model.caffemodel', caffe.TEST)
netNew = Net('/data/mgn_caffe/resnet50/ResNet-50-deploy-3.prototxt',
caffe.TEST)
for k1, v1 in netNew.params.items():
for k, v in net.params.items():
#print (k, v[0].data.shape)
#print np.size(net.params[k])
for i in range(np.size(net.params[k])):
if (k1 == k + '_3'):
print(k1)
netNew.params[k1][i].data[:] = np.copy(
net.params[k][i].data[:])
for i in range(np.size(net.params[k])):
if (k1 == k):
print(k1)
imageToTest = cv.resize(oriImg, (0, 0),
fx=scale,
fy=scale,
interpolation=cv.INTER_CUBIC)
imageToTest_padded, pad = util.padRightDownCorner(
imageToTest, model['stride'], model['padValue'])
print(imageToTest_padded.shape)
net.blobs['data'].reshape(*(1, 3, imageToTest_padded.shape[0],
imageToTest_padded.shape[1]))
# net.forward() # dry run
net.blobs['data'].data[...] = np.transpose(
np.float32(imageToTest_padded[:, :, :, np.newaxis]),
(3, 2, 0, 1)) / 256 - 0.5
start_time = time.time()
output_blobs = net.forward()
print('At scale %d, The CNN took %.2f ms.' %
(m, 1000 * (time.time() - start_time)))
# extract outputs, resize, and remove padding
heatmap = np.transpose(
np.squeeze(net.blobs[list(output_blobs.keys())[1]].data),
(1, 2, 0)) # output 1 is heatmaps
heatmap = cv.resize(heatmap, (0, 0),
fx=model['stride'],
fy=model['stride'],
interpolation=cv.INTER_CUBIC)
heatmap = heatmap[:imageToTest_padded.shape[0] -
pad[2], :imageToTest_padded.shape[1] - pad[3], :]
heatmap = cv.resize(heatmap, (oriImg.shape[1], oriImg.shape[0]),
interpolation=cv.INTER_CUBIC)
def visualize(filename, Bx, data, scale):
img = data[0].copy()
img[0] += 103
img[1] += 116
img[2] += 123
img = np.array(np.rollaxis(img, 0, 3), \
dtype=np.uint8).copy()
for bx in Bx * scale:
cv2.rectangle(img, (bx[0], bx[1]), (bx[2], bx[3]), \
(0, 0, 255), 2)
return img
if __name__ == '__main__':
from caffe import Net, TRAIN
net = Net('caffe-materials/practice6/net-test.pt',
'caffe-materials/practice6/net_trained.cm', TRAIN)
for count in range(100):
net.forward()
Bx, score = reconstruct(net.blobs['bbox'].data,
net.blobs['obj_score'].data,
net.blobs['label'].data)
img = visualize('data/pvtdb/VOC2007/JPEGImages/%06d.jpg' % count, \
Bx, net.blobs['data'].data, 32)
cv2.imshow('%06d.jpg' % count, img)
key = cv2.waitKey(0)
cv2.destroyAllWindows()
if key == 27:
break
#! python
#coding=utf-8
import sys
from caffe import Net
caffe_root = your / caffe / root
sys.path.insert(0, caffe_root + 'python')
n_orig = Net('orig_model.prototxt', 'orig_model.caffemodel', 0)
n_new = Net('new_model.prototxt', 'orig_model.caffemodel', 0)
for orig_name in n_orig.params:
#n_new.params[orig_name + '_p'] = n_orig.params[orig_name]
n_new.params[new_name] = n_orig.params[orig_name]
n_new.save('new_model.caffemodel')