因为比赛过程中，不停尝试新的想法，所以代码比较乱！有时间会抽空整理一下！
# coding:utf-8
import caffe
import numpy as np
import collections
class Pruner(object):
def __init__(self, net):
self._net = net
self.conv_data = collections.OrderedDict()
def _prune(self, conv_param, del_kernels=None, not_del_filters=False, th=0.0001):
weight, bias = conv_param
weight = weight.data
bias = bias.data
origin_channels = weight.shape[0]
# thre = 0.05
# delete filters
if not not_del_filters:
abs_mean = np.abs(weight).mean(axis=(1, 2, 3))
# abs_mean = np.sum(np.abs(weight), axis=(1, 2, 3))
del_filters = np.where(abs_mean < th)[0]
# abs_mean_sort = sorted(abs_mean)
# del_filters_2 = np.where(abs_mean < abs_mean_sort[int(origin_channels*thre)])[0]
# if len(del_filters_1) > len(del_filters_2):
# del_filters = del_filters_1
# else:
# del_filters = del_filters_2
weight = np.delete(weight, del_filters, axis=0)
bias = np.delete(bias, del_filters, axis=0)
else:
del_filters = np.array([])
# delete kernels
if del_kernels is not None:
weight = np.delete(weight, del_kernels, axis=1)
return weight, bias, del_filters, origin_channels
def prune_conv(self, name, bottom=None, not_del_filters=False, th=0.0001):
if bottom is None:
self.conv_data[name] = self._prune(self._net.params[name], not_del_filters=not_del_filters, th=th)
else:
self.conv_data[name] = self._prune(self._net.params[name], self.conv_data[bottom][2], not_del_filters=not_del_filters, th=th)
def prune_fc(self, name):
weight, bias = self._net.params[name]
weight = weight.data
bias = bias.data
print(bias.shape)
abs_sum = np.sum(np.abs(weight), axis=0)
del_filters = np.where(abs_sum < 3)[0]
print(del_filters)
print(len(del_filters))
weight = np.delete(weight, del_filters, axis=1)
# bias = np.delete(bias, del_filters, axis=1)
self.conv_data[name] = weight, bias
def prune_copy(self, name, name_, bottom=None):
weight, bias = self._net.params[name]
weight = weight.data
bias = bias.data
origin_channels = weight.shape[0]
del_filters = self.conv_data[name_][2]
del_kernels = self.conv_data[bottom][2]
weight = np.delete(weight, del_filters, axis=0)
bias = np.delete(bias, del_filters, axis=0)
if del_kernels is not None:
weight = np.delete(weight, del_kernels, axis=1)
self.conv_data[name] = weight, bias, del_filters, origin_channels
def prune_bias(self, name, ind):
weight, bias = self._net.params[name]
bias = bias.data
del_filters = ind
bias = np.delete(bias, del_filters, axis=0)
return bias
def prune_id(self, name, ind, bottom=None, bias_l=None):
weight, bias = self._net.params[name]
weight = weight.data
bias = bias.data
origin_channels = weight.shape[0]
del_filters = ind
del_kernels = []
if bottom is not None:
del_kernels = self.conv_data[bottom][2]
weight = np.delete(weight, del_filters, axis=0)
bias = np.delete(bias, del_filters, axis=0)
if bias_l is not None:
for b in bias_l:
bias = bias + b
if del_kernels is not None:
weight = np.delete(weight, del_kernels, axis=1)
self.conv_data[name] = weight, bias, del_filters, origin_channels
def prune_concat(self, name, bottoms, not_del_filters=False, is_delete=False):
offsets = [0] + [self.conv_data[b][3] for b in bottoms]
for i in range(1, len(offsets)):
offsets[i] += offsets[i - 1]
del_filters = [self.conv_data[b][2] + offsets[i] for i, b in enumerate(bottoms)]
if is_delete:
del_filters.append(range(192, 256))
# if is_delete_1:
# del_filters.append(range(64, 96))
del_filters_new = np.concatenate(del_filters)
self.conv_data[name] = self._prune(self._net.params[name], del_filters_new, not_del_filters=not_del_filters)
def fc_concat(self, name, bottoms):
assert ('fc' in name)
weight, bias = self._net.params[name]
weight = weight.data
bias = bias.data
offset_1 = 9*9*512
offset_2 = 9*9
del_filters = []
for ind in self.conv_data[bottoms[0]][2]:
del_filters += range(offset_2*ind, offset_2 * (ind+1))
# for ind in self.conv_data[bottoms[1]][2]:
# del_filters += range(offset_2*ind+offset_1, offset_2 * (ind+1)+offset_1)
# for ind in self.conv_data[bottoms[1]][2]:
# del_filters += range(offset_2*ind+offset_1*2, offset_2 * (ind+1)+offset_1*2)
del_filters += range(offset_1 * 1, offset_1 * 2)
del_filters += range(offset_1 * 2, offset_1 * 3)
weight = np.delete(weight, del_filters, axis=1)
self.conv_data[name] = weight, bias
def save(self, new_model, output_weights):
net2 = caffe.Net(new_model, caffe.TEST)
for key in net2.params.keys():
if key in self.conv_data:
net2.params[key][0].data[...] = self.conv_data[key][0]
net2.params[key][1].data[...] = self.conv_data[key][1]
else:
net2.params[key][0].data[...] = self._net.params[key][0].data
net2.params[key][1].data[...] = self._net.params[key][1].data
net2.save(output_weights)
def test(self, net, input_image=" ", output_tensor="fc5_"):
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2, 0, 1))
transformer.set_channel_swap('data', (2, 1, 0))
image = caffe.io.load_image(input_image)
transformed_image = transformer.preprocess('data', (image * 255 - 127.5) * 0.0078125)
net.blobs['data'].reshape(1, 3, 128, 128)
net.blobs['data'].data[...] = transformed_image
output = net.forward(end=output_tensor)
return output[output_tensor]
def get_feature_id(self, output_tensor):
image_path = './images/5.JPEG'
output = self.test(self._net, image_path, output_tensor)
h = output.shape[2]
w = output.shape[3]
channels = output.shape[1]
# assert channels % 3 == 0
output = output.reshape(channels, h, w)
# print(output_tensor.shape)
sum_l1 = []
ind_2_2 = []
ind_4_2 = []
ind_6_2 = []
ind = []
for i in range(0, channels):
sum_ = np.sum(abs(output[i, :, :]))
sum_l1.append(sum_)
if (sum_ < 7.98):
ind_2_2.append(i)
ind.append(ind_2_2)
# for i in range(channels / 3, 2*channels / 3):
# sum_ = np.sum(abs(output[i, :, :]))
# sum_l1.append(sum_)
# if (sum_ < 0.0001):
# ind_4_2.append(i-channels / 3)
# ind.append(ind_4_2)
#
# for i in range(2*channels / 3, 3*channels / 3):
# sum_ = np.sum(abs(output[i, :, :]))
# sum_l1.append(sum_)
# if (sum_ < 0.0001):
# ind_6_2.append(i-2*channels / 3)
# ind.append(ind_6_2)
# ind_b = list(set(ind[0]).intersection(set(ind[1])))
# ind_b = list(set(ind_b).intersection(set(ind[2])))
print(len(ind))
return ind
def main():
caffe.set_mode_gpu()
net = caffe.Net("./models/svd/TestModel.prototxt", "./models/svd/TestModel.caffemodel", caffe.TEST)
# net = caffe.Net("./models/merge/merge_bn.prototxt", "./models/merge/merge_bn.caffemodel", caffe.TEST)
pruner = Pruner(net)
pruner.prune_conv("conv1_1_1")
pruner.prune_conv("conv1_2_1")
pruner.prune_conv("conv1_2_2", "conv1_2_1")
pruner.prune_conv("conv1_3_1")
pruner.prune_conv("conv1_3_2", "conv1_3_1")
pruner.prune_conv("conv1_3_3", "conv1_3_2")
pruner.prune_concat("conv2_1", ("conv1_1_1", "conv1_2_2", "conv1_3_3"))
pruner.prune_conv("conv2_2", "conv2_1")
pruner.prune_conv("conv2_3", "conv2_2")
pruner.prune_conv("conv2_4", "conv2_3")
pruner.prune_conv("conv2_5", "conv2_4")
pruner.prune_conv("conv2_6", "conv2_5")
# pruner.prune_conv("conv2_7", "conv2_6")
# pruner.prune_conv("conv2_8", "conv2_6")
pruner.prune_concat("conv3_1_1", ("conv2_2", "conv2_4", "conv2_6"), is_delete=True)
pruner.prune_concat("conv3_1_1b", ("conv2_2", "conv2_4", "conv2_6"), True, True)
# residual
# pruner.prune_copy("conv3_1_2", "conv3_1_1b", "conv3_1_1")
# pruner.prune_conv("conv3_2_1", "conv3_1_1b")
# pruner.prune_copy("conv3_2_2", "conv3_1_1b", "conv3_2_1")
# pruner.prune_conv("conv3_3_1", "conv3_1_1b")
# pruner.prune_copy("conv3_3_2", "conv3_1_1b", "conv3_3_1")
# pruner.prune_conv("conv3_4_1", "conv3_1_1b")
# pruner.prune_copy("conv3_4_2", "conv3_1_1b", "conv3_4_1")
# pruner.prune_conv("conv3_5_1", "conv3_1_1b")
# pruner.prune_copy("conv3_5_2", "conv3_1_1b", "conv3_5_1")
# pruner.prune_conv("conv3_6_1", "conv3_1_1b")
# pruner.prune_copy("conv3_6_2", "conv3_1_1b", "conv3_6_1")
pruner.prune_conv("conv3_1_2", "conv3_1_1", not_del_filters=True)
pruner.prune_conv("conv3_2_1")
pruner.prune_conv("conv3_2_2", "conv3_2_1", not_del_filters=True)
pruner.prune_conv("conv3_3_1")
pruner.prune_conv("conv3_3_2", "conv3_3_1", not_del_filters=True)
pruner.prune_conv("conv3_4_1")
pruner.prune_conv("conv3_4_2", "conv3_4_1", not_del_filters=True)
pruner.prune_conv("conv3_5_1")
pruner.prune_conv("conv3_5_2", "conv3_5_1", not_del_filters=True)
pruner.prune_conv("conv3_6_1")
pruner.prune_conv("conv3_6_2", "conv3_6_1", not_del_filters=True)
#pruner.prune_concat("conv4_1_1", ("conv3_2_2", "conv3_4_2", "conv3_6_2"))
#pruner.prune_concat("conv4_1_1b", ("conv3_2_2", "conv3_4_2", "conv3_6_2"))
# pruner.prune_copy("conv4_1_2", "conv4_1_1b", "conv4_1_1")
# pruner.prune_conv("conv4_2_1", "conv4_1_1b")
# pruner.prune_copy("conv4_2_2", "conv4_1_1b", "conv4_2_1")
# pruner.prune_conv("conv4_3_1", "conv4_1_1b")
# pruner.prune_copy("conv4_3_2", "conv4_1_1b", "conv4_3_1")
# pruner.prune_conv("conv4_4_1", "conv4_1_1b")
# pruner.prune_copy("conv4_4_2", "conv4_1_1b", "conv4_4_1")
# pruner.prune_conv("conv4_5_1", "conv4_1_1b")
# pruner.prune_copy("conv4_5_2", "conv4_1_1b", "conv4_5_1")
# pruner.prune_conv("conv4_6_1", "conv4_1_1b")
# pruner.prune_copy("conv4_6_2", "conv4_1_1b", "conv4_6_1")
pruner.prune_conv("conv4_1_1")
pruner.prune_conv("conv4_1_2", "conv4_1_1", not_del_filters=True)
pruner.prune_conv("conv4_2_1")
pruner.prune_conv("conv4_2_2", "conv4_2_1", not_del_filters=True)
pruner.prune_conv("conv4_3_1")
pruner.prune_conv("conv4_3_2", "conv4_3_1", not_del_filters=True)
pruner.prune_conv("conv4_4_1")
pruner.prune_conv("conv4_4_2", "conv4_4_1", not_del_filters=True)
pruner.prune_conv("conv4_5_1")
pruner.prune_conv("conv4_5_2", "conv4_5_1", not_del_filters=True)
pruner.prune_conv("conv4_6_1")
pruner.prune_conv("conv4_6_2", "conv4_6_1", not_del_filters=True)
# pruner.prune_concat("conv5_1_1", ("conv4_2_2", "conv4_4_2", "conv4_6_2"))
# pruner.prune_concat("conv5_1_1b", ("conv4_2_2", "conv4_4_2", "conv4_6_2"), True)
ind = pruner.get_feature_id(output_tensor='conv5_1_1b')
print(len(ind[0]))
bias_l = []
bias1 = pruner.prune_bias("conv5_1_1", ind[0])
bias_l.append(bias1)
bias2 = pruner.prune_bias("conv5_1_2", ind[0])
bias_l.append(bias2)
# pruner.prune_conv("conv5_1_1b")
# pruner.prune_concat("conv5_1_1b", ("conv4_2_2", "conv4_4_2"), is_delete_4=True)
pruner.prune_id("conv5_1_1b", ind[0], bias_l=bias_l)
# pruner.prune_id("conv5_1_2", ind[0], "conv5_1_1")
# pruner.prune_conv("conv5_2_1", "conv5_1_2")
# pruner.prune_id("conv5_2_2", ind[0], "conv5_2_1")
# pruner.prune_conv("conv5_3_1", "conv5_2_2")
# pruner.prune_id("conv5_3_2", ind[0], "conv5_3_1")
# pruner.prune_conv("conv5_4_1", "conv5_3_2")
# pruner.prune_conv("conv5_4_1", "conv5_2_2")
# pruner.prune_id("conv5_4_2", ind[0], "conv5_4_1")
# pruner.prune_conv("conv5_5_1", "conv5_4_2")
# pruner.prune_id("conv5_5_2", ind[0], "conv5_5_1")
# pruner.prune_conv("conv5_6_1", "conv5_5_2")
# pruner.prune_id("conv5_6_2", ind[0], "conv5_6_1")
pruner.fc_concat('fc5_svd', ("conv5_1_1b",))
# pruner.fc_concat('fc5_', ("conv5_1_1b",))
# pruner.prune_conv("conv5_1_1")
# pruner.prune_conv("conv5_1_2", "conv5_1_1", not_del_filters=True)
# pruner.prune_conv("conv5_2_1")
# pruner.prune_conv("conv5_2_2", "conv5_2_1", not_del_filters=True)
# pruner.prune_conv("conv5_3_1")
# pruner.prune_conv("conv5_3_2", "conv5_3_1", not_del_filters=True)
# pruner.prune_conv("conv5_4_1")
# pruner.prune_conv("conv5_4_2", "conv5_4_1", not_del_filters=True)
# pruner.prune_conv("conv5_5_1")
# pruner.prune_conv("conv5_5_2", "conv5_5_1", not_del_filters=True)
# pruner.prune_conv("conv5_6_1")
# pruner.prune_conv("conv5_6_2", "conv5_6_1", not_del_filters=True)
#pruner.prune_fc("fc5_")
print([(k, v[0].shape[0]) for k, v in pruner.conv_data.items() if v[0] is not None])
# You should modify the number of channels in new prototxt before save
# pruner.save("./models/prune_0.05/150/TestModel.prototxt", "./models/prune_0.05/150/TestModel.caffemodel")
pruner.save("./models/prune_fc/TestModel.prototxt", "./models/prune_fc/TestModel.caffemodel")
if __name__ == '__main__':
main()
775
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
