load pre-trained alexnet model

[lgarithm]

• http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/

[lgarithm]

[i] 1    34848        shape(11,11,3,96)
[i] 2    96           shape(96)
[i] 3    614400       shape(5,5,96,256)
[i] 4    256          shape(256)
[i] 5    884736       shape(3,3,256,384)
[i] 6    384          shape(384)
[i] 7    1327104      shape(3,3,384,384)
[i] 8    384          shape(384)
[i] 9    884736       shape(3,3,384,256)
[i] 10   256          shape(256)
[i] 11   37748736     shape(9216,4096)
[i] 12   4096         shape(4096)
[i] 13   16777216     shape(4096,4096)
[i] 14   4096         shape(4096)
[i] 15   4096000      shape(4096,1000)
[i] 16   1000         shape(1000)
[i] total dim: 62378344

[lgarithm]

1        conv1_W                  float32                  34848            (11, 11, 3, 96)
2        conv1_b                  float32                  96               (96,)
3        conv2_W                  float32                  307200           (5, 5, 48, 256)
4        conv2_b                  float32                  256              (256,)
5        conv3_W                  float32                  884736           (3, 3, 256, 384)
6        conv3_b                  float32                  384              (384,)
7        conv4_W                  float32                  663552           (3, 3, 192, 384)
8        conv4_b                  float32                  384              (384,)
9        fc6_W                    float32                  37748736         (9216, 4096)
10       fc6_b                    float32                  4096             (4096,)
11       fc7_W                    float32                  16777216         (4096, 4096)
12       fc7_b                    float32                  4096             (4096,)

#!/usr/bin/env python3
#
# inspect nbvlc_alexnet.npy
#
from functools import reduce
import operator as op

import numpy as np

layers = np.load(open("bvlc_alexnet.npy", "rb"), encoding="latin1").item()

layer_names = ['conv%d' % i for i in range(1, 5)] + ['fc%d' % i for i in range(6, 8)]

def list_params():
    for name in layer_names:
        w, b = layers[name]
        yield '%s_W' % name, w
        yield '%s_b' % name, b

for idx, (name, p) in enumerate(list_params()):
    dim = reduce(op.mul, p.shape, 1)
    print('%-8d %-24s %-24s %-16d %s' % (idx + 1, name, p.dtype, dim, p.shape))