def quantize_net_with_dict(net, layers, codebook, use_stochastic=False, timing=False):
start_time = time.time()
codeDict = {}
maskCode = {}
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
if use_stochastic:
codes = stochasitc_quantize2(W.flatten(), codebook[layer])
else:
codes, _ = scv.vq(W.flatten(), codebook[layer])
W_q = np.reshape(codebook[layer][codes], W.shape)
net.params[layer][0].data[...] = W_q
maskCode[layer] = np.reshape(codes, W.shape)
codeBookSize = len(codebook[layer])
a = maskCode[layer].flatten()
b = xrange(len(a))
codeDict[layer] = {}
for i in xrange(len(a)):
codeDict[layer].setdefault(a[i], []).append(b[i])
if timing:
print "Update codebook time:%f" % (time.time() - start_time)
return codeDict, maskCode
def quantize_net_with_dict(net,
layers,
codebook,
use_stochastic=False,
timing=False):
start_time = time.time()
codeDict = {}
maskCode = {}
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
if use_stochastic:
codes = stochasitc_quantize2(W.flatten(), codebook[layer])
else:
codes, _ = scv.vq(W.flatten(), codebook[layer])
W_q = np.reshape(codebook[layer][codes], W.shape)
net.params[layer][0].data[...] = W_q
maskCode[layer] = np.reshape(codes, W.shape)
codeBookSize = len(codebook[layer])
a = maskCode[layer].flatten()
b = xrange(len(a))
codeDict[layer] = {}
for i in xrange(len(a)):
codeDict[layer].setdefault(a[i], []).append(b[i])
if timing:
print "Update codebook time:%f" % (time.time() - start_time)
return codeDict, maskCode
def quantize_net(net, codebook):
layers = codebook.keys()
print "================Perform quantization=============="
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
codes, _ = scv.vq(W.flatten(), codebook[layer])
W_q = np.reshape(codebook[layer][codes], W.shape)
np.copyto(net.params[layer][0].data, W_q)
def quantize_net(net, codebook, use_stochastic=False):
layers = codebook.keys()
print "================Perform quantization=============="
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
if use_stochastic:
codes = stochasitc_quantize2(W.flatten(), codebook[layer])
else:
codes, _ = scv.vq(W.flatten(), codebook[layer])
W_q = np.reshape(codebook[layer][codes], W.shape)
np.copyto(net.params[layer][0].data, W_q)
def quantize_net(net, codebook, use_stochastic=False):
layers = codebook.keys()
print "================Perform quantization=============="
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
if use_stochastic:
codes = stochasitc_quantize2(W.flatten(), codebook[layer])
else:
codes, _ = scv.vq(W.flatten(), codebook[layer])
W_q = np.reshape(codebook[layer][codes], W.shape)
np.copyto(net.params[layer][0].data, W_q)
def get_codes(net, codebook):
layers = codebook.keys()
codes_W = {}
codes_b = {}
print "================Perform quantization=============="
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
b = net.params[layer][1].data
codes, _ = scv.vq(W.flatten(), codebook[layer])
codes = np.reshape(codes, W.shape)
codes_W[layer] = np.array(codes, dtype=np.uint32)
W_q = np.reshape(codebook[layer][codes], W.shape)
np.copyto(net.params[layer][0].data, W_q)
codes, _ = scv.vq(b.flatten(), codebook[layer])
codes = np.reshape(codes, b.shape)
codes_b[layer] = np.array(codes, dtype=np.uint32)
b_q = np.reshape(codebook[layer][codes], b.shape)
np.copyto(net.params[layer][1].data, b_q)
return codes_W, codes_b
def get_codes(net, codebook):
layers = codebook.keys()
codes_W = {}
codes_b = {}
print "================Perform quantization=============="
for layer in layers:
print "Quantize layer:", layer
W = net.params[layer][0].data
b = net.params[layer][1].data
codes, _ = scv.vq(W.flatten(), codebook[layer])
codes = np.reshape(codes, W.shape)
codes_W[layer] = np.array(codes, dtype=np.uint32)
W_q = np.reshape(codebook[layer][codes], W.shape)
np.copyto(net.params[layer][0].data, W_q)
return codes_W
def recover_all(net, dir_t, idx=0):
layers = net.params.keys()
net.copy_from(dir_t + 'caffemodel%d' % idx)
codebook = pickle.load(open(dir_t + 'codebook%d' % idx))
maskCode = {}
codeDict = {}
for layer in layers:
W = net.params[layer][0].data
codes, _ = scv.vq(W.flatten(), codebook[layer])
maskCode[layer] = np.reshape(codes, W.shape)
codeBookSize = len(codebook[layer])
a = maskCode[layer].flatten()
b = xrange(len(a))
codeDict[layer] = {}
for i in xrange(len(a)):
codeDict[layer].setdefault(a[i], []).append(b[i])
return codebook, maskCode, codeDict
def recover_all(net, dir_t, idx=0):
layers = net.params.keys()
net.copy_from(dir_t + 'caffemodel%d' % idx)
codebook = pickle.load(open(dir_t + 'codebook%d' % idx))
maskCode = {}
codeDict = {}
for layer in layers:
W = net.params[layer][0].data
codes, _ = scv.vq(W.flatten(), codebook[layer])
maskCode[layer] = np.reshape(codes, W.shape)
codeBookSize = len(codebook[layer])
a = maskCode[layer].flatten()
b = xrange(len(a))
codeDict[layer] = {}
for i in xrange(len(a)):
codeDict[layer].setdefault(a[i], []).append(b[i])
return codebook, maskCode, codeDict
return ptr, spm, ind
W = np.load('W.npy').astype('f')
act = np.zeros(W.shape[1], dtype='int16')
act[0] = 1
act[1] = 2
act[5] = 1
groundtruth = np.dot(W, act)
groundtruth = (groundtruth * 32).astype('int16')
codebook = np.arange(16).astype('f') / 32
W_codes, _ = scv.vq(W.flatten(), codebook)
W_codes = np.reshape(W_codes, W.shape)
ptr, spm, ind = get_csc_single_nobias(W_codes, bank_num=2, max_jump=16)
data_dir = 'test_data'
os.system("rm -rf " + data_dir)
os.system("mkdir " + data_dir)
os.system("mkdir " + data_dir + '/ptr')
os.system("mkdir " + data_dir + '/spm')
for idx in range(2):
with open("%s/ptr/ptr%d.dat" % (data_dir, idx), 'wb') as f:
f.write('%d\n' % len(ptr[idx]))
for number in ptr[idx]:
f.write('{:016b} '.format(number))
