def save_checkpoint(self, name):
ckp_path = os.path.join(self.model_dir, name)
tmp = {
'network': self.net,
'opt_state': self.train_func.optimizer_state.make_checkpoint(),
'clock': self.clock.make_checkpoint(),
}
io.dump(tmp, ckp_path)
nonlinearity=mgsk.opr.helper.elemwise_trans.Identity())
f = Softmax("cls_softmax", f)
f.init_weights()
net = RawNetworkBuilder(inputs=[data], outputs=[f])
return net
if __name__ == '__main__':
parser = argparse.ArgumentParser(
description='dump pkl model for resnet50',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('-b', '--batch-size',
help='batch size of the model', default=1)
parser.add_argument('-f', '--format', choices=['NCHW', 'NCHW4'],
help='format of conv',
default='NCHW')
parser.add_argument('-o', '--output',
help='output pkl path', required=True)
args = parser.parse_args()
if args.format != 'NCHW':
print('Only suppprt NCHW for float model')
parser.print_help()
sys.exit(1)
from meghair.utils import io
io.dump(get(args), args.output)
def R3(self):
print("entered R3!!!!")
oprs_dict = self.oprs_dict
new_oprs_dict = self.new_oprs_dict
DEBUG = True
convs = self.convs
end = 5
speed_ratio = 4
alldic = ['conv%d_1' % i for i in range(1, end)
] + ['conv%d_2' % i for i in range(3, end)]
pooldic = {
'conv1_2': 'pool1',
'conv2_2': 'pool2',
'conv3_3': 'pool3',
'conv4_3': 'pool4'
}
rankdic = {
'conv1_1': 24,
'conv1_2': 22,
'conv2_1': 41,
'conv2_2': 51,
'conv3_1': 108,
'conv3_2': 89,
'conv3_3': 111,
'conv4_1': 184,
'conv4_2': 276,
'conv4_3': 228,
'conv5_1': 512,
'conv5_2': 512,
'conv5_3': 512
}
for i in rankdic:
if 'conv5' in i:
continue # the break-statemet was giving a bug, so changed it to continue-statement -by Mario
rankdic[i] = int(rankdic[i] * 4. / speed_ratio)
c_ratio = 1.15
t = Timer()
for conv, convnext in zip(
convs[0:], convs[1:] + ['pool5']
): # note that we exclude the first conv, conv1_1 contributes little computation -by Mario
#W_shape = self.param_shape(conv)
#d_c = int(W_shape[0] / c_ratio)
#rank = rankdic[conv]
#d_prime = rank
#if d_c < rank: d_c = rank
d_c = rankdic[conv]
print("channel pruning", conv, convnext)
'''channel pruning'''
if (conv in alldic or conv in pooldic) and (convnext
in self.convs):
t.tic()
if conv in pooldic:
X_name = pooldic[conv]
else:
X_name = conv
print(X_name, convnext)
idxs, W2, B2 = self.dictionary_kernel(X_name, d_c, convnext)
# W2
W_new = self.param_data(convnext)
W_new[:, ~idxs, ...] = 0
W_new[:, idxs, ...] = W2.copy()
self.new_oprs_dict[convnext + ':W'].set_value(W_new)
self.new_oprs_dict[convnext + ':b'].set_value(B2)
t.toc('channel_pruning')
print("channel pruning finished")
io.dump(self.new_net, 'outputs/' + X_name + 'pruned_model.save')
self.val()
io.dump(self.new_net, 'pruned_model.save')
def make_network(model, isize):
data = [
O.DataProvider('input{}'.format(i), shape=isizes[i])
for i in range(len(isizes))
]
f = open(model, 'rb')
engine = f.read()
opr = TensorRTRuntimeOpr(data, engine, 1)
net = RawNetworkBuilder(inputs=[data], outputs=opr.outputs)
return net
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(dest='model')
parser.add_argument(dest='output')
parser.add_argument(
'--isize',
help='input sizes. '
'e.g. for models with two (1,3,224,224) inputs, '
'the option --isize="1,3,224,224;1,3,224,224" should be used')
args = parser.parse_args()
isizes = [str2tuple(x) for x in args.isize.split(';')]
net = make_network(args.model, isizes)
dump(net, args.output)
func_test = Function().compile(network.pred)
import pickle
import gzip
import numpy as np
train_set, valid_set, test_set = pickle.load(gzip.open("mnist.pkl.gz", "rb"),
encoding="latin1")
minibatch_size = network.minibatch_size
l = len(train_set[0])
epoch = 0
for i in range(100000):
j = i % (l // minibatch_size)
minibatch = train_set[0][j * minibatch_size:(j + 1) * minibatch_size]
label = train_set[1][j * minibatch_size:(j + 1) * minibatch_size]
minibatch = np.array(minibatch).reshape(-1, 1, 28, 28)
loss = func(input_mat=minibatch, label=label)
if j == 0:
print("*****")
print("epoch = ", epoch)
epoch += 1
print("loss = ", loss)
res = func_test(
input_mat=np.array(valid_set[0]).reshape(-1, 1, 28, 28))
res = np.argmax(np.array(res), axis=1)
acc = (np.array(res) == np.array(valid_set[1])).mean()
print("acc = ", acc)
dump(network.network, open("1.data", "wb"))