import argparse
import chainer
import mxnet as mx
from chainercv.experimental.links import FCISResNet101
def main():
parser = argparse.ArgumentParser(
description='Script to convert mxnet params to chainer npz')
parser.add_argument(
'mxnet_param_file', metavar='mxnet-param-file',
help='Mxnet param file i.e. fcis_coco-0000.params')
parser.add_argument('--process', action='store_true')
parser.add_argument(
'--dataset', choices=('sbd', 'coco'), type=str, default='sbd')
parser.add_argument(
'--out', '-o', type=str, default=None)
args = parser.parse_args()
if args.dataset == 'sbd':
model = FCISResNet101(
n_fg_class=20,
pretrained_model=None)
elif args.dataset == 'coco':
model = FCISResNet101(
n_fg_class=80,
pretrained_model=None,
anchor_scales=[4, 8, 16, 32],
proposal_creator_params={
'nms_thresh': 0.7,
'n_train_pre_nms': 6000,
'n_train_post_nms': 300,
'n_test_pre_nms': 6000,
'n_test_post_nms': 300,
'force_cpu_nms': False,
'min_size': 2})
params = mx.nd.load(args.mxnet_param_file)
print('mxnet param is loaded: {}'.format(args.mxnet_param_file))
print('start conversion')
if args.process:
tests = [k for k in params.keys() if k.endswith('_test')]
for test in tests:
params[test.replace('_test', '')] = params.pop(test)
model = convert(model, params)
print('finish conversion')
if args.out is None:
out = 'fcis_resnet101_{}_converted.npz'.format(args.dataset)
print('saving to {}'.format(out))
chainer.serializers.save_npz(out, model)
def convert(model, params):
finished_keys = []
for key, value in params.items():
value = value.asnumpy()
param_type, param_name = key.split(':')
if param_type == 'arg':
if param_name.endswith('_test'):
continue
elif param_name.startswith('rpn'):
if param_name == 'rpn_bbox_pred_bias':
value = value.reshape((-1, 4))
value = value[:, [1, 0, 3, 2]]
value = value.reshape(-1)
assert model.rpn.loc.b.shape == value.shape
model.rpn.loc.b.array[:] = value
finished_keys.append(key)
elif param_name == 'rpn_bbox_pred_weight':
value = value.reshape((-1, 4, 512, 1, 1))
value = value[:, [1, 0, 3, 2]]
value = value.reshape((-1, 512, 1, 1))
assert model.rpn.loc.W.shape == value.shape
model.rpn.loc.W.array[:] = value
finished_keys.append(key)
elif param_name == 'rpn_cls_score_bias':
value = value.reshape((2, -1))
value = value.transpose((1, 0))
value = value.reshape(-1)
assert model.rpn.score.b.shape == value.shape
model.rpn.score.b.array[:] = value
finished_keys.append(key)
elif param_name == 'rpn_cls_score_weight':
value = value.reshape((2, -1, 512, 1, 1))
value = value.transpose((1, 0, 2, 3, 4))
value = value.reshape((-1, 512, 1, 1))
assert model.rpn.score.W.shape == value.shape
model.rpn.score.W.array[:] = value
finished_keys.append(key)
elif param_name == 'rpn_conv_3x3_bias':
assert model.rpn.conv1.b.shape == value.shape
model.rpn.conv1.b.array[:] = value
finished_keys.append(key)
elif param_name == 'rpn_conv_3x3_weight':
assert model.rpn.conv1.W.shape == value.shape
model.rpn.conv1.W.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('conv1'):
if param_name == 'conv1_weight':
assert model.extractor.conv1.conv.W.shape \
== value.shape
model.extractor.conv1.conv.W.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('bn_conv1'):
if param_name == 'bn_conv1_beta':
assert model.extractor.conv1.bn.beta.shape \
== value.shape
model.extractor.conv1.bn.beta.array[:] = value
finished_keys.append(key)
elif param_name == 'bn_conv1_gamma':
assert model.extractor.conv1.bn.gamma.shape \
== value.shape
model.extractor.conv1.bn.gamma.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('fcis'):
if param_name == 'fcis_bbox_bias':
value = value.reshape((2, 4, 7 * 7))
value = value[:, [1, 0, 3, 2]]
value = value.reshape(392)
assert model.head.ag_loc.b.shape == value.shape
model.head.ag_loc.b.array[:] = value
finished_keys.append(key)
elif param_name == 'fcis_bbox_weight':
value = value.reshape((2, 4, 7 * 7, 1024, 1, 1))
value = value[:, [1, 0, 3, 2]]
value = value.reshape((392, 1024, 1, 1))
assert model.head.ag_loc.W.shape == value.shape
model.head.ag_loc.W.array[:] = value
finished_keys.append(key)
elif param_name == 'fcis_cls_seg_bias':
assert model.head.cls_seg.b.shape == value.shape
model.head.cls_seg.b.array[:] = value
finished_keys.append(key)
elif param_name == 'fcis_cls_seg_weight':
assert model.head.cls_seg.W.shape == value.shape
model.head.cls_seg.W.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('conv_new_1'):
if param_name == 'conv_new_1_bias':
assert model.head.conv1.b.shape == value.shape
model.head.conv1.b.array[:] = value
finished_keys.append(key)
elif param_name == 'conv_new_1_weight':
assert model.head.conv1.W.shape == value.shape
model.head.conv1.W.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('res'):
block_name, branch_name, prm_name = param_name.split('_')
resblock_name = block_name[:4]
resblock = getattr(model.extractor, resblock_name)
if block_name[4:] == 'a':
blck_name = block_name[4:]
elif block_name[4:] == 'b':
blck_name = 'b1'
elif block_name[4:].startswith('b'):
blck_name = block_name[4:]
elif block_name[4:] == 'c':
blck_name = 'b2'
block = getattr(resblock, blck_name)
if branch_name == 'branch1':
conv_bn_name = 'residual_conv'
elif branch_name == 'branch2a':
conv_bn_name = 'conv1'
elif branch_name == 'branch2b':
conv_bn_name = 'conv2'
elif branch_name == 'branch2c':
conv_bn_name = 'conv3'
conv_bn = getattr(block, conv_bn_name)
if prm_name == 'weight':
assert conv_bn.conv.W.shape == value.shape
conv_bn.conv.W.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('bn'):
block_name, branch_name, prm_name = param_name.split('_')
resblock_name = 'res{}'.format(block_name[2])
resblock = getattr(model.extractor, resblock_name)
if block_name[3:] == 'a':
blck_name = block_name[3:]
elif block_name[3:] == 'b':
blck_name = 'b1'
elif block_name[3:].startswith('b'):
blck_name = block_name[3:]
elif block_name[3:] == 'c':
blck_name = 'b2'
block = getattr(resblock, blck_name)
if branch_name == 'branch1':
conv_bn_name = 'residual_conv'
elif branch_name == 'branch2a':
conv_bn_name = 'conv1'
elif branch_name == 'branch2b':
conv_bn_name = 'conv2'
elif branch_name == 'branch2c':
conv_bn_name = 'conv3'
conv_bn = getattr(block, conv_bn_name)
if prm_name == 'beta':
assert conv_bn.bn.beta.shape == value.shape
conv_bn.bn.beta.array[:] = value
finished_keys.append(key)
elif prm_name == 'gamma':
assert conv_bn.bn.gamma.shape == value.shape
conv_bn.bn.gamma.array[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
else:
print('param: {} is not converted'.format(key))
elif param_type == 'aux':
if param_name.endswith('_test'):
continue
elif param_name.startswith('bn_conv1'):
if param_name == 'bn_conv1_moving_mean':
assert model.extractor.conv1.bn.avg_mean.shape \
== value.shape
model.extractor.conv1.bn.avg_mean[:] = value
finished_keys.append(key)
elif param_name == 'bn_conv1_moving_var':
assert model.extractor.conv1.bn.avg_var.shape \
== value.shape
model.extractor.conv1.bn.avg_var[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
elif param_name.startswith('bn'):
block_name, branch_name, _, prm_name = \
param_name.split('_')
resblock_name = 'res{}'.format(block_name[2])
resblock = getattr(model.extractor, resblock_name)
if block_name[3:] == 'a':
blck_name = block_name[3:]
elif block_name[3:] == 'b':
blck_name = 'b1'
elif block_name[3:].startswith('b'):
blck_name = block_name[3:]
elif block_name[3:] == 'c':
blck_name = 'b2'
block = getattr(resblock, blck_name)
if branch_name == 'branch1':
conv_bn_name = 'residual_conv'
elif branch_name == 'branch2a':
conv_bn_name = 'conv1'
elif branch_name == 'branch2b':
conv_bn_name = 'conv2'
elif branch_name == 'branch2c':
conv_bn_name = 'conv3'
conv_bn = getattr(block, conv_bn_name)
if prm_name == 'mean':
assert conv_bn.bn.avg_mean.shape == value.shape
conv_bn.bn.avg_mean[:] = value
finished_keys.append(key)
elif prm_name == 'var':
assert conv_bn.bn.avg_var.shape == value.shape
conv_bn.bn.avg_var[:] = value
finished_keys.append(key)
else:
print('param: {} is not converted'.format(key))
else:
print('param: {} is not converted'.format(key))
else:
print('param: {} is not converted'.format(key))
return model
if __name__ == '__main__':
main()