这个世界总是有各种各样的幺蛾子，所以我们要做各种各样的转换，就像今天要写的pytorch模型需要被转换成tflite。下面就以pytorch-ssd模型为例，做一次pytorch转tflite的实践。
- pth模型转换成onnx
第一步把torch.save()存下的模型转换成onnx模型，代码如下
import torch
from vision.ssd.mobilenet_v3_ssd_lite import create_mobilenetv3_ssd_lite
model = create_mobilenetv3_ssd_lite(num_classes=2)
torch.load("CARN_model_checkpoint.pt",map_location='cpu')['state_dict'].items()},False)
model.load_state_dict(torch.load("Epoch-85-Loss-0.4889--Epoch-45-Loss-0.4090.pth",map_location='cpu'))
dummy_input = torch.randn(1,3,300,300)
input_names = ["input"]
output_names = ["output"]
torch.onnx.export(model, dummy_input, "ssd_Epoch-45.onnx", verbose=True, input_names=input_names, output_names=output_names,opset_version=11)
- onnx转换成tensorflow pb模型
第二步把onnx模型转换成tensorflow pb模型
git clone https://github.com/onnx/onnx-tensorflow.git
cd onnx-tensorflow
git checkout v1.6.0-tf-1.15
pip install -e .
onnx-tf convert -i /path/to/input.onnx -o /path/to/output.pb
通过第二步操作就生成了pb模型。
- 把nchw格式pb模型转换成nhwc格式pb模型
因为pth和onnx模型都是nchw的layout，转换成pb之后layout没有变，而tflite和tensorflow模型是nhwc的layout格式的，所以需要再增加一步转换，把nchw格式pb模型转换成nhwc格式pb模型，其实原理就是增加tranpose算子，代码如下：
import tensorflow as tf
if not tf.__version__.startswith('1'):
import tensorflow.compat.v1 as tf
from tensorflow.python.tools import optimize_for_inference_lib
graph_def_file = "..\output.pb"
tf.reset_default_graph()
graph_def = tf.GraphDef()
with tf.Session() as sess:
# Read binary pb graph from file
with tf.gfile.Open(graph_def_file, "rb") as f:
data2read = f.read()
graph_def.ParseFromString(data2read)
tf.graph_util.import_graph_def(graph_def, name='')
# Get Nodes
conv_nodes = [n for n in sess.graph.get_operations() if n.type in ['Conv2D','MaxPool','AvgPool']]
for n_org in conv_nodes:
# Transpose input
assert len(n_org.inputs)==1 or len(n_org.inputs)==2
org_inp_tens = sess.graph.get_tensor_by_name(n_org.inputs[0].name)
inp_tens = tf.transpose(org_inp_tens, [0, 2, 3, 1], name=n_org.name +'_transp_input')
op_inputs = [inp_tens]
# Get filters for Conv but don't transpose
if n_org.type == 'Conv2D':
filter_tens = sess.graph.get_tensor_by_name(n_org.inputs[1].name)
op_inputs.append(filter_tens)
# Attributes without data_format, NWHC is default
atts = {key:n_org.node_def.attr[key] for key in list(n_org.node_def.attr.keys()) if key != 'data_format'}
if n_org.type in['MaxPool', 'AvgPool','Conv2D']:
st = atts['strides'].list.i
stl = [st[0], st[2], st[3], st[1]]
atts['strides'] = tf.AttrValue(list=tf.AttrValue.ListValue(i=stl))
if n_org.type in ['MaxPool', 'AvgPool']:
st = atts['ksize'].list.i
stl = [st[0], st[2], st[3], st[1]]
atts['ksize'] = tf.AttrValue(list=tf.AttrValue.ListValue(i=stl))
# Create new Operation
#print(n_org.type, n_org.name, list(n_org.inputs), n_org.node_def.attr['data_format'])
op = sess.graph.create_op(op_type=n_org.type, inputs=op_inputs, name=n_org.name+'_new', dtypes=[tf.float32], attrs=atts)
out_tens = sess.graph.get_tensor_by_name(n_org.name+'_new'+':0')
out_trans = tf.transpose(out_tens, [0, 3, 1, 2], name=n_org.name +'_transp_out')
assert out_trans.shape == sess.graph.get_tensor_by_name(n_org.name+':0').shape
# Update Connections
out_nodes = [n for n in sess.graph.get_operations() if n_org.outputs[0] in n.inputs]
for out in out_nodes:
for j, nam in enumerate(out.inputs):
if n_org.outputs[0] == nam:
out._update_input(j, out_trans)
# Delete old nodes
graph_def = sess.graph.as_graph_def()
for on in conv_nodes:
graph_def.node.remove(on.node_def)
# Write graph
tf.io.write_graph(graph_def, "", graph_def_file.rsplit('.', 1)[0]+'_toco.pb', as_text=False)
第三步后会生成output_toco.pb模型，即为nhwc格式的pb模型。
- 把nhwc格式的pb模型转换成tflite模型
通过tensorflow转换工具把nhwc格式的pb模型转换成tflite模型
tflite_convert.exe --graph_def_file=output_toco.pb --output_file=ssd.tflite --input_arrays=input --output_arrays=output,1099
此时就生成了ssd.tflite模型，之后可用tflite进行前向推理。
