系列文章
【如何训练一个中英翻译模型】LSTM机器翻译seq2seq字符编码（一）
【如何训练一个中英翻译模型】LSTM机器翻译模型训练与保存（二）
【如何训练一个中英翻译模型】LSTM机器翻译模型部署（三）
【如何训练一个中英翻译模型】LSTM机器翻译模型部署之onnx（python）（四）

一、事情准备

这篇是在【如何训练一个中译英翻译器】LSTM机器翻译模型部署之onnx（python）（四）的基础上进行的，要用到文件为：

input_words.txt
target_words.txt
config.json
encoder_model-sim.onnx
decoder_model-sim.onnx

其中的onnx就是用来转为ncnn模型的，这里借助了onnx这个中间商，所以前面我们需要先通过onnxsim对模型进行simplify，要不然在模型转换时会出现op不支持的情况（模型转换不仅有中间商这个例子，目前还可以通过pnnx直接将pytorch模型转为ncnn，感兴趣的小伙伴可以去折腾下）
老规矩，先给出工具：

onnx2ncnn：https://github.com/Tencent/ncnn
netron：https://netron.app

二、模型转换

这里进行onnx转ncnn，通过命令进行转换

onnx2ncnn onnxModel/encoder_model-sim.onnx ncnnModel/encoder_model.param ncnnModel/encoder_model.bin
onnx2ncnn onnxModel/decoder_model-sim.onnx ncnnModel/decoder_model.param ncnnModel/decoder_model.bin

转换成功可以看到：

转换之后可以对模型进行优化，但是奇怪的是，这里优化了不起作用，去不了MemoryData这些没用的op

ncnnoptimize ncnnModel/encoder_model.param ncnnModel/encoder_model.bin ncnnModel/encoder_model.param ncnnModel/encoder_model.bin 1
ncnnoptimize ncnnModel/decoder_model.param ncnnModel/decoder_model.bin ncnnModel/decoder_model.param ncnnModel/decoder_model.bin 1

三、ncnn模型加载与推理（python版）

跟onnx的推理比较类似，就是函数的调用方法有点不同，这里先用python实现，验证下是否没问题，方面后面部署到其它端，比如android。
主要包括：模型加载、推理模型搭建跟模型推理，但要注意的是这里的输入输出名称需要在param这个文件里面获取。

采用netron分别查看encoder与decoder的网络结构，获取输入输出名称:

encoder:
输入输出分别如图

decoder:

输入

输出：

推理代码如下，推理过程感觉没问题，但是推理输出结果相差很大（对比过第一层ncnn与onnx的推理结果了），可能问题出在模型转换环节的精度损失上，而且第二层模型转换后网络输出结果不一致了，很迷，还没找出原因，但是以下的推理是能运行通过，只不过输出结果有问题

import numpy as np
import ncnn# 加载字符
# 从 input_words.txt 文件中读取字符串
with open('config/input_words.txt', 'r') as f:input_words = f.readlines()input_characters = [line.rstrip('\n') for line in input_words]# 从 target_words.txt 文件中读取字符串
with open('config/target_words.txt', 'r', newline='') as f:target_words = [line.strip() for line in f.readlines()]target_characters = [char.replace('\\t', '\t').replace('\\n', '\n') for char in target_words]#字符处理，以方便进行编码
input_token_index = dict([(char, i) for i, char in enumerate(input_characters)])
target_token_index = dict([(char, i) for i, char in enumerate(target_characters)])# something readable.
reverse_input_char_index = dict((i, char) for char, i in input_token_index.items())
reverse_target_char_index = dict((i, char) for char, i in target_token_index.items())
num_encoder_tokens = len(input_characters) # 英文字符数量
num_decoder_tokens = len(target_characters) # 中文文字数量import json
with open('config/config.json', 'r') as file:loaded_data = json.load(file)# 从加载的数据中获取max_encoder_seq_length和max_decoder_seq_length的值
max_encoder_seq_length = loaded_data["max_encoder_seq_length"]
max_decoder_seq_length = loaded_data["max_decoder_seq_length"]# Load the ncnn models for the encoder and decoder
encoderNet = ncnn.Net()
encoderNet.load_param("ncnnModel/encoder_model.param")
encoderNet.load_model("ncnnModel/encoder_model.bin")decoderNet = ncnn.Net()
decoderNet.load_param("ncnnModel/decoder_model.param")
decoderNet.load_model("ncnnModel/decoder_model.bin")def decode_sequence(input_seq):# Encode the input as state vectors.# print(input_seq)ex_encoder = encoderNet.create_extractor()ex_encoder.input("input_1", ncnn.Mat(input_seq))states_value = []_, LSTM_1 = ex_encoder.extract("lstm")_, LSTM_2 = ex_encoder.extract("lstm_1")states_value.append(LSTM_1)states_value.append(LSTM_2)# print(ncnn.Mat(input_seq))# print(vgdgd)# Generate empty target sequence of length 1.target_seq = np.zeros((1, 1, 849))# Populate the first character of target sequence with the start character.target_seq[0, 0, target_token_index['\t']] = 1.# this target_seq you can treat as initial state# Sampling loop for a batch of sequences# (to simplify, here we assume a batch of size 1).stop_condition = Falsedecoded_sentence = ''ex_decoder = decoderNet.create_extractor()while not stop_condition:#print(ncnn.Mat(target_seq))print("---------")ex_decoder.input("input_2", ncnn.Mat(target_seq))ex_decoder.input("input_3", states_value[0])ex_decoder.input("input_4", states_value[1])_, output_tokens = ex_decoder.extract("dense")_, h = ex_decoder.extract("lstm_1")_, c = ex_decoder.extract("lstm_1_1")print(output_tokens)tk = []for i in range(849):tk.append(output_tokens[849*i])tk = np.array(tk)output_tokens = tk.reshape(1,1,849)print(output_tokens)# print(fdgd)print(h)print(c)# output_tokens = np.array(output_tokens)# output_tokens = output_tokens.reshape(1, 1, -1)# # h = np.array(h)# # c = np.array(c)# print(output_tokens.shape)# print(h.shape)# print(c.shape)#output_tokens, h, c = decoder_model.predict([target_seq] + states_value)# Sample a token# argmax: Returns the indices of the maximum values along an axis# just like find the most possible charsampled_token_index = np.argmax(output_tokens[0, -1, :])# find char using indexsampled_char = reverse_target_char_index[sampled_token_index]# and append sentencedecoded_sentence += sampled_char# Exit condition: either hit max length# or find stop character.if (sampled_char == '\n' or len(decoded_sentence) > max_decoder_seq_length):stop_condition = True# Update the target sequence (of length 1).# append then ?# creating another new target_seq# and this time assume sampled_token_index to 1.0target_seq = np.zeros((1, 1, num_decoder_tokens))target_seq[0, 0, sampled_token_index] = 1.print(sampled_token_index)# Update states# update states, frome the front partsstates_value = [h, c]return decoded_sentenceimport numpy as npinput_text = "Call me."
encoder_input_data = np.zeros((1,max_encoder_seq_length, num_encoder_tokens),dtype='float32')
for t, char in enumerate(input_text):print(char)# 3D vector only z-index has char its value equals 1.0encoder_input_data[0,t, input_token_index[char]] = 1.input_seq = encoder_input_data
decoded_sentence = decode_sequence(input_seq)
print('-')
print('Input sentence:', input_text)
print('Decoded sentence:', decoded_sentence)

decoder的模型输出为849*849，感觉怪怪的，然后我们把模型的输入固定下来看看是不是模型的问题。
打开decoder_model.param，把输入层固定下来，0=w 1=h 2=c，那么：
input_2：0=849 1=1 2=1
input_3：0=256 1=1
input_4：0=256 1=1

运行以下命令进行优化

ncnnoptimize ncnnModel/decoder_model.param ncnnModel/decoder_model.bin ncnnModel/decoder_model.param ncnnModel/decoder_model.bin 1

结果如下：

打开网络来看一下：
可以看到输出确实是849849（红色框），那就是模型转换有问题了

仔细看，能够看到有两个shape（蓝色框）分别为849跟8491，这两个不同维度的网络进行BinaryOP之后，就变成849849了，那么，我们把Reshape这个网络去掉试试（不把前面InnerProduct的输入维度有849reshape为8491），下面来看手术刀怎么操作。

我们需要在没经过固定维度并ncnnoptimize的模型上操作（也就是没经过上面0=w 1=h 2=c修改的模型上操作）
根据名字我们找到Reshape那一层：

然后找到与reshape那一层相连接的上一层(红色框)与下一层(蓝色框)

通过红色框与蓝色框里面的名字我们找到了上层与下层分别为InnerProduct与BinaryOp

这时候，把InnerProduct与BinaryOp接上，把Reshape删掉

再改一下最上面的层数，把19改为18，因为我们删掉了一层
保存之后再次执行

ncnnoptimize ncnnModel/decoder_model.param ncnnModel/decoder_model.bin ncnnModel/decoder_model.param ncnnModel/decoder_model.bin 1

执行后可以看到网络层数跟blob数都更新了

这时候改一下固定一下输入层数，并运行ncnnoptimize，再打开netron看一下网络结构，可以看到输出维度正常了

但是通过推理结果还是不对，没找到原因，推理代码如下：

import numpy as np
import ncnn# 加载字符
# 从 input_words.txt 文件中读取字符串
with open('config/input_words.txt', 'r') as f:input_words = f.readlines()input_characters = [line.rstrip('\n') for line in input_words]# 从 target_words.txt 文件中读取字符串
with open('config/target_words.txt', 'r', newline='') as f:target_words = [line.strip() for line in f.readlines()]target_characters = [char.replace('\\t', '\t').replace('\\n', '\n') for char in target_words]#字符处理，以方便进行编码
input_token_index = dict([(char, i) for i, char in enumerate(input_characters)])
target_token_index = dict([(char, i) for i, char in enumerate(target_characters)])# something readable.
reverse_input_char_index = dict((i, char) for char, i in input_token_index.items())
reverse_target_char_index = dict((i, char) for char, i in target_token_index.items())
num_encoder_tokens = len(input_characters) # 英文字符数量
num_decoder_tokens = len(target_characters) # 中文文字数量import json
with open('config/config.json', 'r') as file:loaded_data = json.load(file)# 从加载的数据中获取max_encoder_seq_length和max_decoder_seq_length的值
max_encoder_seq_length = loaded_data["max_encoder_seq_length"]
max_decoder_seq_length = loaded_data["max_decoder_seq_length"]# Load the ncnn models for the encoder and decoder
encoderNet = ncnn.Net()
encoderNet.load_param("ncnnModel/encoder_model.param")
encoderNet.load_model("ncnnModel/encoder_model.bin")decoderNet = ncnn.Net()
decoderNet.load_param("ncnnModel/decoder_model.param")
decoderNet.load_model("ncnnModel/decoder_model.bin")def decode_sequence(input_seq):# Encode the input as state vectors.# print(input_seq)ex_encoder = encoderNet.create_extractor()ex_encoder.input("input_1", ncnn.Mat(input_seq))states_value = []_, LSTM_1 = ex_encoder.extract("lstm")_, LSTM_2 = ex_encoder.extract("lstm_1")states_value.append(LSTM_1)states_value.append(LSTM_2)# print(ncnn.Mat(input_seq))# print(vgdgd)# Generate empty target sequence of length 1.target_seq = np.zeros((1, 1, 849))# Populate the first character of target sequence with the start character.target_seq[0, 0, target_token_index['\t']] = 1.# this target_seq you can treat as initial state# Sampling loop for a batch of sequences# (to simplify, here we assume a batch of size 1).stop_condition = Falsedecoded_sentence = ''ex_decoder = decoderNet.create_extractor()while not stop_condition:#print(ncnn.Mat(target_seq))print("---------")ex_decoder.input("input_2", ncnn.Mat(target_seq))ex_decoder.input("input_3", states_value[0])ex_decoder.input("input_4", states_value[1])_, output_tokens = ex_decoder.extract("dense")_, h = ex_decoder.extract("lstm_1")_, c = ex_decoder.extract("lstm_1_1")print(output_tokens)# print(ghfhf)# tk = []# for i in range(849):#     tk.append(output_tokens[849*i])# tk = np.array(tk)# output_tokens = tk.reshape(1,1,849)# print(output_tokens)# print(fdgd)print(h)print(c)output_tokens = np.array(output_tokens)output_tokens = output_tokens.reshape(1, 1, -1)# # h = np.array(h)# # c = np.array(c)# print(output_tokens.shape)# print(h.shape)# print(c.shape)#output_tokens, h, c = decoder_model.predict([target_seq] + states_value)# Sample a token# argmax: Returns the indices of the maximum values along an axis# just like find the most possible charsampled_token_index = np.argmax(output_tokens[0, -1, :])# find char using indexsampled_char = reverse_target_char_index[sampled_token_index]# and append sentencedecoded_sentence += sampled_char# Exit condition: either hit max length# or find stop character.if (sampled_char == '\n' or len(decoded_sentence) > max_decoder_seq_length):stop_condition = True# Update the target sequence (of length 1).# append then ?# creating another new target_seq# and this time assume sampled_token_index to 1.0target_seq = np.zeros((1, 1, num_decoder_tokens))target_seq[0, 0, sampled_token_index] = 1.print(sampled_token_index)# Update states# update states, frome the front partsstates_value = [h, c]return decoded_sentenceimport numpy as npinput_text = "Call me."
encoder_input_data = np.zeros((1,max_encoder_seq_length, num_encoder_tokens),dtype='float32')
for t, char in enumerate(input_text):print(char)# 3D vector only z-index has char its value equals 1.0encoder_input_data[0,t, input_token_index[char]] = 1.input_seq = encoder_input_data
decoded_sentence = decode_sequence(input_seq)
print('-')
print('Input sentence:', input_text)
print('Decoded sentence:', decoded_sentence)

参考文献：https://github.com/Tencent/ncnn/issues/2586