Julia中使用Flux.train!训练神经网络出现params未定义错误的解决方法

问题描述 #

我编写了以下用于在Julia中训练神经网络的代码，但在尝试使用Flux.train!时遇到错误。

完整代码 #

using Flux, CSV, DataFrames, Random, Statistics, Plots

# Load the data
data = CSV.read("daily_stock_returns.csv")

# Split the data into training and testing sets
train_data = data[1:800, :]
test_data = data[801:end, :]

# Define the input and output variables
inputs = Matrix(train_data[:, 2:end])
outputs = Matrix(train_data[:, 1])

# Define the neural network architecture
n_inputs = size(inputs, 2)
n_hidden = 10
n_outputs = 1
model = Chain(
    Dense(n_inputs, n_hidden, relu),
    Dense(n_hidden, n_outputs)
)

# Define the loss function
loss(x, y) = Flux.mse(model(x), y)

# Define the optimizer
optimizer = ADAM()

# Train the model
n_epochs = 100
for epoch in 1:n_epochs
    Flux.train!(loss, params(model), [(inputs, outputs)], optimizer)
end

# Test the model
test_inputs = Matrix(test_data[:, 2:end])
test_outputs = Matrix(test_data[:, 1])
predictions = Flux.predict(model, test_inputs)
test_loss = Flux.mse(predictions, test_outputs)

# Print the test loss
println("Test loss: $test_loss")

错误详情 #

运行训练循环代码段时：

# Train the model
n_epochs = 100
for epoch in 1:n_epochs
    Flux.train!(loss, params(model), [(inputs, outputs)], optimizer)
end

触发以下错误：

UndefVarError: params not defined

Stacktrace:
 [1] top-level scope
   @ .\In[16]:4
 [2] eval
   @ .\boot.jl:368 [inlined]
 [3] include_string(mapexpr::typeof(REPL.softscope), mod::Module, code::String, filename::String)
   @ Base .\loading.jl:1428

我已尝试卸载并重新安装相关包，但问题未解决。请问该如何修复？

解决方案 #

这个错误源于Flux 0.13+版本已移除params()函数，改用了更简洁的训练API。以下是两种修复方式：

方式1：使用新版Flux训练API（推荐） #

新版Flux.train!无需手动传入参数集合，直接传入模型、数据迭代器和优化器即可。同时推荐用DataLoader做批次处理，提升训练效率：

# 替换原训练循环代码
data_loader = Flux.DataLoader((inputs, outputs), batchsize=32, shuffle=true)
for epoch in 1:n_epochs
    Flux.train!(loss, model, data_loader, optimizer)
end

方式2：兼容旧版API（不推荐，会触发废弃警告） #

如果需要保留旧写法，需显式导入Flux.params：

# 在开头的using语句后添加
using Flux: params

# 原训练代码可保留，但会收到废弃提示
for epoch in 1:n_epochs
    Flux.train!(loss, params(model), [(inputs, outputs)], optimizer)
end

额外修复：预测函数问题 #

你的代码中Flux.predict(model, test_inputs)也会报错，新版Flux中直接调用模型即可完成预测：

# 替换预测代码
predictions = model(test_inputs)

完整修正后的代码示例 #

using Flux, CSV, DataFrames, Random, Statistics, Plots

# Load the data（新版CSV.read需指定DataFrame类型）
data = CSV.read("daily_stock_returns.csv", DataFrame)

# Split the data into training and testing sets
train_data = data[1:800, :]
test_data = data[801:end, :]

# Define the input and output variables
inputs = Matrix(train_data[:, 2:end])
outputs = Matrix(train_data[:, 1])

# Define the neural network architecture
n_inputs = size(inputs, 2)
n_hidden = 10
n_outputs = 1
model = Chain(
    Dense(n_inputs, n_hidden, relu),
    Dense(n_hidden, n_outputs)
)

# Define the loss function
loss(x, y) = Flux.mse(model(x), y)

# Define the optimizer
optimizer = ADAM()

# Train the model with DataLoader
n_epochs = 100
data_loader = Flux.DataLoader((inputs, outputs), batchsize=32, shuffle=true)
for epoch in 1:n_epochs
    Flux.train!(loss, model, data_loader, optimizer)
    # 可选：打印每轮训练损失
    train_loss = loss(inputs, outputs)
    println("Epoch $epoch, Train Loss: $train_loss")
end

# Test the model
test_inputs = Matrix(test_data[:, 2:end])
test_outputs = Matrix(test_data[:, 1])
predictions = model(test_inputs)
test_loss = Flux.mse(predictions, test_outputs)

# Print the test loss
println("Test loss: $test_loss")

内容的提问来源于stack exchange，提问作者Dammon