如何为神经网络建立一个最小且可复制的示例? [英] How do I make a minimal and reproducible example for neural networks?

问题描述

我想知道如何为Stack Overflow制作一个最小且可重现的深度学习示例.我想确保人们有足够的信息来查明我的代码的确切问题.仅提供回溯就足够了吗?

I would like to know how to make a minimal and reproducible deep learning example for Stack Overflow. I want to make sure that people have enough information to pinpoint the exact problem with my code. Is it enough to just provide the traceback?

    c:\users\samuel\appdata\local\programs\python\python35\lib\site-packages\keras\engine\training_utils.py 
                         in standardize_input_data(data, names, shapes, check_batch_axis, exception_prefix)
        135                             ': expected ' + names[i] + ' to have shape ' +
        136                             str(shape) + ' but got array with shape ' +
    --> 137                             str(data_shape))
        138     return data
        139 

还是我应该简单地发布错误消息?

Or should I simply post the error message?

值错误:检查输入时出错:预期density_1_input具有形状 (4，)但形状为(1，)

Value Error: Error when checking input: expected dense_1_input to have shape (4,) but got array with shape (1,)

推荐答案

以下是制作可重复的，最少的深度学习示例的一些技巧.无论是Keras，Pytorch还是Tensorflow，这都是一个很好的建议.

Here are a few tips to make a reproducible, minimal deep learning Example. It's good advice whether it be for Keras, Pytorch, or Tensorflow.

• 我们无法使用您的数据，但是在大多数情况下，这并不重要.我们需要的只是正确的形状.
  • 使用随机生成的正确形状的数字.
    • 例如np.random.randint(0, 256, (100, 30, 30, 3)用于100张 30x30
    尺寸的彩色图片
    • 例如np.random.choice(np.arange(10), 100)用于10个类别的100个样本
    • We can't use your data, but in most cases, it doesn't matter. All we need is the right shape.
      • Use randomly generated numbers of the right shape.
        • E.g., np.random.randint(0, 256, (100, 30, 30, 3) for 100 colored pictures of size 30x30
        • E.g., np.random.choice(np.arange(10), 100) for 100 samples of 10 categories
        • 仅提供运行代码的最低要求.
        • 包括追溯.最有可能指出确切的问题.
        • Include the traceback. It will most likely point out the exact problem.
        • 至少总是提供输入形状.
        • 发布整个神经网络架构.
        • 包括您的库导入.定义所有变量.
        • Post your entire neural network architecture.
        • Include your library imports. Define all variables.

        以下是完美最小且可重现的示例:

        Here is an example of a perfect minimal and reproducible example:

        我有一个错误.运行此代码时，它给了我这个错误:"

        ValueError:检查目标时出错:预期density_2具有形状(10，)但形状为数组

        ValueError: Error when checking target: expected dense_2 to have shape (10,) but got array with shape

        这是我的体系结构，带有生成的数据:"

        import numpy as np
        import keras
        from keras.models import Sequential
        from keras.layers import Dense, Flatten, Conv2D, MaxPooling2D
        
        xtrain, xtest = np.random.rand(2, 1000, 30, 30, 3)
        ytrain, ytest = np.random.choice(np.arange(10), 2000).reshape(2, 1000) 
        
        model = Sequential([
            Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=xtrain.shape[1:]),
            Conv2D(64, (3, 3), activation='relu'),
            MaxPooling2D(pool_size=(2, 2)),
            Flatten(),
            Dense(128, activation='relu'),
            Dense(10, activation='softmax')])
        
        model.compile(loss=keras.losses.categorical_crossentropy,
                      optimizer=keras.optimizers.Adam(),
                      metrics=['accuracy'])
        
        model.fit(xtrain, ytrain,
                  batch_size=16,
                  epochs=10,
                  validation_data=(xtest, ytest))
        

        这篇关于如何为神经网络建立一个最小且可复制的示例?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！