What version of python and tensorflow this project need?

[jingyuLin1999]

I just download this project, and want to use research/slim, but i got some error. so could you please tell me the version of python and tensorflow it need？

[Jiya873]

The research/slim implementation in this project was built using TensorFlow 1.x, and it was tested with Python 3.6/3.7 along with TensorFlow 1.14–1.15. If you’re running a TensorFlow 2.x environment (or a different Python version), you might run into compatibility issues.

[jingyuLin1999]

Thanks for you reply. I use Python 3.6 and TensorFlow 1.13.1, it still no work. then i download code of 2019 MarchMar, the same environment it work. Thanks again！

another question：i use research/slim/train_image_classifier.py train a model about cifar10, how to use the model?

input_tensor = graph.get_tensor_by_name("per_image_standardization:0") 
output_tensor = graph.get_tensor_by_name("CifarNet/Predictions/Softmax:0") 

per_image_standardization:0 or CifarNet/Predictions/Softmax:0 may not be corrcet?

import tensorflow as tf
import numpy as np
from PIL import Image

# 定义检查点路径
checkpoint_path = r"D:/myCode/models-2019-3-27/tfmodel/cifarnet-model/model.ckpt-100000"
image_path = 'D:/myCode/models-2019-3-27/tfmodel/work-images/bird2.png'  # <- 替换为你想测试的图片

# 加载模型并恢复训练好的权重
with tf.Session() as sess:
    # 1. 加载图结构
    saver = tf.train.import_meta_graph(checkpoint_path + '.meta')

    # 2. 恢复模型权重
    saver.restore(sess, checkpoint_path)

    # 获取默认图
    graph = tf.get_default_graph()

    # 4. 加载图片并进行预处理
    def load_image(image_path):
        img = Image.open(image_path).resize((32, 32))  # 根据模型要求调整图片大小
        img = np.array(img)  # 转换为 NumPy 数组
        if img.shape[-1] == 4:  # 如果是 RGBA 图像，将其转换为 RGB
            img = img[:, :, :3]
        img = img / 255.0  # 图像归一化
        return img  # 返回去掉 batch 维度的图像

    # 5. 加载输入图片
    image = load_image(image_path)

    # 3. 获取输入张量和输出张量
    input_tensor = graph.get_tensor_by_name("per_image_standardization:0")  # 标准化后的图像
    output_tensor = graph.get_tensor_by_name("CifarNet/Predictions/Softmax:0")  # Softmax 输出

    # 6. 使用模型进行预测
    prediction = sess.run(output_tensor, feed_dict={input_tensor: image})

    # 输出预测结果
    print(f"Prediction: {prediction}")

    # 获取预测类别（假设你是分类任务）
    predicted_class = np.argmax(prediction, axis=1)
    print(f"Predicted Class: {predicted_class[0]}")

[Jiya873]

When restoring the CifarNet model and trying to access tensors for inference:

• Input tensor name: image:0
• Output tensor name: CifarNet/Predictions/Reshape_1:0 (instead of CifarNet/Predictions/Softmax:0)

Tensor names can vary depending on the model export process. Use for op in graph.get_operations(): print(op.name) to list all tensor names.

[jingyuLin1999]

i test by your advice, and i got ValueError: Cannot feed value of shape (1, 32, 32, 3) for Tensor 'image:0', which has shape '()'
is really image:0?

my code:

import tensorflow as tf
import numpy as np
from PIL import Image
import os

# ---------------------- 配置路径 ----------------------
checkpoint_path = r"D:/myCode/models-2019-3-27/tfmodel/cifarnet-model/model.ckpt-100000"
image_path = r"D:/myCode/models-2019-3-27/tfmodel/work-images/bird2.png"

# ---------------------- 图片预处理函数 ----------------------
def load_image(image_path):
    try:
        img = Image.open(image_path).resize((32, 32))  # CIFAR图像尺寸
        img = np.array(img)
        if img.ndim == 2:  # 灰度图转 RGB
            img = np.stack([img] * 3, axis=-1)
        if img.shape[-1] == 4:  # RGBA 转 RGB
            img = img[:, :, :3]
        img = img / 255.0  # 归一化
        img = np.expand_dims(img, axis=0)  # 添加 batch 维度 -> (1, 32, 32, 3)
        return img.astype(np.float32)
    except Exception as e:
        print(f"[图片预处理出错] {e}")
        return None

# ---------------------- 加载模型 & 预测 ----------------------
if not os.path.exists(checkpoint_path + ".meta"):
    raise FileNotFoundError("未找到模型文件，请检查路径是否正确")

image = load_image(image_path)
if image is None:
    raise ValueError("图片加载失败")

with tf.Session() as sess:
    # 1. 加载图结构
    saver = tf.train.import_meta_graph(checkpoint_path + '.meta')

    # 2. 恢复模型权重
    saver.restore(sess, checkpoint_path)

    # 3. 获取默认图
    graph = tf.get_default_graph()

    # for op in tf.get_default_graph().get_operations():
    #     for output in op.outputs:
    #         print(f"张量名: {output.name}, shape: {output.shape}")

    # 4. 尝试自动获取输入输出张量名（可打印验证）
    input_tensor = graph.get_tensor_by_name("image:0")  # 输入占位符
    output_tensor = graph.get_tensor_by_name("CifarNet/Predictions/Reshape_1:0")  # 输出 softmax 前的结果

    print(f"[Info] 输入张量 shape: {input_tensor.shape}")
    print(f"[Info] 预测张量 shape: {output_tensor.shape}")

    # 5. 执行预测
    prediction = sess.run(output_tensor, feed_dict={input_tensor: image})

    # 6. 输出结果
    predicted_class = np.argmax(prediction, axis=1)[0]
    print(f"\n预测结果: {prediction}")
    print(f"预测类别: {predicted_class}")

[Jiya873]

The tensor names depend on how the graph was built. For example, if you exported an inference graph with export_inference_graph.py, the input placeholder is usually named "input:0" rather than "per_image_standardization:0". To verify the correct tensor names from your trained model, list the graph operations:

with tf.Session() as sess:
    saver = tf.train.import_meta_graph(checkpoint_path + '.meta')
    saver.restore(sess, checkpoint_path)
    graph = tf.get_default_graph()
    for op in graph.get_operations():
        print(op.name)

This will help you determine the appropriate names to use.

[jingyuLin1999]

ok, thanks