saveload test 10 case (#2953)

* add case

* test wz add case, test=model

* add case

framework/e2e/research/saveload_ten_model/model1.py

@@ -0,0 +1,190 @@
+"""
+Module Docstring:
+This module contains functions for saving and loading models using PaddlePaddle.
+"""
+import os
+import paddle
+import paddle.nn as nn
+from paddle.static import InputSpec
+
+# 引用 paddle inference 预测库
+import paddle.inference as paddle_infer
+import numpy as np
+
+# 生成随机数据
+paddle.seed(33)
+np.random.seed(32)
+
+
+def save_model(model, path, inputs):
+    """
+    paddle.save
+    """
+    # 保存模型
+    model.eval()
+
+    save_model_path = os.path.join(path.split("/")[0], path.split("/")[1], "model.pdparams")
+    # paddle.save(model.state_dict(),save_model_path)
+    # 加载模型的参数
+    model.set_state_dict(paddle.load(save_model_path))
+    pred = model(inputs)
+    output3 = paddle.mean(pred)
+    print("Paddle.Save Model Output Mean: " + str(output3.numpy()))
+    return pred, output3
+
+
+def jit_save(model, path, inputs):
+    """
+    保存:paddle.jit.save
+    """
+    # 设置训练模式
+    model.eval()
+    pred = model(inputs)
+    outputs_mean = paddle.mean(pred)
+    print("Save Model Output Mean: " + str(outputs_mean.numpy()))
+    # 保存模型
+    # paddle.jit.save(model, path)
+
+    return pred, outputs_mean
+
+
+def jit_load(model, path, inputs):
+    """
+    加载:paddle.jit.load
+    """
+    # 加载模型
+    loaded_model = paddle.jit.load(path)
+
+    pred = loaded_model(inputs)
+    output_mean = paddle.mean(pred)
+    print("Loaded Model Output Mean: " + str(output_mean.numpy()))
+
+    return pred, output_mean
+
+
+def infer(num_image, path, inputs):
+    """
+    paddle.inference
+    """
+    # 创建 config
+    model_name_or_path = path.split("/demo")[0]
+    model_prefix = path.split("/")[-1]
+    params_path = os.path.join(model_name_or_path, model_prefix + ".pdiparams")
+    if paddle.framework.use_pir_api():
+        model_file = os.path.join(model_name_or_path, model_prefix + ".json")
+    else:
+        model_file = os.path.join(model_name_or_path, model_prefix + ".pdmodel")
+    # inference_config = paddle.inference.Config(model_path, params_path)
+
+    config = paddle_infer.Config(model_file, params_path)
+    # 根据 config 创建 predictor
+    predictor = paddle_infer.create_predictor(config)
+    # 获取输入的名称
+    input_names = predictor.get_input_names()
+    input_handle = predictor.get_input_handle(input_names[0])
+    # 设置输入
+    input_handle.reshape([num_image, 3, 64, 64])
+    input_handle.copy_from_cpu(inputs)
+    # 运行 predictor
+    predictor.run()
+    # 获取输出
+    output_names = predictor.get_output_names()
+    output_handle = predictor.get_output_handle(output_names[0])
+    pred = output_handle.copy_to_cpu()  # numpy.ndarray 类型
+    outputs = np.mean(pred)
+    print("Infer Output Mean: " + str(np.mean(pred)))
+
+    return pred, outputs
+
+
+def are_close(a, b, tol=1e-5):
+    """
+    判断两个浮点数a和b是否足够接近
+    """
+    # 判断两个浮点数a和b是否足够接近，是则返回True；否则返回False；
+    # a, b: 要比较的两个浮点数或numpy数组；
+    # tol: 容差，默认为1e-9。
+    return np.allclose(a, b, atol=tol)
+
+
+class RandomNet(nn.Layer):
+    """
+    一个简单的神经网络
+    """
+
+    def __init__(self):
+        super(RandomNet, self).__init__()
+        # 定义卷积层和ReLU层
+        self.layers = nn.LayerList(
+            [
+                nn.Conv2D(3, 106, kernel_size=[3, 3], padding=1),
+                nn.BatchNorm2D(num_features=106, momentum=0.9, epsilon=1e-05),
+                nn.ReLU(),
+                nn.MaxPool2D(kernel_size=2, stride=2, padding=0),
+                nn.Conv2D(106, 120, kernel_size=[3, 3], padding=1),
+                nn.ReLU(),
+                nn.Conv2D(120, 87, kernel_size=[3, 3], padding=1),
+                nn.BatchNorm2D(num_features=87, momentum=0.9, epsilon=1e-05),
+                nn.ReLU(),
+                nn.Conv2D(87, 96, kernel_size=[3, 3], padding=1),
+                nn.ReLU(),
+            ]
+        )
+
+        self.fc_layers = nn.LayerList(
+            [
+                nn.Linear(in_features=98304, out_features=312),
+                nn.ReLU(),
+                nn.Linear(in_features=312, out_features=914),
+                nn.ReLU(),
+                nn.Linear(in_features=914, out_features=135),
+                nn.ReLU(),
+                nn.Linear(in_features=135, out_features=273),
+                nn.ReLU(),
+                nn.Linear(in_features=273, out_features=10),
+            ]
+        )
+
+    @paddle.jit.to_static
+    def forward(self, x):
+        """前向传播函数"""
+        for layer in self.layers:
+            x = layer(x)
+            if isinstance(layer, nn.MaxPool2D):
+                # 处理池化层输出，这里不需要额外操作
+                pass
+        x = paddle.reshape(paddle.flatten(x), [x.shape[0], -1])
+
+        for layer1 in self.fc_layers:
+            x = layer1(x)
+        return x
+
+
+if __name__ == "__main__":
+
+    # 转换为Paddle张量
+    num_image = 10
+    size = 64
+    max_conv_layers = 5
+
+    assert (size % (2**max_conv_layers)) == 0, "不能整除，有报错风险！"
+    # assert (size % (2 ** max_conv_layers)) != 0, "不应该能够整除，但却整除了！"
+
+    data = np.random.randn(num_image, 3, size, size).astype("float32")
+    label = np.random.randint(0, 10, (10, 1), dtype="int64")
+    inputs = paddle.to_tensor(data)
+    labels = paddle.to_tensor(label)
+    path = "simple/model1/demo1"  # 路径这里用demo，若改infer中对应需要改
+
+    model = RandomNet()
+    # 这里可以添加训练、保存、加载和预测的逻辑
+    # print(model)
+
+    # 模型保存 加载评估 预测部署
+    pred_0, output_0 = save_model(model, path, inputs)
+    pred_1, output_1 = jit_save(model, path, inputs)
+    pred_2, output_2 = jit_load(model, path, inputs)
+    pred_3, output_3 = infer(num_image, path, data)
+
+    assert output_1 == output_0 and are_close(output_2, output_3), "模型运行失败，可能是数据不一致或计算错误"
+    print("恭喜你,模型运行成功，再接再厉！！")

framework/e2e/research/saveload_ten_model/model10.py

@@ -0,0 +1,169 @@
+"""
+Module Docstring:
+This module contains functions for saving and loading models using PaddlePaddle.
+"""
+import os
+import paddle
+import paddle.nn as nn
+from paddle.static import InputSpec
+
+# 引用 paddle inference 预测库
+import paddle.inference as paddle_infer
+import numpy as np
+
+# 生成随机数据
+paddle.seed(33)
+np.random.seed(32)
+
+
+def save_model(model, path, inputs):
+    """保存模型"""
+    # 保存模型
+    model.eval()
+    save_model_path = os.path.join(path.split("/")[0], path.split("/")[1], "model.pdparams")
+    paddle.save(model.state_dict(), save_model_path)
+    # 加载模型的参数
+    model.set_state_dict(paddle.load(save_model_path))
+    pred = model(inputs)
+    output3 = paddle.mean(pred)
+    print("Paddle.Save Model Output Mean: " + str(output3.numpy()))
+    return pred, output3
+
+
+def jit_save(model, path, inputs):
+    """paddle.jit.save"""
+    # 设置训练模式
+    model.eval()
+    pred = model(inputs)
+    outputs = paddle.mean(pred)
+    print(f"Jit.Save Model Output Mean: {outputs.numpy()}")
+    # 保存模型
+    paddle.jit.save(model, path)
+
+    return pred, outputs
+
+
+def jit_load(model, path, inputs):
+    """paddle.jit.load"""
+    # 加载模型
+    loaded_model = paddle.jit.load(path)
+
+    pred = loaded_model(inputs)
+    output_mean = paddle.mean(pred)
+    print(f"Loaded Model Output Mean: {output_mean.numpy()}")
+
+    return pred, output_mean
+
+
+def infer(num_image, path, inputs):
+    """推理"""
+    # 创建 config
+    model_name_or_path = path.split("/demo")[0]
+    model_prefix = path.split("/")[-1]
+    params_path = os.path.join(model_name_or_path, model_prefix + ".pdiparams")
+    if paddle.framework.use_pir_api():
+        model_file = os.path.join(model_name_or_path, model_prefix + ".json")
+    else:
+        model_file = os.path.join(model_name_or_path, model_prefix + ".pdmodel")
+    # inference_config = paddle.inference.Config(model_path, params_path)
+
+    config = paddle_infer.Config(model_file, params_path)
+    # 根据 config 创建 predictor
+    predictor = paddle_infer.create_predictor(config)
+    # 获取输入的名称
+    input_names = predictor.get_input_names()
+    input_handle = predictor.get_input_handle(input_names[0])
+    # 设置输入  这里的64是输入图片的尺寸 需要和前面保持一致的
+    input_handle.reshape([num_image, 3, 64, 64])
+    input_handle.copy_from_cpu(inputs)
+    # 运行 predictor
+    predictor.run()
+    # 获取输出
+    output_names = predictor.get_output_names()
+    output_handle = predictor.get_output_handle(output_names[0])
+    pred = output_handle.copy_to_cpu()  # numpy.ndarray 类型
+    outputs = np.mean(pred)
+    print(f"Infer Output Mean: {outputs}")
+
+    return pred, outputs
+
+
+def are_close(a, b, tol=1e-3):
+    """
+    判断两个浮点数是否足够接近。
+
+    a, b: 要比较的两个浮点数或numpy数组。
+    tol: 容差，默认为1e-9。
+
+    如果a和b足够接近，则返回True；否则返回False。
+    """
+    return np.allclose(a, b, atol=tol)
+
+
+class RandomNet(nn.Layer):
+    """一个简单的神经网络"""
+
+    def __init__(self):
+        super(RandomNet, self).__init__()
+        # 定义卷积层和ReLU层
+        self.layers = nn.LayerList(
+            [
+                nn.Conv2D(3, 76, kernel_size=[3, 3], padding=1),
+                nn.ReLU(),
+                nn.Conv2D(76, 109, kernel_size=[3, 3], padding=1),
+                nn.ReLU(),
+                nn.MaxPool2D(kernel_size=2, stride=2, padding=0),
+                nn.Conv2D(109, 113, kernel_size=[3, 3], padding=1),
+                nn.ReLU(),
+                nn.MaxPool2D(kernel_size=2, stride=2, padding=0),
+                nn.Conv2D(113, 93, kernel_size=[3, 3], padding=1),
+                nn.BatchNorm2D(num_features=93, momentum=0.9, epsilon=1e-05),
+                nn.ReLU(),
+            ]
+        )
+
+        self.fc_layers = nn.LayerList([nn.Linear(in_features=23808, out_features=10)])
+
+    @paddle.jit.to_static
+    def forward(self, x):
+        """前向传播"""
+        for layer in self.layers:
+            x = layer(x)
+            if isinstance(layer, nn.MaxPool2D):
+                # 处理池化层输出，这里不需要额外操作
+                pass
+        x = paddle.reshape(paddle.flatten(x), [x.shape[0], -1])
+
+        for layer1 in self.fc_layers:
+            x = layer1(x)
+        return x
+
+
+if __name__ == "__main__":
+
+    # 转换为Paddle张量
+    num_image = 10
+    size = 64
+    max_conv_layers = 5
+
+    assert (size % (2**max_conv_layers)) == 0, "不能整除，有报错风险！"
+    # assert (size % (2 ** max_conv_layers)) != 0, "不应该能够整除，但却整除了！"
+
+    data = np.random.randn(num_image, 3, size, size).astype("float32")
+    label = np.random.randint(0, 10, (10, 1), dtype="int64")
+    inputs = paddle.to_tensor(data)
+    labels = paddle.to_tensor(label)
+    path = "simple/model10/demo10"  # 路径这里用demo，若改infer中对应需要改
+
+    model = RandomNet()
+    # 这里可以添加训练、保存、加载和预测的逻辑
+    # print(model)
+
+    # 模型保存 加载评估 预测部署
+    pred_0, output_0 = save_model(model, path, inputs)
+    pred_1, output_1 = jit_save(model, path, inputs)
+    pred_2, output_2 = jit_load(model, path, inputs)
+    pred_3, output_3 = infer(num_image, path, data)
+
+    assert output_1 == output_0 and are_close(output_2, output_3), "模型运行失败，可能是数据不一致或计算错误"
+    print("恭喜你,模型运行成功，再接再厉！！")