Created by: wawltor

PR types

Bug fixes

PR changes

Others

Describe

fix the argmin,argmax op for the paddlepaddle 2.0

• Registered the AddCheckPoint for the argmax, argmin, add the flatten attribuate for the argmax/argmin, the flatten means that flattening the tensor as input;At the same time, change the default value from -1 to 3 in the attribute of argmax/argmin, but do not change the operator logic.

• Fixed the output of argmax/argmin, because the paddlepaddle do not support the scalar tensor, change the output of argmin/argmax from the scalar tensor to normal tensor with shape [1] .

• Added the check for dtype of argmin/argmax, when the num of element in argmax/argmin is larger the max value of int32, remind the user to use the dtype of int64.

测试改变Argmax, Argmin的属性后是否能进行正常预测且预测值是否为默认dtype

1 . 测试代码一 在1.8版本进行模型的搭建，同时argmax/argmin使用默认dtype属性来保存模型，默认的返回的结果的dtype为int64, 测试代码如下

import paddle
import numpy as np
import paddle.fluid as F
import paddle.fluid.layers as L
np.random.seed(123)
input_var = F.data(name="input_data", shape=[10, 100], dtype="float32")
max_value = L.argmax(input_var, axis=0)
loss = L.reduce_sum(max_value)
exe = F.Executor(F.CUDAPlace(0))
exe.run(F.default_startup_program())


numpy_data = np.random.rand(10, 100).astype("float32")
for i in range(0, 1):
   result, indices = exe.run(feed={"input_data": numpy_data}, fetch_list=[loss, max_value])
print("the result is {}, and the argmax output dtype:{}".format(result, indices.dtype))

F.io.save_inference_model(dirname="./model_path", feeded_var_names=['input_data'], target_vars=[loss, max_value], executor=exe)
print("save the inference model done in paddle version:{}".format(paddle.__version__))

运行且保存模型如下： 2 . 使用paddle-develop版本进行模型的预测，测试代码如下

mport numpy as np
import paddle
import paddle.fluid as F
import paddle.fluid.layers as L

exe = F.Executor(F.CUDAPlace(0))
exe.run(F.default_startup_program())
[inference_program, feed_target_names, fetch_targets] = (F.io.load_inference_model(dirname="./model_path", executor=exe))
print("laod inference model from paddle version:{}".format(paddle.__version__))
np.random.seed(123)
input_np = np.array(np.random.random((10, 100)), dtype=np.float32)
results, indices = exe.run(inference_program,
        feed={feed_target_names[0]: input_np},
        fetch_list=fetch_targets)
print("the result is {}, indices dtype:{}".format(results, indices.dtype))

测试结果正常，跟1.8.4版本的输出结果一致，且argmax/argmin的输出的indices是int64

3 . 测试结论 无兼容性影响