search

mindspore-ai / mindspore

MindSpore is a new open source deep learning training/inference framework that could be used for mobile, edge and cloud scenarios.
https://gitee.com/mindspore/mindspore
Apache License 2.0
4.3k stars 711 forks source link

Mindspore.nn.Conv2d lacks checks on convolutional kernel shapes #291

Open PhyllisJi opened 5 months ago

PhyllisJi commented 5 months ago

Software Environment:

Describe the current behavior

This layer works fine even with incorrect parameter settings in model-implemented code, and crashes if run on its own.

Describe the expected behavior

The convolution kernel size needs to be checked and inappropriate parameter settings need to be thrown as an exception in advance.

Steps to reproduce the issue

import mindspore
import mindspore.nn as nn
import mindspore.ops as ops
import numpy as np

# Define the Conv2dTranspose layer with specified parameters
conv_transpose = nn.Conv2dTranspose(
    in_channels=8, 
    out_channels=8, 
    kernel_size=(1, 1), 
    stride=(1, 1), 
    pad_mode="pad", 
    padding=(7, 7, 7, 7), 
    output_padding=(0, 0), 
    dilation=(1, 1), 
    group=1, 
    has_bias=True
)

# Create an input tensor with shape (1, 8, 14, 14)
input_tensor = mindspore.Tensor(np.random.randn(1, 8, 14, 14), mindspore.float32)

# Apply the Conv2dTranspose layer to the input tensor
output_tensor = conv_transpose(input_tensor)

# Print the shape of the output tensor
print("Output shape:", output_tensor.shape)
import mindspore
import numpy as np
import os

mindspore.context.set_context(device_target='CPU')

class Model_A5cKhrbOS9qjJeV7sRO4l2ukdpmxk6kx(mindspore.nn.Cell):
    def __init__(self):
        super().__init__()
        self.conv1 = mindspore.nn.Conv2d(in_channels=3, out_channels=3, kernel_size=(3, 3), stride=(2, 2), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.conv2 = mindspore.nn.Conv2d(in_channels=3, out_channels=3, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=3, has_bias=True, data_format="NCHW")
        self.relu1 = mindspore.nn.ReLU()
        self.conv3 = mindspore.nn.Conv2d(in_channels=3, out_channels=4, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu2 = mindspore.nn.ReLU()
        self.conv4 = mindspore.nn.Conv2d(in_channels=4, out_channels=4, kernel_size=(3, 3), stride=(2, 2), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=4, has_bias=True, data_format="NCHW")
        self.relu3 = mindspore.nn.ReLU()
        self.conv5 = mindspore.nn.Conv2d(in_channels=4, out_channels=5, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu4 = mindspore.nn.ReLU()
        self.conv6 = mindspore.nn.Conv2d(in_channels=5, out_channels=5, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=5, has_bias=True, data_format="NCHW")
        self.relu5 = mindspore.nn.ReLU()
        self.conv7 = mindspore.nn.Conv2d(in_channels=5, out_channels=6, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu6 = mindspore.nn.ReLU()
        self.conv8 = mindspore.nn.Conv2d(in_channels=6, out_channels=6, kernel_size=(3, 3), stride=(2, 2), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=6, has_bias=True, data_format="NCHW")
        self.relu7 = mindspore.nn.ReLU()
        self.conv9 = mindspore.nn.Conv2d(in_channels=6, out_channels=7, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu8 = mindspore.nn.ReLU()
        self.conv10 = mindspore.nn.Conv2d(in_channels=7, out_channels=7, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=7, has_bias=True, data_format="NCHW")
        self.relu9 = mindspore.nn.ReLU()
        self.conv11 = mindspore.nn.Conv2d(in_channels=7, out_channels=7, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu10 = mindspore.nn.ReLU()
        self.conv12 = mindspore.nn.Conv2d(in_channels=7, out_channels=7, kernel_size=(3, 3), stride=(2, 2), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=7, has_bias=True, data_format="NCHW")
        self.relu11 = mindspore.nn.ReLU()
        self.conv13 = mindspore.nn.Conv2d(in_channels=7, out_channels=8, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu12 = mindspore.nn.ReLU()
        self.conv14 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=8, has_bias=True, data_format="NCHW")
        self.relu13 = mindspore.nn.ReLU()
        self.conv15 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu14 = mindspore.nn.ReLU()
        self.conv16 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=8, has_bias=True, data_format="NCHW")
        self.relu15 = mindspore.nn.ReLU()
        self.conv17 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu16 = mindspore.nn.ReLU()
        self.conv18 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=8, has_bias=True, data_format="NCHW")
        self.relu17 = mindspore.nn.ReLU()
        self.conv19 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(0, 0, 0, 0), dilation=(1, 1), group=1, has_bias=True, data_format="NCHW")
        self.relu18 = mindspore.nn.ReLU()
        self.conv20 = mindspore.nn.Conv2d(in_channels=8, out_channels=8, kernel_size=(3, 3), stride=(1, 1), pad_mode="pad", padding=(1, 1, 1, 1), dilation=(1, 1), group=8, has_bias=True, data_format="NCHW")
        self.relu19 = mindspore.nn.ReLU()
        self.conv21_mutated = mindspore.nn.Conv2dTranspose(in_channels=8, out_channels=8, kernel_size=(1, 1), stride=(1, 1), pad_mode="pad", padding=(7, 7, 7, 7), output_padding=(0, 0), dilation=(1, 1), group=1, has_bias=True)

    def construct(self, input):
        conv1_output = self.conv1(input)
        conv2_output = self.conv2(conv1_output)
        relu1_output = self.relu1(conv2_output)
        conv3_output = self.conv3(relu1_output)
        relu2_output = self.relu2(conv3_output)
        conv4_output = self.conv4(relu2_output)
        relu3_output = self.relu3(conv4_output)
        conv5_output = self.conv5(relu3_output)
        relu4_output = self.relu4(conv5_output)
        conv6_output = self.conv6(relu4_output)
        relu5_output = self.relu5(conv6_output)
        conv7_output = self.conv7(relu5_output)
        relu6_output = self.relu6(conv7_output)
        conv8_output = self.conv8(relu6_output)
        relu7_output = self.relu7(conv8_output)
        conv9_output = self.conv9(relu7_output)
        relu8_output = self.relu8(conv9_output)
        conv10_output = self.conv10(relu8_output)
        relu9_output = self.relu9(conv10_output)
        conv11_output = self.conv11(relu9_output)
        relu10_output = self.relu10(conv11_output)
        conv12_output = self.conv12(relu10_output)
        relu11_output = self.relu11(conv12_output)
        conv13_output = self.conv13(relu11_output)
        relu12_output = self.relu12(conv13_output)
        conv14_output = self.conv14(relu12_output)
        relu13_output = self.relu13(conv14_output)
        conv15_output = self.conv15(relu13_output)
        relu14_output = self.relu14(conv15_output)
        conv16_output = self.conv16(relu14_output)
        relu15_output = self.relu15(conv16_output)
        conv17_output = self.conv17(relu15_output)
        relu16_output = self.relu16(conv17_output)
        conv18_output = self.conv18(relu16_output)
        relu17_output = self.relu17(conv18_output)
        conv19_output = self.conv19(relu17_output)
        relu18_output = self.relu18(conv19_output)
        conv20_output = self.conv20(relu18_output)
        relu19_output = self.relu19(conv20_output)
        print(relu19_output.shape)
        conv21_output = self.conv21_mutated(relu19_output)
        print(conv21_output.shape)
        fc_output = conv21_output
        return fc_output

def go():
    model = Model_A5cKhrbOS9qjJeV7sRO4l2ukdpmxk6kx()
    x = mindspore.Tensor(np.random.randn(1, 3, 224, 224).astype(np.float32))
    y = model(x)
    flag = True

def train(inp, label):
    ms_model = Model_A5cKhrbOS9qjJeV7sRO4l2ukdpmxk6kx()
    initialize(ms_model)
    ms_input = mindspore.Tensor(inp.astype(np.float32))
    def forward_fn(label):
        ms_output = ms_model(ms_input)
        label = label.astype(np.int32)
        ms_targets = mindspore.Tensor(label)
        loss = mindspore.nn.CrossEntropyLoss(reduction='mean')(ms_output, ms_targets)
        return loss, ms_output

    (ms_loss, ms_output), ms_gradients = mindspore.value_and_grad(forward_fn, None, ms_model.trainable_params(), has_aux=True)(label)
    ms_gradients_dic = {}
    for var, gradient in zip(ms_model.trainable_params(), ms_gradients):
        ms_gradients_dic.setdefault(var.name, gradient.numpy())
    return ms_gradients_dic, ms_loss.numpy().item(), ms_output.numpy()

def initialize(model):
    module_dir = os.path.dirname(__file__)
    for name, param in model.parameters_and_names():
        layer_name, matrix_name = name.rsplit('.', 1)
        matrix_path = module_dir + '/../initializer/' + layer_name + '/' + matrix_name + '.npz'
        data = np.load(matrix_path)
        data = data['matrix']
        weight_tensor = mindspore.Tensor(data).float()
        param.set_data(weight_tensor)

go()

Related log / screenshot

Traceback (most recent call last):
  File "/mnt/AA_MoCoDiff/BR_MoCoDiff/result/extract/test.py", line 143, in <module>
    output_tensor = conv_transpose(input_tensor)
  File "/root/miniconda3/envs/mocodiff/lib/python3.9/site-packages/mindspore/nn/cell.py", line 705, in __call__
    raise err
  File "/root/miniconda3/envs/mocodiff/lib/python3.9/site-packages/mindspore/nn/cell.py", line 702, in __call__
    _pynative_executor.end_graph(self, output, *args, **kwargs)
  File "/root/miniconda3/envs/mocodiff/lib/python3.9/site-packages/mindspore/common/api.py", line 1215, in end_graph
    self._executor.end_graph(obj, output, *args, *(kwargs.values()))
RuntimeError: 
----------------------------------------------------
- cuDNN Error:
----------------------------------------------------
cudnnSetTensorNdDescriptor failed | Error Number: 3 CUDNN_STATUS_BAD_PARAM

----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/ccsrc/plugin/device/gpu/kernel/nn/convolution/conv2d_cudnn_gpu_kernel.h:574 Set4DDesc

Special notes for this issue

VectorSL commented 1 month ago

same as issue228

PhyllisJi commented 1 week ago

same as issue228

MindSpore 2.2.14