Getting check faild exception when trying to train a simple linear regression model

1

hi,
i’m trying to train a really really simple linear regression with gluon on mxnet but the traceback is as follows:

Traceback (most recent call last):
  File "a.py", line 37, in <module>
    output = model(x)
  File "C:\python36\lib\site-packages\mxnet\gluon\block.py", line 413, in __call__
    return self.forward(*args)
  File "C:\python36\lib\site-packages\mxnet\gluon\block.py", line 629, in forward
    return self.hybrid_forward(ndarray, x, *args, **params)
  File "C:\python36\lib\site-packages\mxnet\gluon\nn\basic_layers.py", line 207, in hybrid_forward
    flatten=self._flatten, name='fwd')
  File "<string>", line 78, in FullyConnected
  File "C:\python36\lib\site-packages\mxnet\_ctypes\ndarray.py", line 92, in _imperative_invoke
    ctypes.byref(out_stypes)))
  File "C:\python36\lib\site-packages\mxnet\base.py", line 149, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [08:59:00] c:\projects\mxnet-distro-win\mxnet-build\src\io\../operator/elemwise_op_common.h:123: Check failed: assign(&dattr, (*vec)[i]) Incompatible attr in node  at 1-th input: expected int64, got float32

this is my code to load and train:

from sklearn.model_selection import *
import mxnet as mx
from mxnet import gluon, nd, autograd, metric
import numpy as np
import csv

x, y=[], []

with open("a.csv", newline="\n") as f:
	r=csv.reader(f)
	for row in r:
		x.append(row[0])
		y.append(row[1])

x=np.array(x, dtype = np.int64).reshape(-1, 1).astype(np.int64)
y=np.array(y, dtype = np.int64).astype(np.int64)

x_train, x_test, y_train, y_test=train_test_split(x, y, test_size=0.20, shuffle=False)

# make our regression model (a Dense layour)
model = gluon.nn.Dense(1)

# initialize and fit
batch_size = 10
epochs = 100
train_dataset = mx.gluon.data.ArrayDataset(x_train, y_train)
train_iter = mx.gluon.data.DataLoader(train_dataset, last_batch = "rollover", batch_size = batch_size)
# initialize
model.collect_params().initialize(mx.initializer.Normal(), ctx=mx.cpu())
trainer = mx.gluon.Trainer(model.collect_params(), "sgd", {"learning_rate":0.0001 } )
# train
for e in range(epochs):
	for i, (x, y) in enumerate(train_iter):
		x.attach_grad()
		y.attach_grad()
		with autograd.record():
			output = model(x)
			loss = mx.gluon.loss.L2Loss(output, y)
			loss.backward()
	trainer.step(x.shape[0])

and, here is how my a.csv looks like:

0,0
1,4
2,8
3,12
4,16
5,20
6,24
7,28
8,32
9,36

thanks.

2

Well, it doesn’t seem to work with int64, and I assume this is because usually matrix multiplication should be done with same data types. Since data type of weight coefficients of the dense layer is float, then there is a type mismatch happening.

Dense layer’s weights doesn’t support ints, so I would just recommend to convert your input from int to float. That what is usually done, when you are working with image data where each pixel is of [0 - 255] range.

I also applied a few clean ups and fixes, so my final version of your code looks like below:

from sklearn.model_selection import *
import mxnet as mx
from mxnet import gluon, nd, autograd, metric
import numpy as np
import csv

x, y = [], []

with open("a.csv", newline="\n") as f:
    r = csv.reader(f)
    for row in r:
        x.append(row[0])
        y.append(row[1])

x = np.array(x, dtype=np.float32).reshape(-1, 1).astype(np.float32)
y = np.array(y, dtype=np.float32).astype(np.float32)

x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.20, shuffle=False)

# make our regression model (a Dense layout)
model = gluon.nn.Dense(1)

# initialize and fit
batch_size = 10
epochs = 100

train_dataset = mx.gluon.data.ArrayDataset(x_train, y_train)
train_iter = mx.gluon.data.DataLoader(train_dataset, last_batch="keep", batch_size=batch_size)

# initialize
model.initialize(mx.initializer.Normal(), ctx=mx.cpu())
trainer = mx.gluon.Trainer(model.collect_params(), "sgd", {"learning_rate": 0.0001})

# train
for e in range(epochs):
    for i, (x, y) in enumerate(train_iter):
        with autograd.record():
            output = model(x)
            loss = mx.gluon.loss.L2Loss()(output, y)
        loss.backward()
    trainer.step(x.shape[0])
1 Like
3

hi,
thanks!.
it had fixed the problem!.