cauchyturing / ucr_time_series_classification_deep_learning_baseline Goto Github PK

Wonderful job! I studied your paper these days. Your paper proposed a strong baseline for time series classification in UCR. From the results of the paper, FCN and ResNet have achieved high classification accuracy。The ResNet in your paper is made up of 3 blocks, Have you tried more blocks to improve the performance? Thanks for your kind help!

I have run the code twice and I find that the performance of FCN and Resnet is normal but in MLP the performance is very low.
FCN

0.034404267222644426 0.8465473055839539
0.014002826408698009 0.8491048812866211

MLP

2.7152027130126952 0.21994884312152863

RESNET

9.826100923909018e-05, 0.8184143304824829

But in the paper, it is said to be good. Is there something wrong?

Hi, may I ask what is the value reported as the result? Since you stated that it is Mean Per-Class Error.

Is it actually the training loss?

[Time Series Classification from Scratch with Deep Neural Networks: A Strong Baseline] It can be seen from this paper that the author choose the best model that achieves the lowest training loss and report its performance on the test set. I wonder what if it overfits.

Get increasing validation loss for trace dataset.

may i know where to access the dataset used in this baseline? For trace specifically would be enough. Or what preprocessing was used.

Thanks

How do the models perform for multivariate time series classification?
the commented list in the code only contains univariate time series

All images in the readme are broken

How can we use the model developed so far to forecast it woud be helpful if you could give an insight into this.

I am facing this issue after running resnet code, MLP and FCN is working fine. could you please help?

TypeError Traceback (most recent call last)
in ()
138 x_test = x_test.reshape(x_test.shape + (1, 1,))
139
--> 140 x, y = build_resnet(x_train.shape[1:], 64, nb_classes)
141 model = Model(input=x, output=y)
142 optimizer = keras.optimizers.Adam()

in build_resnet(input_shape, n_feature_maps, nb_classes)
42 shortcut_y = keras.layers.normalization.BatchNormalization()(x)
43 print('Merging skip connection')
---> 44 y = merge([shortcut_y, conv_z], mode='sum')
45 y = Activation('relu')(y)
46

TypeError: 'module' object is not callable

To run the code with a new dataset, which is the preferred format for the file to load into the code?

May I ask two questions of your FCN code please?

1. Traditional sequential data is of three dimensions: (batch_size, sequence_length, sequence_dimension). Why the data is four dimension in your code?

2. In your paper, you stated that "the features are fed into a global average pooling layer instead of a fully connected layer." But you implemented the fully connected layer
   out = keras.layers.Dense(nb_classes, activation='softmax')(full). Why is that?

Thank you in advance.

i think the result is too high.

Hi,
Thanks for the nice example. I am learning the codes of MLP and have a question about it. In line 59, code is “y = keras.layers.Dense(500, activation='relu')(x)”. Would ‘(x)’ be ‘(y)’? Since code in line 58 is “y = keras.layers.Dropout(0.1)(x)”

Hi, lately, I applied your baseline model - FCN with keras, but I found the similar problem with this issue. I wonder if this will effect the performance.

Thanks.

great code thanks
may you clarify :
will it work for multivariate time series prediction both regression and classification
1
where all values are continues values
weight height age target
1 56 160 34 1.2
2 77 170 54 3.5
3 87 167 43 0.7
4 55 198 72 0.5
5 88 176 32 2.3

2
or even will it work for multivariate time series where values are mixture of continues and categorical values
for example 2 dimensions have continues values and 3 dimensions are categorical values

color        weight     gender  height  age  target 

1 black 56 m 160 34 yes
2 white 77 f 170 54 no
3 yellow 87 m 167 43 yes
4 white 55 m 198 72 no
5 white 88 f 176 32 yes

Could you please explain what is the reason to rescale Y here?
https://github.com/cauchyturing/UCR_Time_Series_Classification_Deep_Learning_Baseline/blob/master/FCN.py#L41

Hello, I've copied your Adiac data and FCN.py to run with spyder (Python3.7). But it returned a ValueError: Input tensors to a Functional must come from tf.keras.Input. Received: 0 (missing previous layer metadata) when it ran to get the object get_last_conv. Do you know this problem? And do you have any solutions?

While running ResNet.py on cpu, I am getting following error,

AssertionError: AbstractConv2d Theano optimization failed: there is no implementation available supporting the requested options.

Please help