AifuHan

I solve this problem by using res101 pre-weight replace vgg16, good luck!

> no I do not think so... From the source code, I think if changing the prediction length, one must retrain the model. because the input to rnn must be...

Right, I have reproduced the code and found that the input_size=taget_dim+target_dim*len(lags_seq)+len(time_feat). For example, in the solar dataset, target=137, len(lags_seq)=3, and len(time_feat)=4, so the input_size=137+137x3+4=552. It's the same for traffic_nips. Good...

Please, did you solve your problem? I have the exact same problem with the solar dataset.

> I believe with solar since the values in the dataset are somewhat discrete you will need to set the `dequantize=True,` in the estimator... can you kindly try? @kashif Actually,...

> I believe with solar since the values in the dataset are somewhat discrete you will need to set the `dequantize=True,` in the estimator... can you kindly try? Hi, @kashif...

> @hanlaoshi so as far as I know the `make_evaluation_predictions` required a gluonts prediction object which obviously is not possible here... so the next best would be to obtain the...

> BTW can you redo with the bug fix in the blog #1558 i.e. use the `validation` splitter for the backtesting scenario Hey there! Thanks a bunch for your suggestions....

> so the test-splitter is more for the production use-case in the sense it just takes the very last context window and starts predicting into the unknown future... while the...

Excuse me， Dr. Kashif, I noticed in your publicly available code that the following code is used: `dataset = get_dataset("traffic_nips", regenerate=True) ` The training time steps for the "traffic_nips" dataset...