hecate | Automagically generate thumbnails , animated GIFs | Video Utils library

This is a sample of one line of the OP's XML file
109555912.69

The OP wants to use the IMPORTXML function to report the ID and Name as well as the Job Cost from the XML data. Presently, the OP's formula is:
=importxml("link","//job-base-cost")

There are two options:
1 - One long column
=importxml("link","//@id | //@name | //job-base-cost")

Note //@id and //@name in the xpath query: // indicate nodes in the document (at any level, not just the root level) and @ indicate attributes. The pipe | operator indicates AND. So the plain english query is to display the id, name and job-base-cost.

2 - Three columns (table format)
={IMPORTXML("link","//@name"),IMPORTXML("link","//job-base-cost"),IMPORTXML("link","//@id")}

This creates a series that will display the fields in each of three columns.

Note: there is an arrayformula that uses a single importXML function described in How do I return multiple columns of data using ImportXML in Google Spreadsheets?. Readers may want to look at whether that option can be implemented.

My thanks to @Tanaike for his comment which spurred me to look at how xpath works.

Try implementing this by using a @property declared in the class extension instead of an ivar:

I want to be clear with the OP that this answer does not represent a formal description of the approach so as to describe the prospective approach in an intuitive way.

Suppose that a video is composed by n frames and that each of them can be represented as a 3D tensor (height, width, channel). It is possible to use a Convolutional Neural Networks (CNNs) to generate a latent representation for each frame.

A video can be represented as a sequence of frames (f_1, f_2, ..., f_n). The most suitable neural network architecture for sequence modeling are Recurrent Neural Networks (RNNs). We can use RNNs to encode a sequence of video frames latent representations generated by a CNN. After that, you will have a latent representations (f_1, f_2, ..., f_n) for each frame generated by the RNN which directly depends from the previous ones (this is a well-known property of RNNs).

As you can see in the recently released Youtube-8M dataset, there are high quality thumbnails associated to each video so you can use them as targets. In particular, given the latent representations generated by the RNNs applied on the sequence of frames, you can generate a context vector c which is generated as follows: