A Wednesday 29 July 2009 10:33:47 Kim Hansen escrigué: > > That is of great help! I had tried parsing a dtype in createTable (not > > accepted), but I never thought of trying an np-ndarray with that > > dtype. If that works, it will solve my problem (will try at work > > tomorrow). > > OK, I have now tried creating a table using a nump.recarray in the > description field, and on close reading of the Library reference sec. > 4.2.3 regarding the createTable method I see that this is indeed > documented there and so is the pydoc you get when you do, e.g., a > h5file.createtable? in an IPython shell. > > What mislead me, was the Tutorial chapter 3 in the User's Guide, which > I followed to get a good understanding. First, in sec. 3.1.5 "Creating > a new table" an example is shown where you use an IsDescription > instance as the description parameter, and later in the more > complicated example on p. 34-35 in sec. 3.4 "Multidimensional table > cells and automatic sanity checks", it is also shown that you can use > a dictionary with XXXCol() instances as the dictionary values. > > >From that example I got the impression that if you had a dtype, the > > path to creating a table of that type would be to first create such a > dictionary or a Description instance translating the dtype values, > i.e., '<f8' to the XXXCol() instances, e.g., Float32Col(), which is > why I started looking for a factory method like > Description.from_dtype(). > > To not lead other numpy users along the same erroneous path (because > the Library reference is not the first chaper you start reading from > beginning to end), I could suggest to mention the recarray createtable > path somewhere in the tutorial section in Sec. 3.1.5 or 3.4, > alternatively write a chapter or an appendix specifically targeted at > the PyTables/numpy interplay, as it seems like there are many nifty > tricks you can use to get your work done very easy and fast if you are > working with numpy arrays together with PyTables. > > Finally, I would also suggest to extend the implementation of > createTable such that the description paramer can also accept a > np.dtype instance. Personally, I append to a table in chunks of, say > 1MB to keep down memory usage (I may have 1GB in a single table), and > here I prefer to first define the table, and then append data (seems > like the cleanest thing to do), so I therefore first use > > h5file.createTable(...,description=np.empty(0, dtype=my_dtype), ...) > > to define the table based on a dtype, which seems kinda awkward as > inside this method you are extracting the dtype from array. Having the > possibility to do > > h5file.createTable(...,description=my_dtype, ...) > > seems more intuitive for me.
Ok. Note taken. The reason why `Table.createTable` is not accepting a NumPy dtype as the description for creating a Table is mainly historical: when PyTables was created back in 2002 it was based in numarray, that had not the notion of a structured type as a such (although they were supported indirectly). When NumPy (stable) appeared in late 2006, the structured dtype was introduced, but we have not supported it in the `Table.createTable` yet mainly because the PyTables way was independent of the underlying numerical package. But now that it seems clear that NumPy is here for stay I think it is a good idea to implement support for it. Will try to do this for forthcoming 2.2 > > Finally, this should not be considered as a rant, just some > observations from a new user. Indeed. Thanks for your feedback; it is very important for us. > My general impression is that the "product" is well very polished and > stable, the user's guide is well-written and the package is very, very > nice (it is an answer to my praisers and the end of a long frustrating > search for a package which could need my needs). Wow, looks like a good quote for: http://pytables.org/moin/UserQuotes :) Would you mind me adding it there? > As mentioned in another post, there is an issue with installing LZO > for windows users and other details, but overall very good. I'll see what can be done about this and come back with a solution. > > I am actually thrilled about PyTables, and I have eagerly demonstrated > its capabilities to my co-workers today (first create an HDF5 and then > browse and filter it using VITables), and they are impressed as well. > I think it is amazing that such a product is available under a free > license, and I would like to thank the creators for taking their time > to make it. You are welcome! -- Francesc Alted ------------------------------------------------------------------------------ Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day trial. Simplify your report design, integration and deployment - and focus on what you do best, core application coding. Discover what's new with Crystal Reports now. http://p.sf.net/sfu/bobj-july _______________________________________________ Pytables-users mailing list Pytables-users@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/pytables-users
