I'm studying the multi-layer perceptron algorithm and I'm translating python code to golang.
I have 2 matrices. Let's call this matrix M1:
[[0 0 1 1]
[0 1 0 1]]
Let's call this matrix M2:
[[ 0.00041597 0.02185088 -0.00362142]
[-0.00057384 -0.02866677 0.00488404]
[-0.00056316 -0.02705587 0.00410378]
[ 0.00048268 0.01692128 -0.00262183]]
I'm implementing the dotProduct(M1,M2) in python and it gives me this result
[[ -8.04778516e-05 -1.01345901e-02 1.48194623e-03]
[ -9.11603819e-05 -1.17454886e-02 2.26221011e-03]]
I'm doing it in golang with the same inputs matrix(M1,M2)
but the golang code returns this matrix:
[[-8.047785157755936e-05 -0.010134590118173147 0.0014819462317188985]
[-9.116038191682538e-05 -0.011745488603430228 0.0022622101145935328]]
In python I'm using numpy's dot operation:
resultMatrix = M1.dot(M2)
In go, I'm using this package to work with matrix in go
The problem here is because I calculate others dotProcut calculos with golang and it are all ok
I make N tests with other values, i'm using this package(the same dotProduct method) in others parts of this my code and all has been ok
My Go code at line 128
Tutorial Python code at line 61
Matrix golang package method that implemets the golang dotProduct at line 30
The code in python is not mine, and because this, the code it's written in Portuguese, but my go code is written in English
In python i know that's right because all the neural network works well, but in go I'm not sure
i read the matrix go package method many times and dont get the "bug code implementation", some one know where I'm wrong?
Well, actually the results are pretty much the same. The thing that might confuse you is that formatting is different but still Python's -1.01345901e-02 = -0.0101345901 (see Scientific notation and particularly its E-notation" section) which is pretty close to Go's -0.010134590118173147 and just to make it clear let's align them
Python -1.01345901e-02
Go -0.010134590118173147
So if you have any problems in your code, they probably come from some other source than matrix multiplication.
in our project we have to use the GreenHillsCompiler Suite MULTI. So everything is configured and running. Reading the compiler manual I found an option for the linker which will generate a callgraph.
I added the option to the linker (elxr) in the makefile with
LINK_OPT += -callgraph
which generates a file with the extension ".graph" in the output folder. This files just contains plain text.
Function Function Call Call Count Percent of Total Max Displacement (bits)
#% BEGIN STATIC GRAPH
myFunc out 0 in 3
out 0 100%
in 3 100%
myFunc2 1 33% 0 2514 0
.static00012204 1 33% 0 514 0
.static0001220b 1 33% 0 1300 0
#% END STATIC GRAPH
So the question is: What tool has to be used further?
What we want is an image or a html-document.
That is the graph. It is in text format.
Note that if you want a graphical representation of a function's call graph, from within the MULTI debugger you can right-click a function and select Browse Other -> Browse Static Calls.
I want to try the LDA -c code by Blie .et.al. as it is in this link.
I have compiled the code, and when I run ./lda in my terminal, the following result is displayed.
usage : lda est [initial alpha] [k] [settings] [data] [random/seeded/manual=filename/*] [directory]
lda inf [settings] [model] [data] [name]
Which means that, it has been complied correctly.
However, in spite reading the README.txt file there, I am not being able to succesfully run the LDA code.
Either it says Segmentation fault (core dumped) or killed.
What am I missing? How to use it on the example data they have given?
I have read the stack overflow answer to the question asked here, but it was not useful as I dont know the default values.
P.S: I am a beginer.
Are you using ap.txt instead of ap.dat by any chance? lda-c doesn't take raw sentences or marked up data as input; it takes a sequence of bag of words information for each document. When ap.dat has a line like
186 0:1 6144:1 3586:2 ..., it means that the corresponding document has 186 distinct words, word 0 appears once, word 6144 appears once, word 3586 appears twice, and so on.
This command works for me (using Blei's original code):
./lda est 0.1 10 settings.txt ap.dat random modeldir
(Feel free to tweak the initial alpha (0.1) and number of topics (10) as you wish.)
I am working on a program that outputs PDF documents. Given a sequence of UTF-8 encoded characters and the name of a font that shall be used to render it, I would like to show the appropriate glyphs that make the actual content of the document. I would like to be able to display national characters such as č or ö. It would be great to support ligatures like ae or ffi.
The problem is, I do not know how the actual glyphs to be shown are specified (inside a content stream, for example).
If I, for example, want to display the string "Hello World", I need not to worry about encoding, I simply write (Hello World)Tj. The PDF reader will then use the appropriate font to render this string.
But what if I wanted to show the string
It is difficult to read the PDF specification all day. Prostě dočista nemožné!
with the ligatures ffi, fi and ea and the Czech national symbols ě, č and é in a given font, how would I proceed?
I am trying to get through the PDF specification, but it is not easy.
How do I find out the "code of the glyph" that corresponds to a given character or ligature?
How is this code encoded within a PDF content stream?
Help is much appreciated.
Edit: I may have overestimated the problem. Counting the glyphs that are needed to display a "common European document", I cannot think of a way how this number could exceed 256. If my assumptions are correct, I can remap the encoding of the font completely. This should be sufficient to cover all common symbols of the latin alphabet, numbers, punctuation, common symbols like ( and [ and still I would have plenty of room for national symbols, ligatures and other elements of high-quality typography. (I can implement a priority queue to select the most used ligatures if the total number of glyphs shall exceed 256.)
That being said, I do not think I need to use the CID-keyed fonts.
Still I wander how do I map UTF-8 encoded characters onto glyphs of an arbitrary font. I have the AFM of the font available. For the DejaVu font, for example, character information go like this:
C 63 ; WX 536 ; N question ; B 67 -15 488 743 ;
C 64 ; WX 1000 ; N at ; B 65 -174 930 705 ;
C 65 ; WX 722 ; N A ; B -6 0 732 730 ;
But after the 256th character is mapped, the codes are -1:
C 255 ; WX 564 ; N ydieresis ; B -3 -223 563 767 ;
C -1 ; WX 722 ; N Amacron ; B -6 0 732 899 ;
C -1 ; WX 596 ; N amacron ; B 49 -15 568 746 ;
For example, if I had the sequence 11100010 10000010 10101100 (Euro sign) in my input, how would I know what glyph name it corresponds to so that I can map it in the /Encoding dictionary?
Encoding varies based on the font type. Typically, there is a font resource that is defined as the current font and within that font dictionary is a reference to a base font and a means of describing the encoding (via the /Encoding key). If that key doesn't exist, the encoding will be "standard", but you can use other simple encodings such as /MacRoman and /WinAnsi for the value of the encoding, or you can specify a standard encoding and an encoding delta to show the differences.
Easy so far - as long as you're working with 8-bit characters. For many early apps, they would create a couple different fonts, one with say Roman encoding and another that maps roman characters to unavailable characters. In order to do that, your encoding delta would include references to the ligatures and other typically non-encoded symbols. This works great for Type 1 fonts, but is specifically contraindicated by the spec in the section on TrueType Fonts:
A nonsymbolic font should specify MacRomanEncoding or WinAnsiEncoding as the value of its Encoding entry, with no Differences array
This is vastly different when you want to use, say, Unicode. In which case you would be using a CID font (a font based on character IDs). In that case there is a procedure referenced by the font which is used to map from a character encoding in your string to a character ID in your font (and vice versa). I would strongly recommend that you read and fully understand section 9.7 in the PDF specification on Composite Fonts, which describes everything you need in order to encode UTF16BE into strings to get them to render properly in PDF. It is decidedly non-trivial in that there are a lot of details that if missed will result in a blank rendered page in Acrobat.
As a software engineer who professionally writes code that produces and consumes PDF, let me state that when I get tasked with having to put in special cases in my code to deal with non-spec compliant PDF, a little piece of me dies inside. Please, please, don't even think of releasing any documents you produce into the wild until they pass Preflight at the least. This is not the same as "Acrobat renders it so it must be OK." Let me give you an example - I've seen a number of files in the wild that include fonts that are missing the key elements of the FontDescriptor dictionary, including /Ascent, /Descent, /CapHeight, etc. These render in Acrobat, but are in violation of the spec since each of those is required. I know how Acrobat handles that - it comes with an enormous database of font metrics and looks up the value if it can't find it in the file (heck, it might even ignore the metrics in the file). I don't have that luxury, so I have to do a number of (potentially expensive/invalid) stop gap measures.
You might want to consider using a library to do this work for you - maybe iText which has a decent enough licensing scheme for education because, I get it, you're a student. There are some C based libraries too. Maybe you can figure a way to make GhostScript do your bidding.
If you are unwilling or unable to follow my advice with regards to cleaving to the specification or to use a library which ostensibly does so, please do me the favor of at least filling out the /Creator and /Producer strings in the Document Information Dictionary referenced by the trailer (see sections 14.3.3 and section 7.5.5). That way, when I have to parse/consume/manipulate your documents, I will have a way to directly cast aspersions on your parentage.
Let's go top down and start with the page object - I'm using output from my own library and am stripping out what I think you don't need:
1 0 obj <<
/Type /Page
/Parent 18 0 R
/Resources <<
/Font <<
/U0 13 0 R
>>
/ProcSet [ /PDF /Text ]
>>
/MediaBox [ 0 0 612 792 ]
/Contents 19 0 R
/Dur -1
>>
endobj
U0 is a reference to a font that will be used for unicode text.
The content stream is intended to print the following text: Greek: Γειά σου κόσμος.
BT /U0 24 Tf 72 670 Td
(\000G\000r\000e\000e\000k\000:\000 \003\223\003\265\003\271\003\254\000 \003\303\003\277\003\305\000 \003\272\003\314\003\303\003\274\003\277\003\302)
Tj ET
The font dictionary referenced looks like this:
13 0 obj <<
/BaseFont /DejaVuSansCondensed
/DescendantFonts [ 4 0 R ]
/ToUnicode 14 0 R
/Type /Font
/Subtype /Type0
/Encoding /Identity-H
>>
endobj
Which has the /ToUnicode entry points to a stream containing the following PostScript code:
/CIDInit /ProcSet findresource begin 12 dict begin begincmap /CIDSystemInfo << /Registry (Adobe) /Ordering (UCS) /Supplement 0 >> def /CMapName /Adobe-Identity-UCS def /CMapType 2 def 1 begincodespacerange <0000> <FFFF> endcodespacerange 1 beginbfrange <0000> <FFFF> <0000> endbfrange endcmap CMapName currentdict /CMap defineresource pop end end
which is defined by the CID font specification.
and the DescendantFonts array points to this object:
4 0 obj <<
/Subtype /CIDFontType2
/Type /Font
/BaseFont /DejaVuSansCondensed
/CIDSystemInfo 7 0 R
/FontDescriptor 8 0 R
/DW 1000
/W 9 0 R
/CIDToGIDMap 10 0 R
>>
The CIDToGIDMap is a compressed stream with the actual map, the CIDSystemInfo is <</Registry (Adobe) /Ordering (USC) /Supplement 0>> (it's a reference because I share it among all unicode fonts that I output. The FontDescriptor is a straight forward boiler plate, and the W array is derived from the font metrics.
With all this detail, are you understanding why I don't say lightly, "walk away before you pollute my environment any furhter"?
I'm really beginning to question the nature of the this assignment. Writing a simple PDF is one thing, but writing code that can handle full unicode in any arbitrary OpenType/TrueType font requires you to understand the CID spec and the TrueType spec (hint: I have a full TrueType parser that can extract all the metrics for any glyph in a font so that I can output the /W array).
If, however, you are required to only output to Type 1 fonts, well my friend, your life got a whole lot easier, because you would take your entire UTF8 stream, read it as unicode and for every unique character that comes in, you build a map from a unicode character to a glyph name and an internal character number by using this table. The internal character number essentially the unique index of the character that came in mod. So for example, if you have less than 257 unique characters on the page, you will have exactly one font that is encoded to map to the characters in the order that the arrived. If you had "abcba" for input, the output string in pdf would be (\000\001\002\001\000) and would map to a font with an encoding dictionary with a differences array that would be [0/a/b/c]. If you have n unique characters where n > 256, you're going to have (n / 256) + 1 fonts, each with encodings.
If your teacher/professor wants anything but Type 1 fonts in a short period of time, s/he has unrealistic expectations for the students and/or low expectations for the quality of output. You should ask whether your are required to handle CID fonts and if you are, then your professor is at the very least a sadist. It took me, a seasoned professional, about 4 days to write a TrueType parser for extracting widths. I had the advantage of (1) using a managed language (C#) which cut down on concerns that will be biting your ass in C and was also able to use reflection to automate parsing and (2) when I don't have interruptions, I write solid code about 10-20 times faster than a typical student, so my 32 hours would translate into 320 student hours, more or less (then again, my code has different constraints than yours - it has to consume any crap font it gets gracefully), so let's call it 200 or less if you're allowed to steal something like stb. That's just for getting one particular element in the font descriptor.
I have the following call statement :
038060 CALL PROG USING
038070 DFH
038080 L000
038090 ZONE-E
038100 ZONE-S.
This call is dynamic and use PROG.
PROG is a group defined as :
018630 01 XX00.
018640 10 PROG.
018650 15 XX00-S06 PICTURE X(6)
018660 VALUE SPACE.
018670 15 XX00-S02 PICTURE X(2)
018680 VALUE SPACE.
018690 10 XX00-S92 PICTURE 9(02)
018700 VALUE ZERO.
018710 10 XX00-S91 PICTURE 9(1)
018720 VALUE ZERO.
018730 10 XX00-S9Z PICTURE 9(1)
018740 VALUE ZERO.
018750 10 XX00-9B0 PICTURE X(05)
018760 VALUE SPACE.
018770 10 XX00-0B0 PICTURE X(02)
018780 VALUE SPACE.
018790 10 XX00-BB1 PICTURE X(01)
018800 VALUE SPACE.
018810 10 XX00-SFN PICTURE X(07)
I cut here but there is a lot of field after...
It seems that actual progname to use is stored in :
XX00-S06
and
XX00-S02
I've also other cases where the name is on 3 or 4 fields, and the progname length is not always 8.
So my question is how Cobol know where to pick the good program name in the group? What are the resolution rules?
Configuration : I use Microfocus Net Express compiler and the environment is UniKix.
Dynamic call rules in COBOL are fairly simple. Given something like:
CALL WS-NAME USING...
COBOL will resolve the program name currently stored in WS-NAME against the load module libraries
available to it based on
a linear search. The first matching load module entry point name that matches WS-NAME is used.
It doesn't matter how complex, or simple, the definition of WS-NAME is. The total length used for the name
is whatever the length of WS-NAME is. For example:
01 WS-NAME.
05 WS-NAME-FIRST-PART PIC X(3).
05 WS-NAME-MIDDLE-PART PIC X(2).
05 WS-NAME-LAST-PART PIC X(3).
WS-NAME is composed of 3 subordinate fields giving a total of 8 characters. You can populate these individually or just move
something into WS-NAME as a whole. If the length of WS-NAME is less than 8 characters, the trailing characters will be
set to spaces on any receiving field. For example:
01 WS-SHORT-NAME.
05 WS-SHORT-NAME-FIRST-PART PIC X(4) VALUE 'AAAA'.
05 WS-SHORT-NAME-LAST-PART PIC X(2) VALUE 'BB'.
Here WS-SHORT-NAME is only 6 characters long. MOVING WS-SHORT-NAME to any longer PIC X type variable as in:
MOVE WS-SHORT-NAME TO WS-NAME
Will result in WS-NAME taking on the value 'AAAABBbb' (note the two trailing spaces). During libary search
for a matching entry point name, the trailing spaces are not significant so on the CALL statement you could use
either:
CALL WS-NAME
or
CALL-WS-SHORT-NAME
And they will resolve to the same entry point.
I am not sure what the length rules are for MicroFocus COBOL but, for IBM z/os dynamically called
program names cannot exceed 8 characters (if they do, the name is truncated to 8 characters).
I will add little more to NeilB with specific information about Micro Focus COBOL.
fyi: PROGRAM-ID, ENTRY-POINTS are restricted to 30-31 characters (check your "System Limits and Programming Restrictions" section in the docs).