Creating a dictionary from text file Python Numpy

I have a text file that looks like this:

# Comments 
PARAMETER  0  0
      1045        54
      1705         0                           time 1
         1        10       100   0.000e+00   9999   A
         2        20       200   0.2717072   9999   B
         3        30       300   0.0282928   9999   C
         1       174        92   2999.4514   9999   APEW-1
         2       174        92   54.952499   9999   ART-3A
         1       174        97   5352.1299   9999   APEW-2
         1       173       128   40.455467   9999   APEW-3
         2       173       128   1291.1320   9999   APEW-3
         3       173       128   86.562599   9999   ART-7B
...

I want to create a dictionary that looks like below (basically skipping the header and certain columns and goes to the data that I need):

my_dict = {'A':(1,10,100),'B':(2,20,200), 'C':(3,30,300), 'APEW-1':(1,174,92), ...}

These data point are observation points and their respective values are depth, y, x. Therefore one observation point can have multiple values for different depths (the first column). I am trying to avoid rename the labels by adding a suffix for duplicates. I wonder if there is any way around it. What I want to do with them is to call a observation point name and extract the coordinates. I am not sure if the dictionary is the right tool for this purpose. It is an small dataset and doesn't need to be fast. I am using Numpy, Python 2.7.

loadtxt can do it:

>>> dtype=np.rec.fromrecords([[0, 0, 0, b'APEW-1']]).dtype
>>> x = np.loadtxt(fn, skiprows=4, usecols=(0,1,2,5), dtype=dtype)
>>>
>>> result = {}
>>> for x0, x1, x2, key in x:
...     try:
...         result[key.decode()].append((x0,x1,x2))
...     except KeyError:
...         result[key.decode()] = [(x0,x1,x2)]
... 
>>> result
{'A': [(1, 10, 100)], 'B': [(2, 20, 200)], 'C': [(3, 30, 300)], 'APEW-1': [(1, 174, 92)], 'ART-3A': [(2, 174, 92)], 'APEW-2': [(1, 174, 97)], 'APEW-3': [(1, 173, 128), (2, 173, 128)], 'ART-7B': [(3, 173, 128)]}

Notes:

we abuse rec.fromrecords to create a compund dtype describing the columns, be sure to use a template string as long as the longest you expect
- there is probably an official way of creating compound dtypes that doesn't involve creating a throw-away array but this is easy and works
loadtxt paramemters are self-explanatory, because of the compound dtype it generates a 1d recordd array
if there were no duplicate keys, we could use dict comprehension to translate the record array to dict f0-f3 are the auto generated field names
- to accomodate duplicates we pack the values which are tuples in lists
- most lists contain just one tuple, but some will have more

py2 version: main difference no need to use byte strings / decode, dictionary forgets order of items

>> dtype=np.rec.fromrecords([[0, 0, 0, 'APEW-1']]).dtype
>>> x = np.loadtxt(fn, skiprows=4, usecols=(0,1,2,5), dtype=dtype)
>>>
>>> result = {}
>>> for x0, x1, x2, key in x:
...     try:
...         result[key].append((x0,x1,x2))
...     except KeyError:
...         result[key] = [(x0,x1,x2)]
... 
>>> result
{'A': [(1, 10, 100)], 'B': [(2, 20, 200)], 'C': [(3, 30, 300)], 'APEW-1': [(1, 174, 92)], 'ART-3A': [(2, 174, 92)], 'APEW-2': [(1, 174, 97)], 'APEW-3': [(1, 173, 128), (2, 173, 128)], 'ART-7B': [(3, 173, 128)]}