A bad choice of "window_samples" crashes PyCall

[ifmelchor]

Hi,

I executed the following code:

using SeisIO
using PhaseNet

S = read_data("mseed", MSEED_FILE) # MSEED_FILE is a three-component miniSEED file of less than 5 minutes of duration
detrend!(S)
sync!(S)

window_samples = 100
batch_size = 256
X = seisdata2torch(S,window_samples)
model_P = load_model("P")
probs_P = detect(X,model_P,batch_size)

And I have this error:

PyError ($(Expr(:escape, :(ccall(#= /home/ivan/.julia/packages/PyCall/twYvK/src/pyfncall.jl:43 =# @pysym(:PyObject_Call), PyPtr, (PyPtr, PyPtr, PyPtr), o, pyargsptr, kw))))) <class 'RuntimeError'>
RuntimeError('Sizes of tensors must match except in dimension 1. Expected size 15 but got size 14 for tensor number 1 in the list.')
  File "/home/ivan/.local/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "PyCall", line 1, in <lambda>

Stacktrace:
  [1] pyerr_check
    @ ~/.julia/packages/PyCall/twYvK/src/exception.jl:75 [inlined]
  [2] pyerr_check
    @ ~/.julia/packages/PyCall/twYvK/src/exception.jl:79 [inlined]
  [3] _handle_error(msg::String)
    @ PyCall ~/.julia/packages/PyCall/twYvK/src/exception.jl:96
  [4] macro expansion
    @ ~/.julia/packages/PyCall/twYvK/src/exception.jl:110 [inlined]
  [5] #107
    @ ~/.julia/packages/PyCall/twYvK/src/pyfncall.jl:43 [inlined]
  [6] disable_sigint
    @ ./c.jl:473 [inlined]
  [7] __pycall!
    @ ~/.julia/packages/PyCall/twYvK/src/pyfncall.jl:42 [inlined]
  [8] _pycall!(ret::PyCall.PyObject, o::PyCall.PyObject, args::Tuple{PyCall.PyObject}, nargs::Int64, kw::Ptr{Nothing})
    @ PyCall ~/.julia/packages/PyCall/twYvK/src/pyfncall.jl:29
  [9] _pycall!(ret::PyCall.PyObject, o::PyCall.PyObject, args::Tuple{PyCall.PyObject}, kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ PyCall ~/.julia/packages/PyCall/twYvK/src/pyfncall.jl:11
 [10] (::PyCall.PyObject)(::PyCall.PyObject, ::Vararg{Any}; kwargs::Base.Pairs{Symbol, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
    @ PyCall ~/.julia/packages/PyCall/twYvK/src/pyfncall.jl:86
 [11] (::PyCall.PyObject)(::PyCall.PyObject, ::Vararg{Any})
    @ PyCall ~/.julia/packages/PyCall/twYvK/src/pyfncall.jl:86
 [12] macro expansion
    @ ~/.julia/packages/PhaseNet/ZNXkE/src/detection.jl:120 [inlined]
 [13] macro expansion
    @ ~/.julia/packages/PyCall/twYvK/src/PyCall.jl:660 [inlined]
 [14] detect(X::PyCall.PyObject, model::PyCall.PyObject, batch_size::Int64; device::Nothing)
    @ PhaseNet ~/.julia/packages/PhaseNet/ZNXkE/src/detection.jl:119
 [15] detect(X::PyCall.PyObject, model::PyCall.PyObject, batch_size::Int64)
    @ PhaseNet ~/.julia/packages/PhaseNet/ZNXkE/src/detection.jl:105
 [16] top-level scope
    @ ~/Desktop/phasent_test.jl:16
 [17] include(fname::String)
    @ Base.MainInclude ./client.jl:476
 [18] top-level scope
    @ REPL[1]:1

However, if I set the window_samples to 160, I have no errors, and the code is executed successfully.

So, Are there any combinations of window_samples, batch_size, and npts to bear in mind before applying the detector?

[tclements]

I believe this error is being caused by the maxpool operation. When window_samples = 100, there aren't enough time samples to pool. I would suggest using a minimum window_samples = 256 and increase from there. The batch_size is number of samples in the first dimension of the X tensor to process at once. This is mainly important if you're using a GPU. For instance, here is a SeisData object with 34.8 seconds of data at 100 Hz:

julia> S = read_data("seisio", seisiopath)[1:3]
SeisData with 3 channels (3 shown)
    ID: KI.ABSH01..HNE                     KI.ABSH01..HNN                     KI.ABSH01..HNZ
  NAME: KI.ABSH01..EW2                     KI.ABSH01..NS2                     KI.ABSH01..UD2
   LOC: 44.5276 N, 142.844 E, 105.0 m      44.5276 N, 142.844 E, 105.0 m      44.5276 N, 142.844 E, 105.0 m
    FS: 100.0                              100.0                              100.0
  GAIN: 1.0                                1.0                                1.0
  RESP: a0 1.0, f0 1.0, 0z, 0p             a0 1.0, f0 1.0, 0z, 0p             a0 1.0, f0 1.0, 0z, 0p
 UNITS: cm/s^2                             cm/s^2                             cm/s^2
   SRC:
  MISC: 0 entries                          0 entries                          0 entries
 NOTES: 3 entries                          3 entries                          3 entries
     T: 2012-12-07T17:18:43 (0 gaps)       2012-12-07T17:18:43 (0 gaps)       2012-12-07T17:18:43 (0 gaps)
     X: -8.015e-04                         +2.753e-03                         -5.642e-04
        +8.900e-04                         +6.775e-04                         -3.672e-04
            ...                                ...                                ...
        -1.340e-02                         -5.468e-02                         -7.772e-02
        (nx = 34800)                       (nx = 34800)                       (nx = 34800)

Using a window_samples = 400, there are 87 total batches in X:

julia> window_samples = 400

julia> X = seisdata2torch(S,window_samples)

julia> X.shape
(87, 3, 400)

Since X is fairly small, we can match the batch_size to the size of X's first dimension:

batch_size = X.shape[1]

For much larger X, say using an entire day of continuous data, I would increase the batch_size until running out of GPU or CPU memory, which will depend on your hardware.