some questions about the net.prototxt

[zkself]

hello ! i'am very appreciate your nice work!

i have some question about your net.prototxt.
First,
layer {
name: "embedding"
type: "Embed"
bottom: "input"
top: "embedding"
param {
name: "embed_param"
}
propagate_down: false
embed_param {
num_output: 1000
input_dim: 10010
bias_term: false
weight_filler {
type: "gaussian"
std: 0.00999999977648
}
}
}

why
propagate_down: false?
is that means the embedding layer is pretrained? but,i have not found the clue in your code.

[peteanderson80]

The propagate_down flag affects the inputs to this layer, it does not stop this layer from learning: https://github.com/BVLC/caffe/blob/04ab089db018a292ae48d51732dd6c66766b36b6/src/caffe/proto/caffe.proto#L347-L354
The embedding layer is not pretrained and the propagate_down flag is not actually necessary here. I think it is left over from some experiments I was doing that involved backpropagating through beam search or something.

[zkself]

thanks for your reply By the way, the caffe code is too hard to learn. I am trying to rewrite your up down code in pytorch On 03/06/2019 22:47, Peter Anderson wrote: The propagate_down flag affects the inputs to this layer, it does not stop this layer from learning: https://github.com/BVLC/caffe/blob/04ab089db018a292ae48d51732dd6c66766b36b6/src/caffe/proto/caffe.proto#L347-L354 The embedding layer is not pretrained and the propagate_down flag is not actually necessary here. I think it is left over from some experiments I was doing that involved backpropagating through beam search or something. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[zkself]

@peteanderson80
i have encountered another problem when i run the train.sh. the error output is listed blew:

HDF5-DIAG: Error detected in HDF5 (1.8.16) thread 139667420518144:
#000: ../../../src/H5G.c line 467 in H5Gopen2(): unable to open group
major: Symbol table
minor: Can't open object
#1: ../../../src/H5Gint.c line 320 in H5G__open_name(): group not found
major: Symbol table
minor: Object not found
#2: ../../../src/H5Gloc.c line 430 in H5G_loc_find(): can't find object
major: Symbol table
minor: Object not found
#3: ../../../src/H5Gtraverse.c line 861 in H5G_traverse(): internal path traversal failed
major: Symbol table
minor: Object not found
#4: ../../../src/H5Gtraverse.c line 641 in H5G_traverse_real(): traversal operator failed
major: Symbol table
minor: Callback failed
#5: ../../../src/H5Gloc.c line 385 in H5G_loc_find_cb(): object 'data' doesn't exist
major: Symbol table
minor: Object not found
F0311 21:32:04.921723 33935 net.cpp:793] Check failed: data_hid >= 0 (-1 vs. 0) Error reading weights from snapshots/caption_lstm/lstm_scst_iter_1000.caffemodel.h5
*** Check failure stack trace: ***
@ 0x7f06736175cd google::LogMessage::Fail()
@ 0x7f0673619433 google::LogMessage::SendToLog()
@ 0x7f067361715b google::LogMessage::Flush()
@ 0x7f0673619e1e google::LogMessageFatal::~LogMessageFatal()
@ 0x7f0673bb0848 caffe::Net<>::CopyTrainedLayersFromHDF5()
@ 0x7f0673bb4d04 caffe::Net<>::CopyTrainedLayersFrom()
@ 0x7f067430ee8c caffe::Net_Init_Load()
@ 0x7f0674329a43 boost::python::objects::signature_py_function_impl<>::operator()()
@ 0x7f0672d935cd boost::python::objects::function::call()
@ 0x7f0672d937c8 (unknown)
@ 0x7f0672d9b613 boost::python::handle_exception_impl()
@ 0x7f0672d90999 (unknown)
@ 0x7f06db017973 (unknown)
@ 0x7f06db02680d (unknown)
@ 0x7f06db017973 (unknown)
@ 0x7f06db070764 _PyExc_Init
@ 0x7f06db06d01b method_get
@ 0x7f06db017973 (unknown)
@ 0x7f06db0add49 PyType_Ready
@ 0x7f06db0b217e slot_nb_xor
@ 0x7f06db0b36c9 type_setattro
@ 0x7f06db0b38ea type_new
@ 0x7f06db0ccbad _PyUnicodeUCS4_ToNumeric
@ 0x7f06db0cdd28 _PyUnicodeUCS4_ToNumeric
@ 0x7f06db0cef48 warn_explicit
@ 0x7f06db0e116c ast_for_exprlist.isra.16
@ 0x7f06da277830 libc_start_main
@ 0x55d114aac07f (unknown)
./experiments/caption_lstm/train.sh: line 27: 33935 Aborted (core dumped) python ./scripts/beam_decode.py --gpu ${GPU_ID:0:1} --model ${BASE_DIR}${NET_NAME}/decoder.prototxt --weights=${SNAPSHOT_DIR}${NET_NAME}/lstm_scst_iter${SCST_MAX_IT}.caffemodel.h5 --vocab ${DATA_DIR}${VOCAB_FILE} --outfile ${OUT_DIR}/${NET_NAME}/scst_iter${SCST_MAX_IT}.json

[GaoYifanGHB]

Check failed: data_hid >= 0 (-1 vs. 0) Error reading weights from snapshots/caption_lstm/lstm_scst_iter_1000.caffemodel.h5

I have the same problem, have you solved it? @zkself

[wonderseen]

@GaoYifanGHB Hi, Gao.
I got a broken connection to the published model lstm_scst_iter_1000.caffemodel.h5 (listed in author's code). Could you please share the way how you get the model :)?
Thank you.

[YoussefSaadYoussef]

I have the same problem
Check failed: data_hid >= 0 (-1 vs. 0) Error reading weights from snapshots/caption_lstm/lstm_scst_iter_1000.caffemodel.h5

did you solve it? @GaoYifanGHB @zkself