why there is always error "ValueError: Input Pipeline definition must have a class property." about seq2seq HOT 5 CLOSED

I think the problem here is that you are not passing the input_pipeline_dev argument, which is required. Ideally you should always have a validation set, not just a training set ;)

It's probably a good idea to throw a better error message and to make the development data optional. Will do that.

For now I guess a workaround is to just copy a small subset of your training data and use that as validation data to get around the error. Thanks for reporting.

from seq2seq.

Unfortunately the Tensorflow Experiment class does not allow you to disable evaluation, so you need to pass an input_pipeline_dev. I at least updated the code to throw a better error message though.

from seq2seq.

Thanks for your kindly reply! @dennybritz

from seq2seq.

Same problem. Thank you.

from seq2seq.

@dennybritz i am having the same problem but i am using both dev data as well as train set.
config.yml
`model: AttentionSeq2seq
model params:
vocab_source: ${HOME}/nil/beam/seq2seq/tmp/vocab.50k.bpe.32000.en
vocab_target: ${HOME}/nil/beam/seq2seq/tmp/vocab.50k.bpe.32000.hi
input_pipeline_train:
input_pipeline_train:
class: ParallelTextInputPipeline
params:
- source_files: ${HOME}/nil/beam/seq2seq/tmp/test.bpe.32000.en
- target_files: ${HOME}/nil/beam/seq2seq/tmp/test.bpe.32000.hi
input_pipeline_dev:
class: ParallelTextInputPipeline
params:
- source_files: ${HOME}/nil/beam/seq2seq/tmp/IITB.en-hi.bpe.32000.en
- target_files: ${HOME}/nil/beam/seq2seq/tmp/IITB.en-hi.bpe.32000.hi
train_steps: ${HOME}/nil/beam/seq2seq/tmp/train
output_dir: ${HOME}/nil/beam/seq2seq/tmp/model
'

beam.yml
model: AttentionSeq2Seq
model_params:
attention.class: seq2seq.decoders.attention.AttentionLayerBahdanau
attention.params:
num_units: 512
bridge.class: seq2seq.models.bridges.ZeroBridge
embedding.dim: 512
encoder.class: seq2seq.encoders.BidirectionalRNNEncoder
encoder.params:
rnn_cell:
cell_class: GRUCell
cell_params:
num_units: 512
dropout_input_keep_prob: 0.8
dropout_output_keep_prob: 1.0
num_layers: 2
decoder.class: seq2seq.decoders.BeamSearchDecoder
decoder.params:
rnn_cell:
cell_class: GRUCell
cell_params:
num_units: 512
dropout_input_keep_prob: 0.8
dropout_output_keep_prob: 1.0
num_layers: 2
optimizer.name: Adam
optimizer.params:
epsilon: 0.0000008
optimizer.learning_rate: 0.0001
source.max_seq_len: 50
source.reverse: false
target.max_seq_len: 50

training the set: python -m bin.train --config_paths="./example_configs/beam.yml,example_configs/train_seq2seq.yml,./config.yml"

from seq2seq.