philipus
commented
4 years ago very strange habit... with train.R form https://cran.r-project.org/web/packages/cloudml/vignettes/deployment.html
export_savedmodel works but with my code (the part below) not. Any Idea what the problem is?!
model <- keras_model( inputs = list(encoder_inputs, decoder_inputs), outputs = decoder_outputs )
Compile model
model %>% compile(optimizer='adam', loss='mean_absolute_error')
model %>% fit( list(encoder_input_data, decoder_input_data), decoder_target_data, batch_size=batch_size, epochs=epochs, validation_split=0.2)
export_savedmodel(model, "savedmodel")
export_savedmodel using cloudml did not work properly (see below) on google AI Platform.
How do you save and download the model for further steps (applying model on new data) on local computer?!
Error: ValueError: Attempted to save a function __inference_lstm_layer_call_fn_8346 which references a symbolic Tensor Tensor("dropout/mul_1:0", shape=(?, 1), dtype=float32) that is not a simple constant. This is not supported. Detailed traceback: File "//usr/local/lib/python2.7/dist-packages/tensorflow_core/python/saved_model/save.py", line 860, in save meta_graph_def, saveable_view, signatures) File "//usr/local/lib/python2.7/dist-packages/tensorflow_core/python/saved_model/save.py", line 556, in _fill_meta_graph_def object_map, resource_map, asset_info = saveable_view.map_resources() File "//usr/local/lib/python2.7/dist-packages/tensorflow_core/python/saved_model/save.py", line 291, in map_resources "supported.").format(concrete_function.name, capture)) py_call_impl(callable, dots$args, dots$keywords) tensorflow::tf$saved_model$save(obj = object, export_dir = export_dir_base, ...) export_savedmodel.keras.engine.training.Model(model, "savedmodel") export_savedmodel(model, "savedmodel") eval(ei, envir) eval(ei, envir) withVisible(eval(ei, envir)) source(file = file, local = envir, echo = echo, encoding = encoding) withCallingHandlers({ source(file = file, local = envir, echo = echo, encoding = encoding) write_run_property("completed", TRUE) }, error = function(e) { write_run_metadata("error", list(message = e$message, traceback = capture_stacktrace(sys.calls()))) stop(e) }) force(expr) tryCatch({ tfruns::training_run(file = deploy$entrypoint, context = deploy$context, config = "cloudml", flags = deploy$overlay, encoding = "UTF-8", echo = TRUE, view = FALSE, run_dir = run_dir) }, error = function(e) { training_error <<- e })