[ https://issues.apache.org/jira/browse/MADLIB-1426?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Domino Valdano updated MADLIB-1426: ----------------------------------- Description: Whenever I try to run {{madlib_keras_fit_multiple_model()}} on a system without GPU's, it always fails in evaluate complaining that device `gpu0` is not available. This happens regardless of whether {{use_gpus=False}} or use_gpus=True. My platform is OSX 10.14.1 with latest version of madlib (1.17.0) and gpdb5. I think I've also seen this happen on CentOS in gpdb6, so I believe this is a bug that affects all platforms, but not entirely sure of that. Possibly specific to OSX or gpdb5. The problem happens in {{internal_keras_eval_transition()}} in {{madlib_keras.py_in}}. With {{use_gpus=False}}, it runs: {{ with K.tf.device(device_name): res = segment_model.evaluate(x_val, y_val) }} ``` I added a {{plpy.info}} statement to print {{device_name}} at the beginning of this function. I also printed the value of {{use_gpus}} on master before training begins. While {{use_gpus}} is set to false, the {{device_name}} on the segments is set to {{/gpu:0}}. This is the bug (it should be set to {{/cpu:0}}). This is the error message that happens: ``` INFO: 00000: {{use_gpus = False}} ... INFO: 00000: device_name = /gpu:0 (seg1 slice1 127.0.0.1:25433 pid=90300) CONTEXT: PL/Python function "internal_keras_eval_transition" LOCATION: PLy_output, plpython.c:4773 psql:../run_fit_mult_iris.sql:1: ERROR: XX000: plpy.SPIError: tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation group_deps: Operation was explicitly assigned to /device:GPU:0 but available devices are [ /job:localhost/replica:0/task:0/device:CPU:0 ]. Make sure the device specification refers to a valid device. (plpython.c:5038) (seg0 slice1 127.0.0.1:25432 pid=90299) (plpython.c:5038) DETAIL: [[{{node group_deps}} = NoOp[_device="/device:GPU:0"](^loss/mul, ^metrics/acc/Mean)]] Traceback (most recent call last): PL/Python function "internal_keras_eval_transition", line 6, in <module> return madlib_keras.internal_keras_eval_transition(**globals()) PL/Python function "internal_keras_eval_transition", line 782, in internal_keras_eval_transition PL/Python function "internal_keras_eval_transition", line 1112, in evaluate PL/Python function "internal_keras_eval_transition", line 391, in test_loop PL/Python function "internal_keras_eval_transition", line 2714, in __call__ PL/Python function "internal_keras_eval_transition", line 2670, in _call PL/Python function "internal_keras_eval_transition", line 2622, in _make_callable PL/Python function "internal_keras_eval_transition", line 1469, in _make_callable_from_options PL/Python function "internal_keras_eval_transition", line 1351, in _extend_graph PL/Python function "internal_keras_eval_transition" CONTEXT: Traceback (most recent call last): PL/Python function "madlib_keras_fit_multiple_model", line 23, in <module> fit_obj = madlib_keras_fit_multiple_model.FitMultipleModel(**globals()) PL/Python function "madlib_keras_fit_multiple_model", line 42, in wrapper PL/Python function "madlib_keras_fit_multiple_model", line 216, in __init__ PL/Python function "madlib_keras_fit_multiple_model", line 230, in fit_multiple_model PL/Python function "madlib_keras_fit_multiple_model", line 270, in train_multiple_model PL/Python function "madlib_keras_fit_multiple_model", line 302, in evaluate_model PL/Python function "madlib_keras_fit_multiple_model", line 417, in compute_loss_and_metrics PL/Python function "madlib_keras_fit_multiple_model", line 739, in get_loss_metric_from_keras_eval PL/Python function "madlib_keras_fit_multiple_model" LOCATION: PLy_elog, plpython.c:5038 ``` was: Whenever I try to run `madlib_keras_fit_multiple_model()` on a system without GPU's, it always fails in evaluate complaining that device `gpu0` is not available. This happens regardless of whether use_gpus=False or use_gpus=True. My platform is OSX 10.14.1 with latest version of madlib (1.17.0) and gpdb5. I think I've also seen this happen on CentOS in gpdb6, so I believe this is a bug that affects all platforms, but not entirely sure of that. Possibly specific to OSX or gpdb5. The problem happens in `internal_keras_eval_transition()` in `madlib_keras.py_in`. With `use_gpus=False`, it calls: ``` with K.tf.device(device_name): res = segment_model.evaluate(x_val, y_val) ``` with `device_name='/gpu0'` ``` I know this because I added a plpy.info statement to print `device_name` at the beginning of this function. I also printed the value of `use_gpus` on master before training begins: ``` INFO: 00000: use_gpus = False ``` This is what the error looks like: ``` INFO: 00000: device_name = /gpu:0 (seg1 slice1 127.0.0.1:25433 pid=90300) CONTEXT: PL/Python function "internal_keras_eval_transition" LOCATION: PLy_output, plpython.c:4773 psql:../run_fit_mult_iris.sql:1: ERROR: XX000: plpy.SPIError: tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a device for operation group_deps: Operation was explicitly assigned to /device:GPU:0 but available devices are [ /job:localhost/replica:0/task:0/device:CPU:0 ]. Make sure the device specification refers to a valid device. (plpython.c:5038) (seg0 slice1 127.0.0.1:25432 pid=90299) (plpython.c:5038) DETAIL: [[{{node group_deps}} = NoOp[_device="/device:GPU:0"](^loss/mul, ^metrics/acc/Mean)]] Traceback (most recent call last): PL/Python function "internal_keras_eval_transition", line 6, in <module> return madlib_keras.internal_keras_eval_transition(**globals()) PL/Python function "internal_keras_eval_transition", line 782, in internal_keras_eval_transition PL/Python function "internal_keras_eval_transition", line 1112, in evaluate PL/Python function "internal_keras_eval_transition", line 391, in test_loop PL/Python function "internal_keras_eval_transition", line 2714, in __call__ PL/Python function "internal_keras_eval_transition", line 2670, in _call PL/Python function "internal_keras_eval_transition", line 2622, in _make_callable PL/Python function "internal_keras_eval_transition", line 1469, in _make_callable_from_options PL/Python function "internal_keras_eval_transition", line 1351, in _extend_graph PL/Python function "internal_keras_eval_transition" CONTEXT: Traceback (most recent call last): PL/Python function "madlib_keras_fit_multiple_model", line 23, in <module> fit_obj = madlib_keras_fit_multiple_model.FitMultipleModel(**globals()) PL/Python function "madlib_keras_fit_multiple_model", line 42, in wrapper PL/Python function "madlib_keras_fit_multiple_model", line 216, in __init__ PL/Python function "madlib_keras_fit_multiple_model", line 230, in fit_multiple_model PL/Python function "madlib_keras_fit_multiple_model", line 270, in train_multiple_model PL/Python function "madlib_keras_fit_multiple_model", line 302, in evaluate_model PL/Python function "madlib_keras_fit_multiple_model", line 417, in compute_loss_and_metrics PL/Python function "madlib_keras_fit_multiple_model", line 739, in get_loss_metric_from_keras_eval PL/Python function "madlib_keras_fit_multiple_model" LOCATION: PLy_elog, plpython.c:5038 ``` > Without GPU's, FitMultipleModel fails in evaluate() > --------------------------------------------------- > > Key: MADLIB-1426 > URL: https://issues.apache.org/jira/browse/MADLIB-1426 > Project: Apache MADlib > Issue Type: Bug > Components: Deep Learning > Reporter: Domino Valdano > Priority: Major > > Whenever I try to run {{madlib_keras_fit_multiple_model()}} on a system > without GPU's, it always fails in evaluate complaining that device `gpu0` is > not available. This happens regardless of whether {{use_gpus=False}} or > use_gpus=True. > My platform is OSX 10.14.1 with latest version of madlib (1.17.0) and gpdb5. > I think I've also seen this happen on CentOS in gpdb6, so I believe this is a > bug that affects all platforms, but not entirely sure of that. Possibly > specific to OSX or gpdb5. > The problem happens in {{internal_keras_eval_transition()}} in > {{madlib_keras.py_in}}. > With {{use_gpus=False}}, it runs: > {{ > with K.tf.device(device_name): > res = segment_model.evaluate(x_val, y_val) > }} > ``` > I added a {{plpy.info}} statement to print {{device_name}} at the beginning > of this function. I also printed the value of {{use_gpus}} on master before > training begins. While {{use_gpus}} is set to false, the {{device_name}} on > the segments is set to {{/gpu:0}}. This is the bug (it should be set to > {{/cpu:0}}). > This is the error message that happens: > ``` > INFO: 00000: {{use_gpus = False}} > ... > INFO: 00000: device_name = /gpu:0 (seg1 slice1 127.0.0.1:25433 pid=90300) > CONTEXT: PL/Python function "internal_keras_eval_transition" > LOCATION: PLy_output, plpython.c:4773 > psql:../run_fit_mult_iris.sql:1: ERROR: XX000: plpy.SPIError: > tensorflow.python.framework.errors_impl.InvalidArgumentError: Cannot assign a > device for operation group_deps: Operation was explicitly assigned to > /device:GPU:0 but available devices are [ > /job:localhost/replica:0/task:0/device:CPU:0 ]. Make sure the device > specification refers to a valid device. (plpython.c:5038) (seg0 slice1 > 127.0.0.1:25432 pid=90299) (plpython.c:5038) > DETAIL: > [[{{node group_deps}} = NoOp[_device="/device:GPU:0"](^loss/mul, > ^metrics/acc/Mean)]] > Traceback (most recent call last): > PL/Python function "internal_keras_eval_transition", line 6, in <module> > return madlib_keras.internal_keras_eval_transition(**globals()) > PL/Python function "internal_keras_eval_transition", line 782, in > internal_keras_eval_transition > PL/Python function "internal_keras_eval_transition", line 1112, in evaluate > PL/Python function "internal_keras_eval_transition", line 391, in test_loop > PL/Python function "internal_keras_eval_transition", line 2714, in __call__ > PL/Python function "internal_keras_eval_transition", line 2670, in _call > PL/Python function "internal_keras_eval_transition", line 2622, in > _make_callable > PL/Python function "internal_keras_eval_transition", line 1469, in > _make_callable_from_options > PL/Python function "internal_keras_eval_transition", line 1351, in > _extend_graph > PL/Python function "internal_keras_eval_transition" > CONTEXT: Traceback (most recent call last): > PL/Python function "madlib_keras_fit_multiple_model", line 23, in <module> > fit_obj = madlib_keras_fit_multiple_model.FitMultipleModel(**globals()) > PL/Python function "madlib_keras_fit_multiple_model", line 42, in wrapper > PL/Python function "madlib_keras_fit_multiple_model", line 216, in __init__ > PL/Python function "madlib_keras_fit_multiple_model", line 230, in > fit_multiple_model > PL/Python function "madlib_keras_fit_multiple_model", line 270, in > train_multiple_model > PL/Python function "madlib_keras_fit_multiple_model", line 302, in > evaluate_model > PL/Python function "madlib_keras_fit_multiple_model", line 417, in > compute_loss_and_metrics > PL/Python function "madlib_keras_fit_multiple_model", line 739, in > get_loss_metric_from_keras_eval > PL/Python function "madlib_keras_fit_multiple_model" > LOCATION: PLy_elog, plpython.c:5038 > ``` -- This message was sent by Atlassian Jira (v8.3.4#803005)