"KeyError: 'program_parsed'" when running the program

[dlshu]

Hi. I tried to test the pretrained model using the test data from CLEVR dataset but had a error 'KeyError: 'program_parsed''. The command I typed in is as follows.

jac-crun 0 scripts/main.py --mode run-experiment --task CLEVR --model VCML --experiment synonym_generalization --log_dir ../data/log --data_dir ../data/ --pretrained --in_epoch test

and the error message shown is as follows.

0:00:52.63 | => epoch 0
|
0:00:52.85 | -----> Testing
Traceback (most recent call last):
File "scripts/main.py", line 44, in
main()
File "scripts/main.py", line 34, in main
run_experiment.run(args)
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/run_experiment.py", line 181, in run
processes[0]._target(*processes[0]._args)
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/run_experiment.py", line 151, in ready_go
train(coach, args)
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/utils/train.py", line 223, in train
run_epoch(coach, args, coach.epoch)
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/utils/train.py", line 192, in run_epoch
True, False
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/utils/train.py", line 142, in any_epoch
inner()
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/utils/train.py", line 128, in inner
loss, outputs = run_batch(data, model, args)
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/scripts/utils/train.py", line 82, in run_batch
losses, outputs, debugs, objects = model(data)
File "/home/dule/anaconda3/envs/VCML/lib/python3.6/site-packages/torch/nn/modules/module.py", line 547, in call
result = self.forward(*input, **kwargs)
File "/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/models/model/vcml_model.py", line 87, in forward
program = data['program_parsed']
KeyError: 'program_parsed'

/home/dule/Desktop/NN/AAAI_21/nuero-symbolic_ai/VCML/models/model/vcml_model.py(87)forward()
85 program_encoded = data['program_encoded']
86 else:
---> 87 program = data['program_parsed']
88 program_encoded = data['program_parsed_encoded']
89

ipdb>

Any suggestions how to fix this issue? Thank you very much!

[anette123]

Getting the same error. If I modify 'inner()' function in train.py when 'is_train=False' to estimate 'program parsed' and 'program_parsed_encoded' just like when 'is_train=True' , the code will proceed further, but will fail as the downloaded parser cannot parse test questions. That is, the first 10 questions from the test set:

# Out[93]: TEST QUESTIONS
# ['Is small a synonym of small?',
#  'Is sphere a synonym of sphere?',
#  'Is shiny a synonym of shiny?',
#  'Is sphere a synonym of shiny?',
#  'Is sphere a synonym of small?',
#  'Is shiny a synonym of shiny?',
#  'Is shiny a synonym of sphere?',
#  'Is small a synonym of shiny?',
#  'Is shiny a synonym of small?',
#  'Is small a synonym of small?']

And the parser will process them in the following way (showing the first two):

# Out[96]:
# [[{'operation': '<END>', 'argument': 'small'},
#   {'operation': '<END>', 'argument': 'small'},
#   {'operation': '<END>', 'argument': '<END>'},
#   {'operation': '<END>', 'argument': '<END>'},
#   {'operation': '<END>', 'argument': '<END>'}],
#  [{'operation': '<END>', 'argument': 'sphere'},
#   {'operation': '<END>', 'argument': 'sphere'},
#   {'operation': '<END>', 'argument': '<END>'},
#   {'operation': '<END>', 'argument': '<END>'},
#   {'operation': '<END>', 'argument': '<END>'}],

Could you please have a look at this? I suspect the parser provided cannot process meta concept words like 'synonym'..

[anette123]

Update to my previous comment: have been testing the parser separately, and while in parser.tools.operation.records it contains 'synonym', it is always predicting '10' for operation, which is '' for the test questions. Below listing parser operatons.

['select_concept',
 'synonym',
 'isinstanceof',
 'samekind',
 'select_object',
 'filter',
 'exist',
 'classify',
 '<NULL>',
 '<START>',
 '<END>',
 '<UNK>']

[coldmanck]

Same issue here. @anette123 may I know if you have solved this problem?

[coldmanck]

I'd also like to call for the authors' attention to this issue @vacancy @Glaciohound