batra-mlp-lab / visdial-rl

PyTorch code for Learning Cooperative Visual Dialog Agents using Deep Reinforcement Learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Somehow -visdomServer is overwritten in an unexpected way

okisy opened this issue · comments

commented

This is related to #2
I ran
python -m visdom.server -p 8097
in order to use visdom.

After that, I ran
python evaluate.py -useGPU \ -startFrom checkpoints/abot_sl_ep60.vd \ -qstartFrom checkpoints/qbot_sl_ep60.vd \ -evalMode ABotRank QBotRank \ -visdomServer http://127.0.0.1 \ -visdomServerPort 8097 \ -visdomEnv ABotRank-QBotRank
Even though I clearly designated which server to use and the server port number, somehow they are overwritten as follows. How can I fix this?

{'CELossCoeff': 200,
'batchSize': 20,
'beamSize': 1,
'ckpt_iterid': 152160,
'ckpt_lRate': 4.999614155150613e-05,
'cocoDir': '',
'cocoInfo': '',
'continue': True,
'decoder': 'gen',
'dropout': 0.0,
'embedSize': 300,
'enableVisdom': 1,
'encoder': 'hre-ques-lateim-hist',
'evalModeList': ['ABotRank', 'QBotRank'],
'evalSplit': 'val',
'evalTitle': 'eval',
'featLossCoeff': 1000,
'freezeQFeatNet': 0,
'imgEmbedSize': 300,
'imgFeatureSize': 4096,
'imgNorm': 1,
'inputImg': 'data/visdial/data_img.h5',
'inputJson': 'data/visdial/chat_processed_params.json',
'inputQues': 'data/visdial/chat_processed_data.h5',
'learningRate': 0.001,
'lrDecayRate': 0.9997592083,
'minLRate': 5e-05,
'numEpochs': 65,
'numLayers': 2,
'numRounds': 10,
'numWorkers': 2,
'qdecoder': 'gen',
'qencoder': 'hre-ques-lateim-hist',
'qstartFrom': 'checkpoints/qbot_sl_ep60.vd',
'randomSeed': 32,
'rlAbotReward': 1,
'rnnHiddenSize': 512,
'saveName': 'debug-final5_sl-qbot',
'savePath': 'checkpoints/debug-final5_sl-qbot',
'startFrom': 'checkpoints/abot_sl_ep60.vd',
'trainMode': 'sl-abot',
'useCurriculum': 1,
'useGPU': True,
'useHistory': True,
'useIm': 'late',
'verbose': 1,
'visdomEnv': 'ABotRank-QBotRank',
'visdomServer': 'http://walle.cc.gatech.edu',
'visdomServerPort': 8893,
'vocabSize': 7826}

Thanks for pointing this out, the visdom parameters were being overwritten in evaluate.py with the ones loaded from checkpoint. I have added those parameters to the list of excluded parameters in 58110e2 to prevent this. Let me know if the issue persists.