memray / seq2seq-keyphrase

MIT License
318 stars 109 forks source link

运行 prediction 后的结果和 你之前保存下来的prediction结果不一样 #24

Closed whqwill closed 5 years ago

whqwill commented 6 years ago

你好, 我用你的代码运行了prediction部分, 载入的训练好的模型就是你提供的 experiments.keyphrase-all.one2one.copy.id=20170106-025508.epoch=4.batch=1000.pkl,然后测试数据是 seq2seq-keyphrase/dataset/keyphrase/baseline-data/kp20k,也就是直接运行了一遍没有改动任何参数,然后我evaluate的时候发现上述predict的结果和你提供的predict结果完全不一样,而且运行结果明显错误,预测值和文本完全不相关,然后显示的f1 score为0。 我想知道为什么会这样,难道是载入的模型不对还是需要另外再设置什么参数吗? 谢谢

memray commented 6 years ago

根据你的描述我很难确定问题。看起来就是哪里设置出了问题,比如data没有被正确的load。你能否debug检查一下?

whqwill commented 6 years ago

我刚刚重新下载了一次代码又运行了一遍

我在你代码里 keyphrase_copynet.py 的 601行加入了 这段代码: print ("source_str") print (source_str) print ()

            pre = []

            for phrase in prediction:
                pre.append(" ".join(["None" if idx2word.get(w) is None else idx2word.get(w) for w in phrase]))

            print ("pre")
            print (pre)
            print ()

然后我随便比较了一个比如dataset/keyphrase/baseline-data/kp20k/text/14073.txt,里面的text内容和上述输出一样,但是 输出的prediction的东西里面完全没有你提供的dataset/keyphrase/prediction/CopyRNN_present/kp20k/14073.txt.phrases 里面的短语,压根不一样。

所以 这里数据加载应该没问题,看上去加载的模型是有问题的,这样子的话,我是否需要重新训练模型?

memray commented 6 years ago

输出的内容是什么样呢?不一样在什么地方? 你看看config里这几行是不是appear-only。

whqwill commented 6 years ago

全部都设成 appear-only 的。

这是原文: reinforcement_NN learning_NN agentsNNS .. _NN reinforcement_NN learning_VBG (_NNP rl_NNP )_NNP is_VBZ learning_VBG through_IN direct_JJ experimentationNN .. it_PRP does_VBZ not_RB assume_VB the_DT existence_NN of_IN a_DT teacher_NN that_WDT provides_VBZ examples_NNS upon_IN which_WDT learning_NN of_IN a_DT task_NN takes_VBZ placeNN .. insteadRB ,, in_IN rl_NN experience_NN is_VBZ the_DT only_JJ teacherNN .. with_IN historical_JJ roots_NNS on_IN the_DT study_NN of_IN biological_JJ conditioned_JJ reflexesNNS ,, rl_NN attracts_VBZ the_DT interest_NN of_IN engineers_NNS and_CC computer_NN scientists_NNS because_IN of_IN its_PRP$ theoretical_JJ relevance_NN and_CC potential_JJ applications_NNS in_IN fields_NNS as_RB diverse_JJ as_IN operational_JJ research_NN and_CC intelligent_JJ roboticsNNS .. computationallyRB ,, rl_NN is_VBZ intended_VBN to_TO operate_VB in_IN a_DT learning_NN environment_NN composed_VBN by_IN two_CD subjects_NNS the_DT learner_NN and_CC a_DT dynamic_JJ processNN .. at_IN successive_JJ time_NN stepsNNS ,, the_DT learner_NN makes_VBZ an_DT observation_NN of_IN the_DT process_NN stateNN ,, selects_VBZ an_DT action_NN and_CC applies_VBZ it_PRP back_RB to_TO the_DT processNN .. its_PRP$ goal_NN is_VBZ to_TO find_VB out_RP an_DT action_NN policy_NN that_WDT controls_VBZ the_DT behavior_NN of_IN the_DT dynamic_JJ processNN ,, guided_VBN by_IN signals_NNS (_CD reinforcements_NNS )_NN that_WDT indicate_VBP how_WRB badly_RB or_CC well_RB it_PRP has_VBZ been_VBN performing_VBG the_DT required_JJ taskNN .. these_DT signals_NNS are_VBP usually_RB associated_VBN to_TO a_DT dramatic_JJ condition_NN eLS .. gNN .. ,_, accomplishment_NN of_IN a_DT subtask_FW (_FW reward_NN )_NN or_CC complete_JJ failure_NN (_NN punishment_NN )NN ,, and_CC the_DT learner_NN tries_VBZ to_TO optimize_VB its_PRP$ behavior_NN by_IN using_VBG a_DT performance_NN measure_NN (_CD a_DT function_NN of_IN the_DT received_JJ reinforcements_NNS )VBP .. the_DT crucial_JJ point_NN is_VBZ that_IN in_IN order_NN to_TO do_VB thatDT ,, the_DT learner_NN must_MD evaluate_VB the_DT conditions_NNS (_CD associations_NNS between_IN observed_VBN states_NNS and_CC chosen_JJ actions_NNS )_NN that_WDT led_VBD to_TO rewards_NNS or_CC punishmentsNNS .. starting_VBG from_IN basic_JJ conceptsNNS ,, this_DT tutorial_NN presents_VBZ the_DT many_JJ flavors_NNS of_IN rl_NN algorithmsNNS ,, develops_VBZ the_DT corresponding_JJ mathematical_JJ toolsNNS ,, assess_VBP their_PRP$ practical_JJ limitations_NNS and_CC discusses_VBZ alternatives_NNS that_WDT have_VBP been_VBN proposed_VBN for_IN applying_VBG rl_NN to_TO realistic_JJ tasksNNS ..

你给出的prediction: reinforcement learning learning intelligent robotics direct experimentation dynamic process action policy performance measure learning environment conditioned reflexes learning agents complete failure guided by signals reinforcement learning agents

我运行输出的内容:(以上的短语一个都没有出现,感觉像是输出的另一篇文章) [u'certain ', u'design ', u'high performance ', u'high performance local ', u'theoretic ', u'grouping ', u'compare ', u'complex ', u'considerable ', u'tools ', u'software defined ', u'defined ', u'generated ', u'psychological ', u'erlang ', u'indexing ', u'simulation ', u'existing ', u'software ', u'high programming ', u'approximate ', u'local ', u'author ', u'architecture ', u'high power certain ', u'effect ', u'energy certain ', u'consecutive ', u'software complex ', u'design theoretic ', u'definitions ', u'systems ', u'generation ', u'programming ', u'lies ', u'tolerant ', u'design certain ', '', u'circuits ', u'l ', u'visit ', u'disorders ', u'obtaining ', u'treatments ', u'composite ', u'aspects ', u'existing certain ', u'performance ', u'their ', u'modeling ', u'appropriately ', u'energy ', u'regularization ', u'measurement ', u'foreground ', u'find ', u'process ', u'synthesis certain ', u'regarding consecutive ', u'speaker ', u'control ', u'propose ', u'processes ', u'c ', u'organizations ', u'free ', u'consistent ', u'optimization ', u'computing ', u'application ', u'benchmark ', u'design tools ', u'widely ', u'efficiency ', u'system ', u'applications ', u'algorithms ', u'solutions specific certain ', u'considerable checked ', u'continue ', u'involvement ', u'themselves ', u'checked ', u'high programming certain ', u'like complex ', u'known ', u'after ', u'science ', u'high power aspects ', u'controller ', u'pedagogical ', u'derived ', u'model ', u'transition ', u'high performance certain ', u'considered regularization ', u'system design ', u'technical ', u'like certain ', u'various ', u'over ', u'turret ', u'incorporating local ', u'simulated local ', u'cavity ', u'services ', u'makes ', u'ecvs ', u'existing specific certain ', u'achieved ', u'specific certain ', u'high programming local ', u'solutions ', u'software tools ', u'like energy ', u'known generation ', u'series ', u'high high performance ', u'types ', u'suppliers ', u'classification complex ', u'cell ', u'high performance complex ', u'databases ', u'dota ', u'existing one ', u'field ', u'balance certain ', u'programming certain ', u'device ', u'like local ', u'ipcs ', u'lack ', u'high power local ', u'simulated design ', u'high order ', u'design effect ', u'software design ', u'design control ', u'function ', u'software certain ', u'like software ', u'hematoxylin ', u'high power ', u'( ', u'expression ', u'physical design ', u'camm ', u'functional ', u'existing complex ', u'design application ', u'current ', u'synthesis control ', u'programming efficiency ', u'high complex ', u'defined design ', u'minimum grouping ', u'classification transportation design ', u'camx ', u'human ', u'existing unstructured certain ', u'many certain ', u'go ', u'simulated certain ', u'incorporating generated ', u'high certain ', u'finite level ', u'architecture compare ', u'homogeneous design ', u'students systems ', u'incorporating certain ', u'existing local ', u'soaml ', u'high order certain ', u'processes design ', u'ecdh ', u'implementation local ', u'balance defined ', u'functions scheduling ', u'design , certain ', u'it ', u'design complex ', u'high efficiency ', u'high generated ', u'existing generation ', u'simulated complex ', u'high programming complex ', u'high high performance local ', u'functionality existing ', u'many control ', u'cell local ', u'types design ', u'high solutions ', u'existing unstructured forms ', u'high software ', u'high degree ', u'design propose ', u'high power design ', u'like software defined ', u'high performance aspects ', u'minimum design ', u'functions existing ', u'high expression ', u'like energy certain ', u'high software complex ', u'high performance software ', u'existing like certain ', u'high development ', u'design like certain ', u'high power tolerant ', u'high order local ', u'existing unstructured processes ', u'high programming systems ', u'high performance defined ', u'high specific certain ', u'c . ', u'software software ', u'high performance generated ', u'existing software ', u'high performance design ', u'high power complex ', u'existing like complex ', u'high programming generated ', u'high performance local certain ', u'existing unstructured local ', u'software defined certain ', u'design a certain ', u'finite level design ', u'design like complex ', u'high programming design ', u'like software complex ', u'high expression certain ', u'high solutions certain ', u'high programming software ', u'high software defined ', u'existing unstructured ', u'high performance software defined ', u'existing , certain ', u'high order aspects ', u'existing unstructured generated ', u'software defined design ', u'high high programming ', u'high programming aspects ', u'high solutions local ', u'high software certain ', u'design , ', u'design like ', u'high performance grouping ', u'high performance theoretic ', u'existing based certain ', u'high programming system ', u'high power processes ', u'high order complex ', u'existing like local ', u'high performance compare ', u'high high performance certain ', u'high d ', u'high performance tools ', u'existing like software defined ', u'high high power certain ', u'high performance propose ', u'high performance control ', u'high performance systems ', u'functions scheduling analysis ', u'high d certain ', u'high development certain ', u'software software defined ', u'high performance simulation ', u'design like software ', u'functions scheduling control ', u'existing unstructured software ', u'high expression local ', u'high performance over ', u'high programming difficulties ', u'existing like ', u'high performance energy ', u'high performance application ', u'design based certain ', u'high power grouping ', u'high performance modeling ', u'high performance software complex ', u'design ) certain ', u'high performance generation ', u'high power software ', u'high programming efficiency ', u'high order grouping ', u'high d local ', u'high degree certain ', u'high programming definitions ', u'high performance definitions ', u'high order design ', u'high software design ', u'high programming theoretic ', u'high performance find ', u'existing unstructured software defined ', u'high performance software certain ', u'high performance effect ', u'design like energy ', u'existing like software ', u'high performance science ', u'existing unstructured forms certain ', u'existing like energy ', u'high high performance complex ', u'high performance difficulties ', u'design , software ', u'high high programming certain ', u'existing unstructured forms modification ', u'high order processes ', u'existing unstructured forms generated ', u'high power , certain ', u'design like software defined ', u'high performance local tools ', u'high power software defined ', u'classification transportation software defined ', u'high order software ', u'existing based software defined ', u'high performance software design ', u'high high performance defined ', u'high programming software defined ', u'high performance local software ', u'high high software ', u'high performance energy certain ', u'high performance local generated ', u'high performance local design ', u'high performance local effect ', u'classification transportation software ', u'high performance local systems ', u'high order software defined ', u'high performance incorporating local ', u'high performance , certain ', u'high performance local complex ', u'high power design certain ', u'high high software complex ', u'high high performance aspects ', u'high power certain certain ', u'high high power aspects ', u'high performance a certain ', u'high performance like certain ', u'high high programming local ', u'design like energy certain ', u'high performance propose certain ', u'high performance incorporating certain ', u'high performance complex certain ', u'existing unstructured forms light ', u'high high performance software ', u'high performance local a certain ', u'high high performance design ', u'high performance local , certain ', u'high power local certain ', u'existing based software ', u'high performance software tools ', u'existing unstructured forms systems ', u'design ) high performance ', u'high power software complex ', u'high programming software complex ', u'high performance local theoretic ', u'high power aspects certain ', u'high high software defined ', u'high performance local application ', u'high power processes certain ', u'existing unstructured forms complex ', u'high programming local certain ', u'existing like energy certain ', u'high power software certain ', u'high high performance find ', u'high high power local ', u'high high performance generated ', u'design , high performance ', u'high performance software software ', u'high programming software certain ', u'high high high performance local ', u'high performance local system ', u'high high high performance ', u'existing like software complex ', u'high high programming complex ', u'classification transportation design certain ', u'existing unstructured processes certain ', u'existing unstructured software complex ', u'existing unstructured forms definitions ', u'high performance software detection ', u'design , software defined ', u'high performance design certain ', u'high order software complex ', u'high power aspects , certain ', u'existing unstructured high performance ', u'high performance like software ', u'high power certain a certain ', u'high high performance software defined ', u'high high performance local certain ', u'high performance software software defined ', u'high performance like software defined ', u'high performance local with databases ', u'high high performance , certain ', u'high performance local ) certain ', u'high performance local a design ', u'high power aspects a certain ', u'high performance software defined certain ', u'high performance local , ', u'high performance local a ', u'high performance local a software ', u'existing like software defined certain ', u'high high performance a certain ', u'high performance local with certain ', u'high performance software defined design ', u'high performance local a local ', u'high performance local a complex ', u'design ) high performance local ', u'high performance local software defined ', u'existing unstructured high performance local ', u'high performance local , software ', u'high performance local software certain ', u'high high performance software complex ', u'high performance like software complex ', u'high high performance software certain ', u'high performance local a software defined', u'existing like software defined design ', u'high performance local a generated ', u'high performance local ) ', u'high performance software software complex ', u'high performance software software certain ', u'design like software defined certain ', u'high performance local a high performance', u'high power aspects , ', u'high high high performance certain ', u'high performance local a compare ', u'high performance local , generated ', u'high performance local software complex ', u'high performance local , design ', u'design , high performance local ', u'high high performance software design ', u'high performance local effect certain ', u'high power software defined certain ', u'high performance local , local ', u'high performance software design certain ', u'high performance software complex certain ', u'high power certain a design ', u'high high high performance complex ', u'high performance local ) high performance', u'high performance local ) software ', u'high power certain a local ', u'high performance local ) generated ', u'high high performance a local ', u'high performance local , various ', u'high programming software defined certain ', u'high high performance local tools ', u'high performance local , complex ', u'existing based software defined certain ', u'existing unstructured software defined certain ', u'high high performance local a certain', u'high performance local a defined ', u'high power aspects , software ', u'high performance local design certain ', u'high performance local ) design ', u'high high performance complex certain ', u'high order software defined certain ', u'classification transportation software defined design ', u'high performance local a processes ', u'high performance local , high performance', u'high performance local software tools ', u'high performance local ) local ', u'design ) high performance certain ', u'high power software defined design ', u'existing based software defined design ', u'high high performance defined certain ', u'high high performance , local ', u'high performance like software certain ', u'design like software defined design ', u'classification transportation software defined certain ', u'high power certain a software ', u'high performance software software design ', u'high power aspects a design ', u'high performance local software design ', u'high high performance local design ', u'high performance software defined tools ', u'high performance local , software defined', u'high performance local a control ', u'high high performance local software ', u'high power aspects , design ', u'high performance local a systems ', u'high performance software defined software ', u'high performance local a tools ', u'high high performance local , certain', u'high performance local , processes ', u'high performance local a software complex', u'high performance local a various ', u'high performance local tools a certain', u'high power certain a software defined', u'high power certain a high performance', u'high performance local ) software defined', u'high high performance , software ', u'high high performance a software defined', u'high power aspects a software ', u'high high performance software software ', u'high performance local a solutions ', u'high performance local a software certain', u'high performance local a high power', u'high high performance a software ', u'high performance local a high order', u'high performance local certain a certain', u'high performance local ) high power', u'high high performance a high performance', u'high power aspects , high performance', u'high performance local a high programming', u'high power aspects a software defined', u'high performance local a like certain', u'high performance local ) high programming', u'high high performance local a design', u'high performance local ) high order', u'high high performance local a software', u'high high power certain a certain', u'high performance local a design certain', u'high performance local , high power', u'high performance local a high certain', u'high power aspects , software defined', u'high high performance , software defined', u'high performance software design a certain', u'high performance software tools a certain', u'high performance local design a certain', u'high high performance certain a certain', u'high performance software software defined certain', u'high performance local , high order', u'high performance local , software complex', u'high performance local , high programming', u'high performance local a high software', u'high high high performance local certain', u'high high performance local ) certain', u'high high performance local a ', u'high performance local , certain a', u'high performance local a software design', u'high high performance software software defined', u'high performance local tools a design', u'high high performance software defined certain', u'high high performance a software complex', u'high performance local a software tools', u'high performance local a certain a', u'high performance local , high ', u'high power certain a high power', u'high high performance local a local', u'high high performance local a high', u'high performance like software defined certain', u'high performance local a like software', u'high performance local a high generated', u'high power certain a high order', u'high performance local a certain certain', u'high power certain a software complex', u'high performance local , software certain', u'high performance local ) high generated', u'high performance software software defined design', u'high performance local tools a ', u'high performance local ) high certain', u'high performance local ) software complex', u'high performance local a high high', u'high performance local , high certain', u'high performance local ) high ', u'high high performance local , ', u'high performance local a design ,', u'high performance local certain a design', u'high performance local tools a software', u'high high performance local a complex', u'high power certain a software certain', u'high high high performance local a', u'high performance local a design a', u'high performance software defined software defined', u'high performance local , high generated', u'high power certain a high programming', u'high performance local a high solutions', u'high performance local ) high software', u'high performance local a software software', u'high performance local a software detection', u'high high performance software defined design', u'high performance local certain a ', u'high performance local , high software', u'high performance local a like complex', u'high performance local , certain certain', u'high performance local a high ', u'high performance local a like local', u'high power aspects , high order', u'high performance local a software generation', u'high performance local software defined certain', u'high power aspects , high power', u'high high performance local , software', u'high performance like software defined design', u'high performance local certain a software', u'high performance local a processes certain', u'high high performance a high power', u'high performance software defined design certain']

memray commented 6 years ago

这个预测的输出的词(high performance)和你给的text(reinforcement learning)完全没关系,因此我觉得还是data feeding哪里出错了。我现在没有环境运行Theano,因此不好意思我没法帮你调试找错。你看看喂进model的Tensor内容确实是这个文本本身吗(用vocab将word_id变回string)?也许loader哪里的逻辑出问题导致读取了错误的数据。

whqwill commented 6 years ago

你好,我发现问题所在了,数据载入的时候出问题了,因为你的test给的全是reinforcement_NN learning_NN agents_NNS 这样的文本,但是predict部分并没有预处理这样的文本,所以它直接把它当做无效词处理了,我去掉了后面的词性后,输出的预测值要合理很多。但是还是和你给出的不太一样。

这是原文: ['reinforcement', 'learning', 'agents', '.', '', 'reinforcement', 'learning', '(', 'rl', ')', 'is', 'learning', 'through', 'direct', 'experimentation', '.', 'it', 'does', 'not', 'assume', 'the', 'existence', 'of', 'a', 'teacher', 'that', 'provides', 'examples', 'upon', 'which', 'learning', 'of', 'a', 'task', 'takes', 'place', '.', 'instead', ',', 'in', 'rl', 'experience', 'is', 'the', 'only', 'teacher', '.', 'with', 'historical', 'roots', 'on', 'the', 'study', 'of', 'biological', 'conditioned', 'reflexes', ',', 'rl', 'attracts', 'the', 'interest', 'of', 'engineers', 'and', 'computer', 'scientists', 'because', 'of', 'its', 'theoretical', 'relevance', 'and', 'potential', 'applications', 'in', 'fields', 'as', 'diverse', 'as', 'operational', 'research', 'and', 'intelligent', 'robotics', '.', 'computationally', ',', 'rl', 'is', 'intended', 'to', 'operate', 'in', 'a', 'learning', 'environment', 'composed', 'by', 'two', 'subjects', 'the', 'learner', 'and', 'a', 'dynamic', 'process', '.', 'at', 'successive', 'time', 'steps', ',', 'the', 'learner', 'makes', 'an', 'observation', 'of', 'the', 'process', 'state', ',', 'selects', 'an', 'action', 'and', 'applies', 'it', 'back', 'to', 'the', 'process', '.', 'its', 'goal', 'is', 'to', 'find', 'out', 'an', 'action', 'policy', 'that', 'controls', 'the', 'behavior', 'of', 'the', 'dynamic', 'process', ',', 'guided', 'by', 'signals', '(', 'reinforcements', ')', 'that', 'indicate', 'how', 'badly', 'or', 'well', 'it', 'has', 'been', 'performing', 'the', 'required', 'task', '.', 'these', 'signals', 'are', 'usually', 'associated', 'to', 'a', 'dramatic', 'condition', 'e', '.', 'g', '.', ',', 'accomplishment', 'of', 'a', 'subtask', '(', 'reward', ')', 'or', 'complete', 'failure', '(', 'punishment', ')', ',', 'and', 'the', 'learner', 'tries', 'to', 'optimize', 'its', 'behavior', 'by', 'using', 'a', 'performance', 'measure', '(', 'a', 'function', 'of', 'the', 'received', 'reinforcements', ')', '.', 'the', 'crucial', 'point', 'is', 'that', 'in', 'order', 'to', 'do', 'that', ',', 'the', 'learner', 'must', 'evaluate', 'the', 'conditions', '(', 'associations', 'between', 'observed', 'states', 'and', 'chosen', 'actions', ')', 'that', 'led', 'to', 'rewards', 'or', 'punishments', '.', 'starting', 'from', 'basic', 'concepts', ',', 'this', 'tutorial', 'presents', 'the', 'many', 'flavors', 'of', 'rl', 'algorithms', ',', 'develops', 'the', 'corresponding', 'mathematical', 'tools', ',', 'assess', 'their', 'practical', 'limitations', 'and', 'discusses', 'alternatives', 'that', 'have', 'been', 'proposed', 'for', 'applying', 'rl', 'to', 'realistic', 'tasks', '.']

我运行的输出: (并没有出现reinforcement learning,intelligent robotics 这种短语,虽然有单个的reinforcement, learning,intelligent 出现了 )

[u'reflexes ', u'engineers ', u'accomplishment ', u'rl experience ', u'historical ', u'rl attracts ', u'rl ', u'assess ', u'dynamic process ', u'realistic ', u'assess their ', u'conditioned reflexes ', u'rl attracts the interest ', u'starting ', u'subtask ( ', u'guided process ', u'computer scientists ', u'performance ', u'signals minimizes ', u'corresponding mathematical ', u'biological conditioned reflexes ', u'corresponding computation ', u'dramatic ', u'teacher ', u'learner tries ', u'corresponding local ', u'basic ', u'attracts the interest ', u'diverse ', u'rl algorithms ', u'existence ', u'rewards ', u'existence process ', u'find ', u'tries investigations ', u'how ', u'corresponding process ', u'attracts ', u'controls ', u'punishments ', u'dynamic ', u'not ', u'received design ', u'corresponding information ', u'many find ', u'e ', u'g ', u'guided ', u'learning error ', u'tries ', u'rl rl ', u'process ', u'operate ', u'rl reflexes ', u'rl process ', u'subtask ', u'signals starting ', u'potential error ', u'descent ', u'upon ', u'corresponding find ', u'rl learning ', u'tries tries ', u'task process ', u'experience ', u'learning find ', u'badly ', u'rl makes ', u'examples ', u'algorithms ', u'rl scientists ', u'rl computer ', u'develops ', u'learner tries tries ', u'many process ', u'corresponding assess ', u'flavors ', u'rl not ', u'conditioned ', u'reinforcement ', u'examples upon ', u'task ', u'only teacher ', u'learning starting ', u'( rl ', u'tries to ', u'learner tries to ', u'how badly ', u'reflexes process ', u'many flavors ', u'assess process ', u'( ', u'intelligent ', u'do ', u'tries reflexes ', u'observation process ', u'chosen ', u'signals find ', u'failure ', u'corresponding starting ', u'rl how ', u'learner ', u'rl guided ', u'conditioned guided ', u'subjects ', u'performance measure ', u'instead problem ', u'rl attracts attracts ', u'function ', u'upon process ', u'process low ', u'rewards algorithm ', u'existence starting ', u'g trade ', u'learner tries reflexes ', u'learning process ', u'g analysis ', u'complete failure ', u'conditioned process ', u'corresponding how ', u'computer ', u'many starting ', u'biological reflexes ', u'rewards algorithms ', u'rewards process ', u'received how ', u'task intelligent ', u'learning ', u'corresponding learning error ', u'task find ', u'learner tries investigations ', u'experimentation ', u'received reinforcements ', u'tries starting ', u'subtask starting ', u'learning how ', u'usually associated ', u'received find ', u'corresponding information systems ', u'corresponding basic ', u'rl potential ', u'rl attracts attracts the interest ', u'guided engineers ', u'task teacher ', u'conditioned robust ', u'rl attracts experience ', u'learning ( ', u'learning intelligent ', u'instead ', u'conditions ', u'performance analysis ', u'tries realistic ', u'instead problems ', u'learner starting ', u'g process ', u'assess their process ', u'performance evaluation ', u'task how ', u'biological conditioned ', u'rl intelligent ', u'existence find ', u'rl attracts rl ', u'corresponding algorithms ', u'rl diverse ', u'computer engineers ', u'rl reinforcements ', u'rl subjects ', u'task scientists ', u'learner tries starting ', u'tries g ', u'rl dynamic ', u'existence assess ', u'rl assess ', u'rl operate ', u'received dynamic ', u'rl attracts interest ', u'chosen actions ', u'discusses rl ', u'subtask ( rl ', u'corresponding information low ', u'existence potential error ', u'corresponding develops ', u'corresponding g ', u'process state ', u'( ( ', u'corresponding dynamic ', u'( reinforcements ', u'subtask ( reward ', u'rl attracts process ', u'existence learning error ', u'learner task ', u'subtask tries ', u'rl task ', u'rl attracts reflexes ', u'attracts the rl ', u'attracts rl ', u'corresponding their ', u'rl attracts learning ', u'many tries ', u'task dynamic ', u'provides examples upon ', u'existence learning ', u'attracts the guided ', u'failure ( ', u'complete various error ', u'measure ( ', u'discusses tries ', u'corresponding information system ', u'existence tries ', u'rl attracts the guided ', u'attracts attracts ', u'learner tries realistic ', u'corresponding guided process ', u'rl attracts how ', u'rl attracts scientists ', u'existence tries investigations ', u'corresponding learning ', u'signals ( ', u'learner tries g ', u'rl attracts realistic ', u'corresponding their process ', u'biological conditioned rl ', u'learning ( rl ', u'rl attracts the rl ', u'rl attracts e ', u'corresponding dynamic process ', u'provides examples ', u'biological rl ', u'existence corresponding mathematical ', u'rl attracts computer ', u'rl attracts g ', u'rl attracts point ', u'rl attracts teacher ', u'corresponding guided ', u'biological conditioned process ', u'rl attracts the rl attracts the', u'rl attracts the rl attracts ', u'rl computer scientists ', u'tries tries reflexes ', u'biological conditioned guided ', u'chosen actions process ', u'attracts attracts the interest ', u'biological conditioned robust ', u'rl corresponding mathematical ', u'rl attracts makes ', u'learner tries optimize ', u'rl attracts guided ', u'attracts the computer ', u'tries tries tries ', u'learner tries tries tries ', u'subtask ( tries ', u'rl attracts the engineers ', u'learner tries tries reflexes ', u'rl attracts the computer ', u'rl attracts the ', u'learner tries to g ', u'rl attracts the learning ', u'biological conditioned tries ', u'rl attracts attracts attracts ', u'rl attracts the process ', u'rl attracts rl attracts ', u'rl attracts the operate ', u'learner tries tries to ', u'rl attracts the assess ', u'rl attracts the makes ', u'rl attracts the existence ', u'rl rl attracts ', u'learner tries tries starting ', u'attracts the learner ', u'rl attracts attracts interest ', u'rl attracts the rl attracts attracts', u'learner tries to starting ', u'learner tries tries investigations ', u'learner tries must ', u'learner tries to optimize ', u'existence tries tries ', u'rl attracts to interest ', u'attracts the rl attracts ', u'attracts the rl experience ', u'rl attracts the operational ', u'rl attracts the realistic ', u'rl attracts attracts attracts the interest', u'rl attracts attracts the rl attracts', u'learner tries to makes ', u'biological conditioned reflexes reflexes ', u'rl attracts the reflexes ', u'learner tries tries g ', u'rl attracts the do ', u'rl attracts to ', u'rl attracts the fields ', u'rl attracts the diverse ', u'rl attracts the attracts ', u'rl attracts the dynamic ', u'rl attracts the interest makes ', u'rl attracts computer scientists ', u'learner tries to learning ', u'rl attracts the intelligent ', u'learner tries to tries ', u'rl attracts the find ', u'learner tries tries realistic ', u'rl attracts the task ', u'rl attracts the learner ', u'rl attracts attracts learning ', u'rl attracts the performing ', u'biological conditioned reflexes process ', u'rl attracts the g ', u'rl attracts rl experience ', u'rl attracts attracts rl ', u'rl attracts attracts experience ', u'biological conditioned tries reflexes ', u'rl attracts the rl experience ', u'rl attracts attracts e ', u'rl attracts the experience ', u'rl attracts the observation ', u'rl attracts the interest process ', u'provides examples upon process ', u'attracts the rl attracts the interest', u'rl attracts rl attracts the interest', u'learner tries to find ', u'rl attracts attracts the ', u'rl attracts the develops ', u'rl attracts the potential ', u'rl attracts is interest ', u'rl attracts the point ', u'rl rl attracts the interest ', u'rl attracts the interest engineers ', u'rl attracts the interest attracts ', u'rl attracts the how ', u'learner tries to reflexes ', u'rl attracts is learning ', u'learner tries tries diverse ', u'rl attracts the e ', u'rl attracts the received ', u'biological conditioned reflexes , rl attracts', u'learner tries to process ', u'attracts the computer scientists ', u'learner tries to e ', u'learner tries tries tries tries tries', u'rl attracts the interest scientists ', u'learner tries tries historical ', u'rl attracts the computer scientists ', u'rl attracts the conditions ', u'rl attracts rl learning ', u'rl attracts the action ', u'learner tries tries point ', u'rl attracts learning error ', u'learner tries tries optimize ', u'learner tries to computer ', u'rl attracts attracts makes ', u'learner tries tries tries tries ', u'rl attracts the function ', u'rl attracts learning how ', u'tries tries tries reflexes ', u'learner tries tries find ', u'learner tries to rl ', u'rl attracts attracts g ', u'rl attracts attracts the rl ', u'rl attracts the interest teacher ', u'rl attracts guided process ', u'rl attracts attracts point ', u'rl attracts the only ', u'learner tries tries rl ', u'rl attracts attracts reflexes ', u'rl attracts attracts realistic ', u'rl attracts attracts computer ', u'rl attracts the tries ', u'learner tries to diverse ', u'rl attracts rl attracts attracts the', u'tries tries tries tries ', u'rl attracts attracts the guided ', u'learner tries to realistic ', u'rl attracts the accomplishment ', u'learner tries to existence ', u'learner tries tries accomplishment ', u'rl attracts the corresponding ', u'learner tries tries tries reflexes ', u'rl attracts attracts how ', u'learner tries to operate ', u'rl attracts the must ', u'rl attracts the interest learning ', u'rl attracts to computer ', u'rl attracts attracts the engineers ', u'rl attracts the learner makes ', u'attracts attracts the rl ', u'subtask ( tries tries ', u'rl attracts is rl ', u'rl attracts attracts the makes ', u'rl attracts the optimize ', u'rl attracts to rl attracts ', u'learner tries to . ', u'rl attracts the interest reflexes ', u'rl attracts the interest how ', u'rl attracts attracts the computer ', u'learner tries tries to g ', u'rl attracts the existence makes ', u'rl attracts the to ', u'rl attracts the signals ', u'subtask ( reward tries ', u'rl attracts attracts to ', u'rl attracts the interest not ', u'rl attracts the attracts the interest', u'rl attracts the study ', u'rl attracts rl rl ', u'rl attracts the interest point ', u'rl attracts attracts the existence ', u'rl attracts the rl makes ', u'rl attracts the interest g ', u'rl attracts the limitations ', u'biological conditioned tries to ', u'rl attracts the the interest ', u'rl attracts the interest function ', u'rl attracts the interest rl ', u'rl attracts attracts attracts attracts the', u'rl attracts to rl attracts attracts', u'rl attracts the interest experience ', u'rl attracts to rl ', u'learner tries tries tries to ', u'rl attracts the interest guided ', u'rl attracts attracts rl attracts the', u'rl attracts the well ', u'rl attracts to rl attracts the', u'rl attracts the guided process ', u'rl attracts the interest attracts the', u'biological conditioned rl attracts ', u'rl attracts rl attracts attracts ', u'rl experience is rl ', u'learner tries tries tries g ', u'rl attracts attracts rl attracts ', u'biological conditioned reflexes rl attracts attracts', u'rl attracts the rl attracts to', u'rl attracts the corresponding mathematical ', u'rl attracts attracts the learning ', u'rl attracts attracts the interest makes', u'rl attracts the interest intelligent ', u'learner tries to tries to ', u'rl attracts the interest performing ', u'rl attracts rl attracts attracts attracts', u'rl attracts the interest computer ', u'rl attracts attracts attracts attracts ', u'rl attracts the interest upon ', u'learner tries tries tries starting ', u'learner tries to tries tries tries', u'rl attracts attracts the interest attracts', u'rl attracts the rl attracts rl', u'rl attracts attracts the operate ', u'learner tries tries tries tries to', u'rl rl attracts attracts the interest', u'rl attracts the rl attracts interest', u'rl attracts the interest ( ', u'rl attracts attracts the attracts ', u'learner tries to tries tries ', u'rl attracts attracts attracts the rl', u'learner tries to tries tries to', u'rl attracts the rl attracts makes', u'learner tries tries to tries tries', u'rl attracts attracts rl attracts attracts', u'rl experience is rl attracts ', u'rl attracts the rl attracts experience', u'rl experience is rl attracts the', u'learner tries tries tries tries reflexes', u'learner tries to optimize makes ', u'rl attracts attracts the learner ', u'rl attracts rl attracts the rl', u'rl attracts the attracts attracts ', u'rl attracts the rl attracts is', u'learner tries tries to optimize ', u'rl attracts the rl attracts e', u'biological conditioned reflexes rl attracts the', u'learner tries tries tries to tries', u'attracts the rl attracts the rl', u'learner tries tries to tries to', u'learner tries to tries to tries', u'learner tries to learner tries to', u'learner tries to rl attracts attracts', u'biological conditioned rl attracts the interest', u'rl attracts the attracts attracts the', u'rl attracts attracts the attracts the', u'rl attracts the engineers attracts the', u'learner tries tries to tries ', u'rl attracts rl rl attracts attracts', u'attracts attracts the rl attracts ', u'rl attracts the rl attracts scientists', u'rl attracts the rl attracts )', u'rl attracts attracts attracts the ', u'attracts attracts the rl attracts the', u'rl attracts attracts the interest engineers', u'learner tries to learner tries tries', u'rl attracts the interest a interest', u'learner tries tries tries tries g', u'rl attracts attracts the interest scientists', u'rl attracts the interest attracts attracts', u'rl attracts rl rl attracts the', u'rl attracts the interest rl attracts', u'rl attracts the rl attracts learning', u'rl attracts the rl attracts performing', u'learner tries to optimize tries to', u'rl attracts the interest , rl', u'rl attracts attracts the interest process', u'rl attracts attracts the interest a', u'rl attracts the rl attracts a', u'rl attracts attracts the existence makes', u'biological conditioned reflexes rl attracts ', u'rl attracts the engineers attracts ', u'rl attracts attracts the learner makes', u'learner tries tries tries to g', u'learner tries to rl attracts ', u'rl attracts attracts the computer scientists', u'rl attracts rl rl attracts ', u'learner tries to rl attracts the', u'rl attracts attracts the interest learning', u'learner tries to tries to optimize', u'learner tries to optimize tries tries', u'rl attracts attracts the interest is', u'rl attracts attracts the interest teacher', u'rl attracts attracts the interest (', u'attracts the rl attracts the existence', u'biological conditioned rl attracts the rl', u'biological conditioned reflexes , rl ', u'rl attracts attracts the rl experience', u'rl attracts the rl attracts .', u'rl attracts the rl attracts an', u'rl attracts attracts attracts attracts attracts', u'learner tries tries tries to optimize', u'rl attracts attracts attracts the existence', u'rl attracts attracts the attracts attracts', u'rl attracts attracts the interest performing', u'rl attracts the rl attracts computer', u'rl attracts attracts the interest rl', u'biological conditioned reflexes , rl experience', u'rl attracts the engineers attracts attracts', u'learner tries to optimize tries ', u'rl attracts the rl attracts point', u'rl attracts the interest is the', u'learner tries tries tries reflexes reflexes', u'rl attracts the interest is rl', u'learner tries to learner tries ', u'rl attracts attracts attracts the makes', u'rl experience is rl attracts attracts', u'rl attracts the interest is learning', u'learner tries tries tries tries (', u'rl attracts the interest in interest', u'rl attracts attracts the interest point', u'learner tries tries tries to rl', u'learner tries tries tries tries optimize', u'rl attracts the interest is ', u'learner tries to tries to .', u'rl attracts attracts the interest g', u'rl attracts the interest a rl', u'rl attracts the interest is interest', u'learner tries tries tries tries historical', u'rl attracts the rl attracts actions', u'rl attracts the interest , interest', u'learner tries tries tries to existence', u'learner tries to tries to makes', u'rl attracts attracts the interest ,', u'rl attracts attracts the rl makes', u'learner tries tries tries to makes', u'rl attracts the rl attracts chosen', u'rl attracts the rl attracts associations', u'rl attracts rl attracts the attracts', u'rl attracts attracts the interest function', u'rl attracts the rl attracts teacher', u'learner tries tries to optimize tries', u'learner tries tries tries tries rl', u'rl attracts attracts attracts the engineers', u'rl attracts the rl attracts reflexes', u'rl attracts the attracts the rl', u'rl attracts attracts the interest and', u'learner tries to tries to existence', u'rl attracts rl attracts the existence', u'learner tries to tries to g', u'rl attracts attracts the interest in', u'rl attracts the interest a learning']

memray commented 6 years ago

恩我eval的时候用的带postag的数据,predict是用的另外一个pipeline对testdata进行预处理的。这个结果看起来是非常noisy,我也说不清具体是什么原因导致了这个差别,可能的原因:(1)我有过滤掉大量只包含一个单词的phrase,只保留一个加入到最后输出;(2)如之前所说可能是输入和我的pipeline有区别,比如e.g.我印象中是整个加入vocab的。你有看在其他数据上的效果吗?是否都和这个一样?

whqwill commented 6 years ago

有一些数据会好一点,但是我测了一下前五十个数据,evaluate出来的top 5 或者top 10 的f1 都只有0.1左右

whqwill commented 6 years ago

还有就是 过滤大量一个单词的phrase具体是在哪里做的呀, 还有整个加入vocab是什么意思?

memray commented 6 years ago

e.g.在我的vocab里应该是一个词,而不是['e', '.', 'g', '.']4个词。不过这个应该不是最关键的原因。