20260127_試行錯誤の結晶を1冊に。著者が解説先輩データサイエンティストからの指南書 / author's_commentary_ds_instructions_guide

ࢼߦࡨޡͷ݁থΛ1࡭ʹɻஶऀ͕ղઆ  ઌഐσʔλαΠΤϯςΟετ͔Βͷࢦೆॻ Junki Asano / Toma Tanaka / Mizuho Yanagi
2026/1/27 Forkwell Library #118

Copyright © Junki Asano / Toma Tanaka / Mizuho Yanagi
All Rights Reserved 1ষɿ࣮຿Ͱੜ͖ൈͨ͘ΊͷΤϯδχΞϦϯάεΩϧ 2ষɿ؀ڥߏங 3ষɿίʔυͷ඼࣭؅ཧ 4ষɿσʔλͷ඼࣭֬ೝ 5ষɿػցֶशϞσϧͷ࣮ݧ؅ཧ 6ষɿϓϩτλΠϓ։ൃ 2 ઌഐσʔλαΠΤϯςΟετ͔Βͷࢦೆॻʢٕज़ධ࿦ࣾ; 2025ʣ ࣮ફతͳΤϯδχΞϦϯάεΩϧΛ·ͱΊͨॻ੶Λࣥච͠·ͨ͠ʂ

All Rights Reserved 3 ࣗݾ঺հ ઙ໺ ७قʢ@nash_efpʣ גࣜձࣾϒϨΠϯύου σʔλαΠΤϯςΟετ ాத ౙഅʢ@fuyu_quantʣ ىۀՈʢʁʣ 㛲 ઘึʢ@suikabar_umaiʣ גࣜձࣾλΠϛʔ σʔλαΠΤϯςΟετ 2019೥ 2022೥ 2025೥ ϒϨΠϯύου ʹೖࣾ λΠϛʔʹೖࣾ LLMΞϓϦέʔγϣϯ   ։ൃͳͲ ϒϨΠϯύου ʹೖࣾ ϒϨΠϯύου ʹೖࣾ લ৬Ͱ DSʹͳΔ ϓϩδΣΫτϚωδϝϯτ σʔλϚωδϝϯτ ෼ੳɺޮՌݕূͳͲ ىۀ४උͱ ݚڀ׆ಈ

7 ॻ੶ͷղઆ

All Rights Reserved ໨࣍ 8 ୈ4ষɹσʔλͷ඼࣭֬ೝ ɹ4.1ɹσʔλͷ඼࣭֬ೝͷॏཁੑ ɹ4.2ɹ෼ੳ։࢝લͷσʔλ֬ೝ ɹ4.3ɹPanderaʹΑΔσʔλϑϨʔϜͷόϦσʔγϣϯ ɹ4.4ɹσʔλ඼࣭؅ཧͷܧଓతͳऔΓ૊Έ ୈ5ষɹػցֶशϞσϧͷ࣮ݧ؅ཧ ɹ5.1ɹ࣮ݧ؅ཧͷҙٛ ɹ5.2ɹHydraʹΑΔύϥϝʔλ؅ཧ ɹ5.3ɹMLflowʹΑΔ࣮ݧͷൺֱ ୈ6ষɹϓϩτλΠϓ։ൃ ɹ6.1ɹϓϩτλΠϓ։ൃͷҙٛ ɹ6.2ɹStreamlitʹΑΔϓϩτλΠϓ։ൃ ୈ1ষɹ࣮຿Ͱੜ͖ൈͨ͘ΊͷΤϯδχΞϦϯάεΩϧ ɹ1.1ɹσʔλαΠΤϯςΟετΛऔΓר͘؀ڥͷมભ ɹ1.2ɹϓϩδΣΫτͰٻΊΒΕΔΤϯδχΞϦϯάεΩϧ ୈ2ষɹ؀ڥߏங ɹ2.1ɹ෼ੳͷ౔୆ͱͯ͠ͷ؀ڥߏங ɹ2.2ɹϦϙδτϦͷߏ଄Λ੔͑Δ ɹ2.3ɹVS CodeͰͷ։ൃ؀ڥͷ੔උ ɹ2.4ɹDev Container ʹΑΔԾ૝؀ڥߏங ɹ2.5ɹuv ʹΑΔύοέʔδ؅ཧ ୈ3ষɹίʔυͷ඼࣭؅ཧ ɹ3.1ɹNotebook͚ͩͷσʔλαΠΤϯςΟετ͔Βͷଔۀ ɹ3.2ɹίʔυ඼࣭ͱ͸ ɹ3.3ɹ඼࣭ͷߴ͍ίʔυͱͦͷ࣮ݱํ๏ ɹ3.4ɹίʔυϨϏϡʔʹΑΔ඼࣭؅ཧ

9 1ষ ࣮຿Ͱੜ͖ൈͨ͘Ίͷ ΤϯδχΞϦϯάεΩϧ

All Rights Reserved 18 σʔλαΠΤϯςΟετ͕࣮຿Ͱ௚໘͢Δ՝୊ ʮͱʹ͔͘ಈ͚͹Α͍ʯͰ   ॻ͍ͨίʔυʹޙ͔Βۤ͠Ή ࣮຿Ͱѻ͏σʔλͷɺ༧ظͤ͵ όά΍ܽଛʹۤ࿑͢Δ ࢼߦࡨޡΛ͢Δ΋ɺ্ख͍ͬ͘ ͨաڈͷ݁ՌΛ࠶ݱͰ͖ͳ͍ ෼ੳ݁Ռ΍Ϟσϧ͕   ҙࢥܾఆ·Ͱ݁ͼ͔ͭͳ͍ ରԠํ๏Λ֤ষͰղઆ 3ষɿίʔυͷ඼࣭؅ཧ 4ষɿσʔλͷ඼࣭֬ೝ 5ষɿػցֶशϞσϧͷ࣮ݧ؅ཧ 6ষɿϓϩτλΠϓ։ൃ

All Rights Reserved 20 ʮ࣮຿ͰॳΊͯ௚໘͢Δ՝୊ʯ͕ൃੜ͢ΔݪҼ σʔλαΠΤϯςΟετ͕   εΩϧΛशಘ͢Δϓϩηε σʔλαΠΤϯεϓϩδΣΫτ ಛ༗ͷ೉͠͞ ίʔυͷ඼࣭΍ӡ༻؅ཧΛݟਾ࣮͑ͨ૷ ΛֶͿ༏ઌॱҐ͕௿͍ ·ͣ͸ʮཧ࿦Λਖ਼࣮͘͠૷Ͱ͖Δ͜ͱʯ ͕༏ઌ͞ΕΔ ଟ͘ͷσʔλαΠΤϯςΟετ͕ॳΊʹֶͿ͜ͱ ͸લॲཧ΍ϞσϧߏஙɺͦΕΒΛNotebookͰ࣮ߦ ࢼߦࡨޡ͕લఏʹ͋Γɺෆ࣮֬ੑ͕ߴ͍ ݩσʔλʹෆඋ͕͋ͬͨΓɺޡͬͨ݁ՌΛ   ग़ྗͯ͠΋Τϥʔʹ͸ͳΒͳ͍ ग़ྗͨ݁͠ՌΛҙࢥܾఆ΍αʔϏεʹ   ద༻͢ΔͨΊʹ͸ɺॲཧ΍޻෉͕ඞཁͰ͋Δ

All Rights Reserved ֤ষͰରԠࡦΛղઆ 21 ୈ4ষɹσʔλͷ඼࣭֬ೝ ɹ4.1ɹσʔλͷ඼࣭֬ೝͷॏཁੑ ɹ4.2ɹ෼ੳ։࢝લͷσʔλ֬ೝ ɹ4.3ɹPanderaʹΑΔσʔλϑϨʔϜͷόϦσʔγϣϯ ɹ4.4ɹσʔλ඼࣭؅ཧͷܧଓతͳऔΓ૊Έ ୈ5ষɹػցֶशϞσϧͷ࣮ݧ؅ཧ ɹ5.1ɹ࣮ݧ؅ཧͷҙٛ ɹ5.2ɹHydraʹΑΔύϥϝʔλ؅ཧ ɹ5.3ɹMLflowʹΑΔ࣮ݧͷൺֱ ୈ6ষɹϓϩτλΠϓ։ൃ ɹ6.1ɹϓϩτλΠϓ։ൃͷҙٛ ɹ6.2ɹStreamlitʹΑΔϓϩτλΠϓ։ൃ ୈ1ষɹ࣮຿Ͱੜ͖ൈͨ͘ΊͷΤϯδχΞϦϯάεΩϧ ɹ1.1ɹσʔλαΠΤϯςΟετΛऔΓר͘؀ڥͷมભ ɹ1.2ɹϓϩδΣΫτͰٻΊΒΕΔΤϯδχΞϦϯάεΩϧ ୈ2ষɹ؀ڥߏங ɹ2.1ɹ෼ੳͷ౔୆ͱͯ͠ͷ؀ڥߏங ɹ2.2ɹϦϙδτϦͷߏ଄Λ੔͑Δ ɹ2.3ɹVS CodeͰͷ։ൃ؀ڥͷ੔උ ɹ2.4ɹDev Container ʹΑΔԾ૝؀ڥߏங ɹ2.5ɹuv ʹΑΔύοέʔδ؅ཧ ୈ3ষɹίʔυͷ඼࣭؅ཧ ɹ3.1ɹNotebook͚ͩͷσʔλαΠΤϯςΟετ͔Βͷଔۀ ɹ3.2ɹίʔυ඼࣭ͱ͸ ɹ3.3ɹ඼࣭ͷߴ͍ίʔυͱͦͷ࣮ݱํ๏ ɹ3.4ɹίʔυϨϏϡʔʹΑΔ඼࣭؅ཧ

All Rights Reserved 24 ͍࣮͟຿ʹऔΓ૊ΜͰࢥͬͨ͜ͱ ΈΜͳ΋ࠔ͍ͬͯΔ͸͔ͣͩΒ ࣮຿ͰֶΜͩ͜ͱΛ·ͱΊͯॻ੶ʹ͠Α͏ʂ ୈ1ষɹ࣮຿Ͱੜ͖ൈͨ͘ΊͷΤϯδχΞϦϯάεΩϧ ɹ1.1ɹσʔλαΠΤϯςΟετΛऔΓר͘؀ڥͷมભ ɹ1.2ɹϓϩδΣΫτͰٻΊΒΕΔΤϯδχΞϦϯάεΩϧ ୈ2ষɹ؀ڥߏங ɹ2.1ɹ෼ੳͷ౔୆ͱͯ͠ͷ؀ڥߏங ɹ2.2ɹϦϙδτϦͷߏ଄Λ੔͑Δ ɹ2.3ɹVS CodeͰͷ։ൃ؀ڥͷ੔උ ɹ2.4ɹDev Container ʹΑΔԾ૝؀ڥߏங ɹ2.5ɹuv ʹΑΔύοέʔδ؅ཧ ୈ3ষɹίʔυͷ඼࣭؅ཧ ɹ3.1ɹNotebook͚ͩͷσʔλαΠΤϯςΟετ͔Βͷଔۀ ɹ3.2ɹίʔυ඼࣭ͱ͸ ɹ3.3ɹ඼࣭ͷߴ͍ίʔυͱͦͷ࣮ݱํ๏ ɹ3.4ɹίʔυϨϏϡʔʹΑΔ඼࣭؅ཧ ୈ4ষɹσʔλͷ඼࣭֬ೝ ɹ4.1ɹσʔλͷ඼࣭֬ೝͷॏཁੑ ɹ4.2ɹ෼ੳ։࢝લͷσʔλ֬ೝ ɹ4.3ɹPanderaʹΑΔσʔλϑϨʔϜͷόϦσʔγϣϯ ɹ4.4ɹσʔλ඼࣭؅ཧͷܧଓతͳऔΓ૊Έ ୈ5ষɹػցֶशϞσϧͷ࣮ݧ؅ཧ ɹ5.1ɹ࣮ݧ؅ཧͷҙٛ ɹ5.2ɹHydraʹΑΔύϥϝʔλ؅ཧ ɹ5.3ɹMLflowʹΑΔ࣮ݧͷൺֱ ୈ6ষɹϓϩτλΠϓ։ൃ ɹ6.1ɹϓϩτλΠϓ։ൃͷҙٛ ɹ6.2ɹStreamlitʹΑΔϓϩτλΠϓ։ൃ

25 2ষ ؀ڥߏங

All Rights Reserved 30 ϦϙδτϦߏ଄Λ੔͑ΔϝϦοτ ద੾ͳϦϙδτϦߏ଄ʹ͢Δ͜ͱͰνʔϜͰͷ࿈ܞ΍௕ظతͳӡ༻͕͠΍͘͢ͳΔ શମ૾ͷཧղ͕༰қ ϓϩδΣΫτͷશମ૾΍ߏ଄Λ೺Ѳ͠΍͘͢ͳΓೝࣝͷڞ ༗͕εϜʔζʹͳΔ νʔϜ։ൃͷ෼୲͕͠΍͍͢ ࣮૷ͷ੹೚ൣғ͕໌֬ʹͳΓɺมߋ࣌ͷίϯϑϦΫτ͕཈ ੍Ͱ͖Δ อकɾ֦ு͕͠΍͍͢ ࢓༷มߋ΍ػೳ௥Ճͷࡍʹ΋ॊೈʹରԠͰ͖ΔͨΊɺٕज़ తෛ࠴ͷ஝ੵΛආ͚ΒΕΔ ࣗಈԽπʔϧͷಋೖɾӡ༻͕͠΍͍͢ CI/CDπʔϧͳͲΛಋೖ͠΍͘͢ӡ༻ޮ཰ΛߴΊΒΕΔ

All Rights Reserved 35 Dev ContainerʹͰ͖Δ͜ͱ ։ൃ࣌ͷ࣮ߦ؀ڥ΍VSCodeͷઃఆΛڞ௨ԽͰ͖Δ VSCodeͷઃఆ • ϦϯλʔϑΥʔϚολʔͷઃఆͳͲΛڞ௨Խ VSCodeͷ֦ுػೳ • ϓϩδΣΫτશମͰڞ௨ར༻͢Δ֦ுػೳΛઃఆ VSCodeͷσόοΨͷઃఆ • ։ൃऀؒͰڞ௨ͷσόοά؀ڥΛ࡞੒Ͱ͖Δ ࣮ߦ؀ڥͷڞ௨Խ VSCodeͷઃఆͷڞ௨Խ OSϨϕϧ͔Βڞ௨ͷ࣮ߦ؀ڥΛ࡞੒ Ͱ͖Δ OSͷڞ௨Խ ඞཁͳݴޠ΍πʔϧͷڞ௨Խ

All Rights Reserved 37 ύοέʔδϚωʔδϟͷൺֱ uv͸ύοέʔδ؅ཧػೳҎ֎ʹ΋ଟ͘ͷػೳΛ༗͓ͯ͠Γɺඇৗʹߴ଎Ͱಈ࡞͢Δ πʔϧ ಛ௃ pip ࠷΋جຊతͳύοέʔδΠϯετʔϥ Pythonͷඪ४ػೳͱͯ͠ར༻Մೳ Poetry ύοέʔδ؅ཧ͚ͩͰ͸ͳ͘ɺԾ૝؀ڥ؅ཧɾґଘؔ܎ΛࣗಈͰղ ܾɾύοέʔδެ։ͳͲͷػೳ΋උΘ͍ͬͯΔ uv PoetryͷػೳΛΧόʔ Rust੡Ͱඇৗʹߴ଎ʢpipͷ10~100ഒߴ଎ʣ ※ uv͸όʔδϣϯ͕0.9.27(2026/1/27)Ͱϝδϟʔόʔδϣϯ͕0Ͱ͋ΔͨΊϓϩδΣΫτͰ࢖͏৔߹ʹ͸ҎԼͷՄೳੑʹ஫ҙ͢Δඞཁ͕͋Γ·͢ɻ ▪ ҆ఆੑ͕อূ͞Ε͍ͯͳ͍ ▪ ϦϦʔε४උલ ▪ ະ׬੒·ͨ͸ࢼݧతͳར༻

All Rights Reserved 38 uvʹͰ͖Δ͜ͱ uv͸Python։ൃʹ͓͍ͯॏཁͱͳΔҎԼͷػೳΛඋ͍͑ͯΔ Pythonύοέʔδͷ؅ཧ ؅ཧ͍ͯ͠ΔPythonύο έʔδ΍Pythonͷόʔδ ϣϯΛඥ͚ͮͨɺԾ૝؀ ڥΛ࡞੒Ͱ͖Δ ։ൃͨ͠ίʔυΛPythonύ οέʔδͷܗࣜʹ·ͱΊ PyPI(Pythonͷύοέʔδ ڞ༗αʔϏε)ʹอଘ͢Δ ͜ͱ͕Ͱ͖Δ Ծ૝؀ڥͷߏங Pythonͷόʔδϣϯ؅ཧ Pythonύοέʔδͷ࡞੒ ɹɹɹ3.11 ɹɹɹ3.12 ෳ਺ͷPythonͷόʔδϣ ϯΛ؅ཧ͢Δ͜ͱ͕Ͱ͖ Δ ؅ཧ ɾ ɾ ɾ PythonύοέʔδΛόʔ δϣϯΛࢦఆͯ͠؅ཧ͢ Δ͜ͱ͕Ͱ͖Δ

39 3ষ ίʔυͷ඼࣭؅ཧ

All Rights Reserved 42 σʔλαΠΤϯςΟετʹٻΊΒΕΔՁ஋ͷมԽ ٻΊΒΕΔ໾ׂ͕มԽ͠Notebook͚ͩͰՁ஋Λग़͠ʹ͍͘ঢ়گʹͳͬͨ Notebookͷݶք • ࣮ݧతͳίʔυͱ׬੒൛ίʔυͷࠞࡏ • ߏ଄Խͷ೉͠͞ͱίʔυͷෳࡶԽ • νʔϜ։ൃ΍Git؅ཧͷ೉͠͞ • σʔλͷ୳ࡧ΍෼ੳ • PoC • ۀ຿ϨϕϧͰͷ։ൃ • ௕ظతͳӡ༻ ࠓ·ͰٻΊΒΕ͍ͯͨεΩϧ ௥ՃͰٻΊΒΕΔεΩϧ

All Rights Reserved 44 Notebook vs Python Script NotebookͱPython Scriptͷదͨ͠༻్Λཧղ͠࢖͍෼͚Δ͜ͱ͕ॏཁ VS ϝϦοτ • ΠϯλϥΫςΟϒʹѻ͑ɺࢼߦࡨޡΛ܁Γฦ͠΍͍͢ దͨ͠༻్ • ୳ࡧతσʔλ෼ੳ • σʔλͷՄࢹԽ • εϐʔυ͕ٻΊΒΕΔΞυϗοΫͳରԠ Notebook ϝϦοτ • ࠶ݱੑ͕ߴ͍ • όʔδϣϯ؅ཧ͕͠΍͍͢ • ίʔυશମͷߏ଄Խ΍࠶ར༻͕༰қ దͨ͠༻్ • σʔλલॲཧ • ػցֶशϞσϧͷ։ൃ • ϓϩτλΠϓ։ൃ Python Script

All Rights Reserved 45 Python Interactive Window Python Interactive WindowͷެࣜυΩϡϝϯτɿ https://code.visualstudio.com/docs/python/jupyter-support-py • VS Code্ͰPython ScriptΛNotebookͷ Α͏ʹ෦෼తʹ࣮ߦͰ͖Δػೳ • ʮ# %%ʯΛ෇͚Δ͚ͩͰ؆୯ʹར༻Ͱ ͖Δ ※ PythonͱJupyterͷ֦ுػೳͷΠϯετʔϧ͕ඞཁͰ͢

All Rights Reserved 48 ؔ਺Λ࣮૷͢Δ্ͰؾΛ͚ͭΔ؍఺ ؔ਺Λ࣮૷͢Δࡍ͸ ࢀরಁաੑɾ෭࡞༻ͷഉআɾؔ਺ͷେ͖͞ ʹ஫ҙ ؍఺ ຬͨ͢΂͖৚݅ ҧ൓͢Δྫ ࢀরಁաੑ ಉ͡ೖྗʹରͯ͠ಉ͡ग़ྗΛฦ͢͜ͱ Ҿ਺Λհͣ͞άϩʔόϧͳઃఆ஋   ʢֶश཰ͳͲʣΛࢀর͍ͯ͠Δ ෭࡞༻ͷഉআ ؔ਺ͷ֎ଆͷঢ়ଶΛมߋ͠ͳ͍͜ͱ Ҿ਺Ͱ౉͞Εͨ pd.DataFrame Λؔ਺಺ Ͱ௚઀ॻ͖׵͑Δʢinplace=Trueʣ ؔ਺ͷେ͖͞ ؔ਺͕࣋ͭॲཧ͕େ͖͗͢ͳ͍͜ͱ Ұͭͷؔ਺Ͱʮܽଛॲཧʯʮਖ਼نԽʯ ͳͲෳ਺ͷෳࡶͳॲཧΛߦ͍ͬͯΔ

All Rights Reserved from abc import ABC, abstractmethod import lightgbm as lgb import pandas as pd from sklearn.linear_model import LinearRegression class ModelInterface(ABC): @abstractmethod # ந৅ϝιουʢܧঝ͢ΔΫϥεʹ࣮૷Λڧ੍͢Δʣ def train(self, X: pd.DataFrame, y: pd.Series) -> None: pass class LinearRegressionModel(ModelInterface): def __init__(self) -> None: ... def train(self, X: pd.DataFrame, y: pd.Series) -> None: ... class LightGBMModel(ModelInterface): def __init__(self, params: dict[str, Any]) -> None: ... def train(self, X: pd.DataFrame, y: pd.Series) -> None: ... 50 ϙϦϞʔϑΟζϜΛҙࣝͨ͠Ϋϥεͷ࣮૷ྫ ಉ͡ΠϯλʔϑΣʔεʹԊ࣮ͬͨ૷ʹΑΓ   ϩδοΫΛަ׵ՄೳʹʢStrategyύλʔϯʣ

All Rights Reserved 51 σʔλαΠΤϯςΟετʹ޲͚ͨؔ਺ͱΫϥεͷ࢖͍෼͚ ಠཱͨ͠ॲཧ͸ؔ਺ɺঢ়ଶΛ࣋ͭৼΔ෣͍͕ඞཁͳ৔߹͸ΫϥεΛ࢖͏ ಛ௃ Ϣʔεέʔε ؔ਺ ঢ়ଶɾσʔλΛ࣋ͨͳ͍ ಠཱͨ͠ॲཧΛߦ͍͍ͨ৔߹ • σʔλͷલॲཧͳͲ Ϋϥε σʔλͱॲཧΛ·ͱ·Γͱͯ͠ѻ͑Δ σʔλͱૢ࡞͕ີʹབྷΈ߹͏৔߹ • MLϞσϧͳͲ • Ϋϥε͸Πϯελϯεม਺Λ௨ͯ͠ϝιουؒʹ҉໧తͳґଘؔ܎Λ࡞Γ͕ͪ • ෳ਺ͷॲཧΛ·ͱΊΔ͚ͩͰ͋Ε͹ϞδϡʔϧԽʢ୯Ұͷ .py ʹ·ͱΊΔʣ͕༗ޮ

All Rights Reserved 54 σʔλαΠΤϯςΟετʹͱͬͯͷίʔυϨϏϡʔ PoC͔Β։ൃ΁εϜʔζʹਐΉͨΊʹϨϏϡʔΛ௨ͨ͠඼࣭୲อ͕ॏཁ • ΦϨΦϨίʔσΟϯά • աڈʹͨ͠มߋͷ಺༰Λࢥ͍ग़ͤͳ͍ ίʔυϨϏϡʔ͕ͳ͍ঢ়ଶ ˚ ඼࣭؅ཧ͕ݸਓʹด͡ଐਓԽ • ίʔυͷ඼࣭Λݟ௚͢ػձ͕Ͱ͖Δ • มߋ΍ͦͷҙਤΛ௥੻Ͱ͖Δ ίʔυϨϏϡʔ͕͋Δঢ়ଶ ̋ νʔϜͰίʔυ඼࣭ʹ੹೚Λ࣋ͯΔ

All Rights Reserved 56 ίʔυϨϏϡʔʛϨϏϡʔͷ৺ಘ ࣗ෼ͷίʔυʹ͸ࣗ෼͕੹೚Λ࣋ͭ ςετͳͲʹΑΓϩδοΫͷਖ਼֬͞͸ࣗ਎Ͱ୲อ ίϛοτɾPRͷҙਤΛ໌֬Խ͢Δ ద੾ͳཻ౓ɺίϛοτϝοηʔδ ϨϏϡΞʔ΁ͷཁ๬ͷ໌ࣔ͢Δ PRςϯϓϨʔτػೳͷ׆༻ ϨϏϡΠʔͷ৺ಘ ϨϏϡʔίϝϯτͷҙਤΛ໌֬Խ͢Δ prefixʹΑΔཁٻͷ໌ࣔʢMUST / IMOͳͲʣ ϙδςΟϒͳϑΟʔυόοΫ΋͢Δ ϨϏϡʔ͸৆ࢍͷ৔Ͱ΋͋Δ ϨϏϡʔҎ֎ͷબ୒ࢶ΋࣋ͭ ྠಡձ΍ϖΞϓϩΛ௨ͨ͠εΩϧͷ౔୆ͮ͘Γ ϨϏϡΞʔͷ৺ಘ ϨϏϡΠʔɾϨϏϡΞʔͷ૒ํ͕ࣗ਎ͷ໾ׂΛཧղ͢Δ͜ͱ͕ॏཁ

58 4ষ σʔλͷ඼࣭֬ೝ

All Rights Reserved 61 σʔλ඼࣭ͷධՁ࣠ʢࢀߟʣ ධՁ࣠ આ໌ ۩ମྫ ༗ޮੑ σʔλͷ஋͕ఆٛ͞ΕͨྖҬͷ஋ͱҰக͍ͯ͠Δ͔ ೥ྸʹෛͷ஋͕ͳ͍͔ ׬શੑ ඞཁͳσʔλ͕͢΂ͯଘࡏ͍ͯ͠Δ͔ ࢦఆͨ͠શظؒͷσʔλ͕ἧ͍ͬͯΔ͔ Ұ؏ੑ σʔλͷ಺༰ʹໃ६΍৯͍ҧ͍͕ͳ͍͔ ΧϥϜͷதͰ஋ʹදه༳Ε͕ͳ͍͔ ੔߹ੑ σʔλ͕Ϗδωεϧʔϧ΍੍໿ʹج͍ͮͯਖ਼͘͠   ؔ܎෇͚ΒΕ͍ͯΔ͔ ෳ਺ςʔϒϧؒͰID৘ใ͕ਖ਼͘͠࿈ܞͰ͖͍ͯΔ͔ ࠷৽ੑ σʔλ͕ݱ࣮ੈքͷ࠷৽ͷঢ়ଶΛ൓ө͍ͯ͠Δ͔ ސ٬Ϛελ͸࠷৽ͷσʔλͰ͋Δ͔ ଥ౰ੑ σʔλύλʔϯ͕ظ଴ʹ߹க͍ͯ͠Δ͔ ೥ྸͷ෼෍͕ײ֮஋ͱେ͖ͳဃ཭͕ͳ͍͔ Ұҙੑ ಉ࣮͡ମΛද͢σʔλ͕ɺಉ͡ςʔϒϧ಺ʹෳ਺   ଘࡏ͍ͯ͠ͳ͍͔ ॏෳ͍ͯ͠ΔϨίʔυ͸ͳ͍͔

All Rights Reserved 63 PanderaʹΑΔσʔλϑϨʔϜͷόϦσʔγϣϯ PanderaʹΑΔόϦσʔγϣϯΛ௨ͯ͠σʔλ͕ظ଴͢Δ੍໿Λຬ͍ͨͯ͠Δ͜ͱΛ୲อ͢Δ import pandera as pa from pandera.typing import DataFrame from .consts import MAX_PREDICT_DAYS from .schema import InferInputSchema, TaxiDatasetSchema @pa.check_types def preprocess_for_infer( df: DataFrame[TaxiDatasetSchema], max_predict_days: int = MAX_PREDICT_DAYS, ) -> DataFrame[InferInputSchema]: ... # dfΛdf_resultʹม׵͢Δॲཧ return df_result όϦσʔγϣϯͷ࣮ߦ import numpy as np import pandas as pd import pandera as pa from pandera.typing import Index, Series class TaxiDatasetSchema(pa.DataFrameModel): date: Series[np.datetime64] area: Series[pd.CategoricalDtype] num_trip: Series[int] class InferInputSchema(pa.DataFrameModel): ... εΩʔϚͷఆٛ

65 5ষ ػցֶशϞσϧͷ࣮ݧ؅ཧ   6ষ ϓϩτλΠϓ։ൃ

All Rights Reserved 66 5ষ ػցֶशϞσϧͷ࣮ݧ؅ཧ ࣮ݧ؅ཧ ʹ ࢼߦࡨޡΛ҆શ͔ͭޮ཰తʹճͨ͢Ίͷϓϩηε • PoCͰ͸ࢼߦࡨޡ͕͖ͭ΋ͷ ◦ ϋΠύʔύϥϝʔλͷௐ੔ɺϓϩϯϓτνϡʔχϯάɺetc… • ద੾ͳ࣮ݧ؅ཧΛ͠ͳ͍ͱτϨʔαϏϦςΟͱ࠶ݱੑͷܽ೗Λট͘ ◦ ʮͲ͜ʹԿΛه࿥͔ͨ͠Θ͔Βͳ͍....ʯ ◦ ʮաڈͷ͏·͍࣮ͬͨ͘ݧઃఆΛ࠶ݱͰ͖ͳ͍....ʯ ॻ੶Ͱ͸୅දతͳ࣮ݧ؅ཧπʔϧͱͯ͠ Hydra ͱ MLflow ʹ͍ͭͯ঺հ

All Rights Reserved 67 6ষ ϓϩτλΠϓ։ൃ • σʔλαΠΤϯςΟετͷ໾ׂ͸෼ੳΛ௨ͯ͠ϢʔβʔʹՁ஋Λಧ͚Δ͜ͱ • ϓϩτλΠϓ͸ϢʔβʔʹՁ஋Λఏڙ͢ΔͨΊͷΞϓϩʔνͷҰͭ ◦ ਝ଎ͳϑΟʔυόοΫͷ֫ಘ ◦ ೝࣝᴥᴪͷϦεΫ௿ݮ ◦ ࣮ݱՄೳੑͷݕূ ◦ ҆৺ײ΍ϫΫϫΫײͷఏڙ ϓϩτλΠϓ͸Ϣʔβʔʹ෼ੳ݁ՌͷՁ஋Λಧ͚ΔͨΊͷ৽͍͠ܗ   ͱͯ͠஫໨͞Ε͍ͯΔ ॻ੶Ͱ͸WebΞϓϦΛ࡞੒Ͱ͖ΔϑϨʔϜϫʔΫ Streamlit ʹ͍ͭͯ঺հ

ॻ੶ʹ͸ɺΑΓৄ͍͠ղઆͱ   ࣮ફతͳίʔυΛࡌ͍ͤͯ·͢ʂ

69 ࣥචͷཪ࿩

͓·͚ ॻ੶ͷࣥචϑϩʔΛ͝঺հ

All Rights Reserved 71 ࣥචϓϩηεͷશମ૾ اըॻ࡞੒ ݪߘࣥච ૊൛ɾߍӾ ग़൛ ࣥච͢Δॻ੶ͷίϯ ηϓτΛݻΊΔϑΣ ʔζͰ͢ɻ ॻ੶ͷίϯηϓτ΍ λʔήοτ૚ɺ۩ମ తͳষͷߏ੒ͳͲΛ اըॻͱͯ͠੔ཧ ͠ɺग़൛ࣾʹఏग़͠ ·͢ɻ اը͕ແࣄ௨Ε ͹ɺࣥච࡞ۀ։࢝ Ͱ͢ɻ ਵ࣌ɺग़൛ࣾͷ୲ ౰ऀͷํʹ΋ϑΟ ʔυόοΫΛ௖͖ ͭͭɺݪߘΛॻ͖ ্͛·͢ɻ ׬੒ݪߘΛ࣮ࡍͷ ॻ੶ͷϨΠΞ΢τ ʹམͱ͠ࠐΈ·͢ ʢ͜ͷ࡞ۀΛ૊൛ ͱݴ͍·͢ʣɻ ૊൛ޙ͸ࣥචऀͱ ग़൛ࣾͷؒͰෳ਺ ճͷमਖ਼Λԟ෮ ͠ɺ࠷ऴతʹ಺༰ ͕fix͞Ε·͢ɻ ֬ఆͨ͠ݪߘ͕ҹ ࡮ॴ΁ೖߘ͞Εɺ ੡ຊ͞Ε·͢ɻ ͦͯ͠ɺશࠃͷॻ ళʹ഑ຊ͞Ε੖Ε ͯץߦͱͳΓ· ͢ɻ

All Rights Reserved 73 ݪߘ؅ཧ͸GitHub ڞಉͰࣥච࡞ۀΛਐΊΔʹ͋ͨΓɺGitHubͰMarkdownܗࣜͷݪߘΛ؅ཧ ҆શʹόʔδϣϯ؅ཧ • ࢼߦࡨޡ͕༰қɺ؆୯ʹ෮ݩՄೳ • ࣥච్தʹग़൛ࣾ΁ͷϨϏϡʔ͢Δࡍ͸ɺGit TagͰεφοϓγϣ οτΛه࿥ Pull RequestΛ௨ͨ͠ ϨϏϡʔϑϩʔͷཱ֬ • ֤ষ͝ͱʹϨϏϡʔ୲౰ऀΛΞαΠϯ͠ɺPRΛ௨ͯ͠ϨϏϡʔ • PRͷςϯϓϨʔτػೳ΍GitHub ActionsʹΑΔCIͳͲͰɺ   ࣥච಺༰Λඪ४Խ ݪߘͱαϯϓϧίʔυ Λซͤͯ؅ཧ • ίʔυͱݪߘΛҰݩ؅ཧ • ຊจͱࢀরίʔυͷ಺༰ͷဃ཭Λ๷͗ͳ͕Βਐߦ GItHubͰͷ؅ཧʹΑΔϝϦοτ

All Rights Reserved 74 ϨϏϡʔϓϩηεͷৄࡉ 1ɿඇಉظϨϏϡʔ ֤ষͷ୲౰ऀͰ࣮ࢪ 2ɿಉظϨϏϡʔʢ௨শɿϫΠϫΠϨϏϡʔʣ ि1ճू·ΓɺϝϯόʔશһͰ௨ಡ ؾʹͳΔͱ͜ΖΛࢦఠ͠߹͏ ɹ   1.ͱ2.Λ܁Γฦͯ͠ϒϥογϡΞοϓ ಉظɾඇಉظͷ྆ํͰ࣮ࢪ ҹ࡮ͯ͠ಡΉͱɺදݱͷҧ࿨ײ΍࿦ཧߏ଄ ͷϛεʹؾ͖ͮ΍͍͢ ʢMarkdownΛը໘্ͰॎʹεΫϩʔϧͯ͠ ಡΉͱɺ໨͕׈Δʣ ݪߘ࡞੒ஈ֊Ͱগͳ͘ͱ΋Ұ౓͸ҹ࡮͢Δ ͜ͱΛਪ঑ ࢴ໘Ͱ΋֬ೝ

76 ͓ΘΓʹ

ͥͻ͓ങ͍ٻΊ͍ͩ͘͞ʂ ͝ߪೖ͸ͪ͜Β͔Β https://gihyo.jp/book/2025/978-4-297-15100-3 ॻళɾΦϯϥΠϯετΞͰ޷ධൃചதʂ

79 ࠷ޙʹେࣄͳ͜ͱΛݴ͍·͢ ʢࢲʹͱͬͯʣ

All Rights Reserved 81 ΄͔ʹ΋ओ࠵Meetup͕2݄ͱ3݄ʹ͋Γ·͢ʂͥͻ͝ࢀՃ͍ͩ͘͞ʂ 2026೥3݄5೔ʢ໦ʣ നۚ߭ۀMeetup Vol.22 ɹ Geoσʔλฤ ʢҐஔ৘ใ•஍ཧۭؒʣ 2/26ʢ໦ʣੜ੒AIձ @͓୆৔ 3/5ʢ໦ʣനۚ߭ۀ Meetup @࿡ຊ໦Ұஸ໨ https://seisei-ai.connpass.com/event/378562/ https://brainpad-meetup.connpass.com/event/382273/

Thank You!!

All Rights Reserved 83 ࢀߟ ࣥචऀ͕ॻ੶ʹࠐΊͨࢥ͍ΛޠͬͨΓɺಡऀ͔Βͷ൓ԠΛࡌ͍ͤͯ·͢ʂ [ઌഐσʔλαΠΤϯςΟετ͔Βͷࢦೆॻ | ٕज़ධ࿦ࣾ] https://gihyo.jp/book/2025/978-4-297-15100-3 [ʮઌഐσʔλαΠΤϯςΟετ͔Βͷࢦೆॻʯͷײ૝· ͱΊ - posfie] https://posfie.com/@nash_efp/p/VMShgtZ [ॻධϒϩά·ͱΊɿઌഐσʔλαΠΤϯςΟετ͔Β ͷࢦೆॻʛΞαϊ / asano] https://note.com/nash_efp/n/n344930f80766 [ʲ2025/08/27ൃചʳॻ੶ʮઌഐσʔλαΠΤϯςΟε τ͔ΒͷࢦೆॻʯΛࣥච͠·ͨ͠ʛΞαϊ / asano] https://note.com/nash_efp/n/n0d0ed1590790 [ॳΊͯͷٕज़ॻࣥචͰֶΜ࣮ͩફϊ΢ϋ΢] https://zenn.dev/suikabar/articles/24fa9c2fe999f3 [ʲॻ੶ग़൛ه೦ʳσʔλαΠΤϯςΟετͷʮ͜͏͍͏ ͜ͱΛڭ͑ͯཉ͔ͬͨ͠ʯΛ1࡭ʹʂ࣮ફ஌ΛମܥԽ͠ ͨࢦೆॻͷཪଆΛࣥචਞ͕ޠΔ - Platinum Data Blog by BrainPad ϒϨΠϯύου] https://blog.brainpad.co.jp/entry/2025/09/10/150042

20260127_試行錯誤の結晶を1冊に。著者が解説 先輩データサイエンティストからの指...

20260127_試行錯誤の結晶を1冊に。著者が解説 先輩データサイエンティストからの指南書 / author's_commentary_ds_instructions_guide

Transcript

More Decks by nash_efp

Other Decks in Programming

Featured

20260127_試行錯誤の結晶を1冊に。著者が解説先輩データサイエンティストからの指...

20260127_試行錯誤の結晶を1冊に。著者が解説先輩データサイエンティストからの指南書 / author's_commentary_ds_instructions_guide