Source code for Transcription between Human-readable Synthetic Descriptions and Machine-executable Instructions: An Application of Latest Pre-training Technology. Our operating system is Ubuntu 16.04. For training process, the 3090 GPU is used.
If you want to quickly explore our job or do not have much deep-learning experience, you can simply follow the instructions in this section.
- Step 1: Download the zip or clone the repository to your workspace.
- Step 2: Download the
D2IPTM-base/large.ptandI2DPTM-base/large.ptfrom googledrive. Create a new directory bymkdir save_modeland then put the downloaded model undersave_model/directory. - Step 3: Install Anaconda (py3) and then create a conda environment by the following command (remember to input 'y' when asked
Proceed ([y]/n)?):
conda create -n ChemTrans python=3.9
conda activate ChemTrans
sh scripts/conda_environment.sh
Note that there may be error when installing transformers if you're using MacOS. See here for help.
- Step 4: Check the
interact.pyfile and setif_cuda=Falsein line 25 if there is no GPU available. Run the command:
python interact.py
And then explore the ChemTrans task following the instructions of the program:
# input the synthesis natural language description and then type enter
SI input: >> an oven-dried Schlenk flask equipped with a magnetic stir bar was charged with N-bromosuccinimide (1.26 g, 7.1 mmol, 1.25 equiv.) and sealed. The flask was evacuated and backfilled with nitrogen (this sequence was repeated a total of three times). To the flask were added a solution of S11 (0.89 g, 5.7 mmol, 1.0 equiv.) in acetone (35 mL) via syringe. The flask was opened and silver nitrate (0.10 g, 0.6 mmol, 10 mol%) was added quickly. The flask was re-sealed, and the reaction mixture was stirred at r.t. for 2 h. The solvent was removed in vacuo and the residue was diluted with a mixture of PE:EA (v/v = 1:1). The resulting slurry was filtered through a plug of silica gel and eluted with PE:EA (v/v = 1:1). The filtrate was washed with water and brine, dried over anhydrous Na2SO4 and concentrated in vacuo. The residue was purified by silica gel column chromatography (PE:EA= 40:1 – 4:1) to afford S12 (1.17 g, 88%) as a pink solid.
# the program will return the predicted instructions
[ add ] reagent: ( name: N-bromosuccinimide & type: pure & mole: 7.1 mmol & mass: 1.26 g & concentration: 1.25 equiv & ) & [ add ] reagent: ( name: S11 & mole: 5.7 mmol & mass: 0.89 g & equivalent: 1.0 equiv & ) & reagent: ( name: acetone & type: pure & volume: 35 mL & ) & reagent: ( name: a solution of S11 0.89 g, 5.7 mmol, 1.0 equiv. in acetone 35 mL & type: mixture & ) & [ add ] reagent: ( name: silver nitrate & type: pure & mole: 0.6 mmol & mass: 0.10 g & concentration: 10 mol% & speed: quickly. & ) & [ settemp ] time: 2 h. & [ evaporate ] N/A: removed & [ add ] reagent: ( name: PE:EA & type: mixture & ) & [ filter ] reagent: ( name: silica gel & type: pure & ) & [ column ] reagent(adsorbent): silica gel & [ wash ] reagent: ( name: water & type: pure & ) & reagent: ( name: brine & type: pure & ) & [ dry ] reagent: ( name: Na2SO4 & type: pure & note: anhydrous & ) & [ evaporate ] N/A: concentrated & [ column ] reagent(eluent): PE:EA= 40:1 – 4:1 & [ yield ] appearance: a pink solid & yield: 88% & mass(yield): 1.17 g &
# automatically loop until you type control+C
SI input: >>Or provide input instructions with the help of the program, and get the transcribed descriptions. Remember to set if_mutual=1 in line 26.
# input the corresponding number for the operations and augments
{1: 'add', 2: 'settemp', 3: 'yield', 4: 'wash', 5: 'filter', 6: 'evaporate', 7: 'dry', 8: 'distill', 9: 'extract', 10: 'transfer', 11: 'reflux', 12: 'recrystallize', 13: 'quench', 14: 'column', 15: 'triturate', 16: 'partition'}
choose operation (17 for stop): >> 1
{1: 'temperature', 2: 'reagent'}
choose augments (0 for stop): >> 2
{1: 'name', 2: 'type', 3: 'mass', 4: 'volume', 5: 'speed', 6: 'concentration', 7: 'equivalent', 8: 'batch', 9: 'note', 10: 'temperature', 11: 'mole'}
choose reagent augments (0 for stop): >> 1
# input the value for the augments
value for name: >> N-bromosuccinimide
choose reagent augments (0 for stop): >> 2
value for type: >> pure
choose reagent augments (0 for stop): >> 3
value for mass: >> 1.26 g
choose reagent augments (0 for stop): >> 11
value for mole: >> 7.1 mmol
choose reagent augments (0 for stop): >> 6
value for concentration: >> 1.25 equiv
choose reagent augments (0 for stop): >> 0
choose augments (0 for stop): >> 0
choose operation (17 for stop): >> 2
{1: 'time', 2: 'temperature'}
choose augments (0 for stop): >> 1
value for time: >> 2 h.
choose augments (0 for stop): >> 0
choose operation (17 for stop): >> 6
{1: 'temperature', 2: 'pressure'}
choose augments (0 for stop): >> 1
value for temperature: >> removed
choose augments (0 for stop): >> 0
choose operation (17 for stop): >> 5
{1: 'reagent', 2: 'phase'}
choose augments (0 for stop): >> 1
{1: 'name', 2: 'type', 3: 'mass', 4: 'volume', 5: 'speed', 6: 'concentration', 7: 'equivalent', 8: 'batch', 9: 'note', 10: 'temperature', 11: 'mole'}
choose reagent augments (0 for stop): >> 1
value for name: >> silica gel
choose reagent augments (0 for stop): >> 2
value for type: >> pure
choose reagent augments (0 for stop): >> 0
choose augments (0 for stop): >> 0
choose operation (17 for stop): >> 3
{1: 'reagent name', 2: 'appearance', 3: 'mass(yield)', 4: 'yield'}
choose augments (0 for stop): >> 2
value for appearance: >> a pink solid
choose augments (0 for stop): >> 4
value for yield: >> 88%
choose augments (0 for stop): 3
value for mass(yield): >> 1.17 g
choose augments (0 for stop): 0
choose operation (17 for stop): 17
# check the input, modify and ensure your instructions
Your input instructions are: [ add ] reagent: ( name: N-bromosuccinimide & type: pure & mass: 1.26 g & mole: 7.1 mmol & concentration: 1.25 equiv & ) [ settemp ] time: 2 h. & [ evaporate ] temperature: removed & [ filter ] reagent: ( name: silica gel & type: pure & ) [ yield ] appearance: a pink solid & yield: 88% & mass(yield): 1.17 g &
verify your input: [ add ] reagent: ( name: N-bromosuccinimide & type: pure & mass: 1.26 g & mole: 7.1 mmol & concentration: 1.25 equiv & ) [ settemp ] time: 2 h. & [ evaporate ] N/A: removed & [ filter ] reagent: ( name: silica gel & type: pure & ) [ yield ] appearance: a pink solid & yield: 88% & mass(yield): 1.17 g &
# get the transcribed result
To a solution of tert-butyl 2-(dimethylamino)cyclohexane-1,3-dione (1.17 g, 7.1 mmol, 88%) in tetrahydrofuran (50 mL) is added N-bromosuccinimide (1.26 g, 7.1 mmol, 1.25 equiv). The flask is fitted with a magnetic stirring bar and the solution is stirred under nitrogen for 2 h. The solvent is removed with a rotary evaporator and the residue is filtered through a short pad of silica gel (elution with ethyl acetate-hexanes) to afford the product as a pink solid (1.17 g, 88%).We strongly suggest you to create a conda environment for this project. Installation is going to be finished in a several minutes.
conda create -n ChemTrans python=3.9
conda activate ChemTrans
sh scripts/conda_environment.sh
D2I/I2DPTM and other pre-trained models can be downloaded from googledrive. We recommend you to download the models and put them under save_model/ before running the code.
Pre-training corpus and the pre-processed data can also be downloaded from the above link. Please create a file called predata/ and put the data under it.
The users may be going to use the files below:
- mix_pretraining: D2I knowledge-enhanced training code for all of the tasks.
- dis_pretraining: I2D knowledge-enhanced training code for tasks except augmentation.
- dec_pt.py: Training code for the decoder language modeling.
- tuning.py: Fine-tuning code for ChemTrans task.
- interact.py: Demo code for ChemTrans interaction.
- evalchem.py: Evaluation function code.
- D2I_deval.py: Evaluation tool code.
- I2D_deval.py: Evaluation tool code.
- gen_aug.py: Decoder language modeling data augmentation generation code.
- LLM_test.py: Large language model testing code.
- data/
- token_save.pkl: Special tokens for expanding T5 tokenizer.
- *_inp.txt: Input text for train/dev/test/aug set.
- *_out.txt: Label text for train/dev/test/aug set.
- aug_decoder.txt: Augmented text for decoder language modeling.
- query_instances.txt: The most similar training instances used for LLM testing.
- scripts/
- conda_environment.sh: Conda environment creation bash file.
- decoder_lm.sh: Decoder language modeling bash file.
- finetune.sh: Fine-tuning bash file.
- pretrain.sh: Knowledge-enhanced training bash file.
We strongly recommend you to test our code with a GPU. Usually the downstream fine-tuning process takes no more than an hour (for base ver.).
For the D2I training period, we modify the scripts/pretrain.sh file according to our settings:
version='base'
# choose 'base', 'large' or 'zero' for the model scale, and 'zero' stands for the small Vanilla Transformer setting. Other model scale can be defined in the mix_pretraining.py file.
save='save_model/ckpt_decbasepre'
# path and name for model saving.
datatrain='predata/data_mix.jbl'
# data_mix.jbl for multi-task training, and data_mlm.jbl for purely masked language modeling.
batch=8
lm='--pth_lm data/aug_train.txt'
# set lm='' for post-training without decoder language modeling.
And run the bash file.
CUDA_VISIBLE_DEVICES=0 sh scripts/pretrain.sh
For I2D training period, replace the python file with dis_pretraining.py, and then use tuning code before fine-tuning, while replace the training text with aug_inp/out.txt.
For the fine-tuning period, we modify the scripts/finetune.sh file according to our settings:
mutual=0
# set 0 for D2I and 1 for I2D
version='base'
# choose 'base', 'large' or 'zero' for the model scale
init='--init_checkpoint save_model/ckpt_basepre980.pt'
# set init='' if there is no need for initializing the T5 model
save='save_model/finetune_basepre.pt'
# path and name for model saving.
log='log/ftbasepre_'
# path and name for the predicted context.
few=1
# set the data ratio used for training
And run the bash file.
CUDA_VISIBLE_DEVICES=0 sh scripts/finetune.sh
For the evaluation, run the code according to the log file that the model generated:
python D2I_deval.py LOG_FILE(e.g. log/ftbasepre_)
python I2D_deval.py LOG_FILE(e.g. log/ftrevbasepre_)
Please cite our paper if you find it helpful.
@Article{D3SC02483K,
author ="Zeng, Zheni and Nie, Yi-Chen and Ding, Ning and Ding, Qian-Jun and Ye, Wei-Ting and Yang, Cheng and Sun, Maosong and E, Weinan and Zhu, Rong and Liu, Zhiyuan",
title ="Transcription between human-readable synthetic descriptions and machine-executable instructions: an application of the latest pre-training technology",
journal ="Chem. Sci.",
year ="2023",
pages ="-",
publisher ="The Royal Society of Chemistry",
doi ="10.1039/D3SC02483K",
url ="http://dx.doi.org/10.1039/D3SC02483K"
}