|
--- |
|
language: |
|
- en |
|
license: mit |
|
tags: |
|
- chemistry |
|
- SMILES |
|
- yield |
|
datasets: |
|
- ORD |
|
metrics: |
|
- r_squared |
|
--- |
|
|
|
# Model Card for ReactionT5v2-yield |
|
|
|
This is a ReactionT5 pre-trained to predict yields of reactions. You can use the demo [here](https://huggingface.co/spaces/sagawa/ReactionT5_task_yield). |
|
|
|
### Model Sources |
|
|
|
<!-- Provide the basic links for the model. --> |
|
|
|
- **Repository:** https://github.com/sagawatatsuya/ReactionT5v2 |
|
- **Paper:** https://arxiv.org/abs/2311.06708 |
|
- **Demo:** https://huggingface.co/spaces/sagawa/ReactionT5_task_yield |
|
|
|
## Uses |
|
|
|
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. --> |
|
|
|
## How to Get Started with the Model |
|
|
|
Use the code below to get started with the model. |
|
|
|
```python |
|
import torch |
|
import torch.nn as nn |
|
from transformers import AutoTokenizer, T5ForConditionalGeneration, AutoConfig, PreTrainedModel |
|
import logging |
|
logging.getLogger('transformers').setLevel(logging.ERROR) |
|
|
|
class ReactionT5Yield(PreTrainedModel): |
|
config_class = AutoConfig |
|
def __init__(self, config): |
|
super().__init__(config) |
|
self.config = config |
|
self.model = T5ForConditionalGeneration.from_pretrained(self.config._name_or_path) |
|
self.model.resize_token_embeddings(self.config.vocab_size) |
|
self.fc1 = nn.Linear(self.config.hidden_size, self.config.hidden_size//2) |
|
self.fc2 = nn.Linear(self.config.hidden_size, self.config.hidden_size//2) |
|
self.fc3 = nn.Linear(self.config.hidden_size//2*2, self.config.hidden_size) |
|
self.fc4 = nn.Linear(self.config.hidden_size, self.config.hidden_size) |
|
self.fc5 = nn.Linear(self.config.hidden_size, 1) |
|
|
|
self._init_weights(self.fc1) |
|
self._init_weights(self.fc2) |
|
self._init_weights(self.fc3) |
|
self._init_weights(self.fc4) |
|
self._init_weights(self.fc5) |
|
|
|
def _init_weights(self, module): |
|
if isinstance(module, nn.Linear): |
|
module.weight.data.normal_(mean=0.0, std=0.01) |
|
if module.bias is not None: |
|
module.bias.data.zero_() |
|
elif isinstance(module, nn.Embedding): |
|
module.weight.data.normal_(mean=0.0, std=0.01) |
|
if module.padding_idx is not None: |
|
module.weight.data[module.padding_idx].zero_() |
|
elif isinstance(module, nn.LayerNorm): |
|
module.bias.data.zero_() |
|
module.weight.data.fill_(1.0) |
|
|
|
def forward(self, inputs): |
|
encoder_outputs = self.model.encoder(**inputs) |
|
encoder_hidden_states = encoder_outputs[0] |
|
outputs = self.model.decoder(input_ids=torch.full((inputs['input_ids'].size(0),1), |
|
self.config.decoder_start_token_id, |
|
dtype=torch.long), encoder_hidden_states=encoder_hidden_states) |
|
last_hidden_states = outputs[0] |
|
output1 = self.fc1(last_hidden_states.view(-1, self.config.hidden_size)) |
|
output2 = self.fc2(encoder_hidden_states[:, 0, :].view(-1, self.config.hidden_size)) |
|
output = self.fc3(torch.hstack((output1, output2))) |
|
output = self.fc4(output) |
|
output = self.fc5(output) |
|
return output*100 |
|
|
|
|
|
model = ReactionT5Yield.from_pretrained('sagawa/ReactionT5v2-yield') |
|
tokenizer = AutoTokenizer.from_pretrained('sagawa/ReactionT5v2-yield') |
|
inp = tokenizer(['REACTANT:CC(C)n1ncnc1-c1cn2c(n1)-c1cnc(O)cc1OCC2.CCN(C(C)C)C(C)C.Cl.NC(=O)[C@@H]1C[C@H](F)CN1REAGENT: PRODUCT:O=C(NNC(=O)C(F)(F)F)C(F)(F)F'], return_tensors='pt') |
|
print(model(inp)) # tensor([[19.1666]], grad_fn=<MulBackward0>) |
|
``` |
|
|
|
## Training Details |
|
|
|
### Training Procedure |
|
|
|
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. --> |
|
We used [Open Reaction Database (ORD) dataset](https://drive.google.com/file/d/1fa2MyLdN1vcA7Rysk8kLQENE92YejS9B/view?usp=drive_link) for model training. In addition, we used palladium-catalyzed Buchwald-Hartwig [C-N cross-coupling reactions dataset](https://yzhang.hpc.nyu.edu/T5Chem/index.html)'s test split to prevent data leakage. |
|
The command used for training is the following. For more information about data preprocessing and training, please refer to the paper and GitHub repository. |
|
|
|
```python |
|
python train.py \ |
|
--train_data_path='../data/preprocessed_ord_train.csv' \ |
|
--valid_data_path='../data/preprocessed_ord_valid.csv' \ |
|
--test_data_path='../data/preprocessed_ord_test.csv' \ |
|
--CN_test_data_path='../data/C_N_yield/MFF_Test1/test.csv' \ |
|
--epochs=100 \ |
|
--batch_size=32 \ |
|
--output_dir='./' |
|
``` |
|
|
|
### Results |
|
|
|
| **R^2** | **DFT** | **MFF** | **Yield-BERT**| **T5Chem** | **CompoundT5**| **ReactionT5** (without finetuning) | **ReactionT5** | |
|
| ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | ------------- | |
|
| Random 70/30 | 0.92 | 0.927 ± 0.007 | 0.951 ± 0.005 | 0.970 ± 0.003 | 0.971 ± 0.002 | 0.831 ± 0.012 | 0.947 ± 0.003 | |
|
| Test 1 | 0.80 | 0.851 | 0.838 | 0.811 | 0.855 | 0.846 | 0.872 | |
|
| Test 2 | 0.77 | 0.713 | 0.836 | 0.907 | 0.852 | 0.869 | 0.917 | |
|
| Test 3 | 0.64 | 0.635 | 0.738 | 0.789 | 0.712 | 0.779 | 0.811 | |
|
| Test 4 | 0.54 | 0.184 | 0.538 | 0.627 | 0.547 | 0.843 | 0.830 | |
|
| Avg. Tests 1–4| 0.69 ± 0.104 | 0.596 ± 0.251 | 0.738 ± 0.122 | 0.785 ± 0.094 | 0.741 ± 0.126 | 0.834 ± 0.034 | 0.857 ± 0.041 | |
|
|
|
|
|
## Citation |
|
|
|
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. --> |
|
arxiv link: https://arxiv.org/abs/2311.06708 |
|
``` |
|
@misc{sagawa2023reactiont5, |
|
title={ReactionT5: a large-scale pre-trained model towards application of limited reaction data}, |
|
author={Tatsuya Sagawa and Ryosuke Kojima}, |
|
year={2023}, |
|
eprint={2311.06708}, |
|
archivePrefix={arXiv}, |
|
primaryClass={physics.chem-ph} |
|
} |
|
``` |
