Is there a script file for model inference in GEM? about paddlehelix HOT 3 OPEN

Hi gaoshan, thank you for using paddlehelix! As for the model inference, you can refer to the
Part III: Downstream Inference of the pretrained_compound tutorial. Hope this can be helpful to you.

from paddlehelix.

Hi gaoshan, thank you for using paddlehelix! As for the model inference, you can refer to the Part III: Downstream Inference of the pretrained_compound tutorial. Hope this can be helpful to you.

Thanks! I will try

from paddlehelix.

I refer to the above tutorial pointed by @Noisyntrain to run an inference case for a finetuned qm7 model, However , it looks the tutorial code does not work for qm7 model. Here is my code , just a few slight changes from the tutorial example ，

All the json config file of inference is same with that of model training

def main(args):
compound_encoder_config = load_json_config( './model_configs/geognn_l8.json')
task_type = 'regr'
dataset_name = 'qm7'
task_names = get_downstream_task_names(dataset_name, './chemrl_downstream_datasets/qm7')

model_config = load_json_config( 'model_configs/down_mlp3.json' )
model_config['task_type'] = task_type
model_config['num_tasks'] = len(task_names)

compound_encoder = GeoGNNModel(compound_encoder_config)
model = DownstreamModel(model_config, compound_encoder)
model.set_state_dict(paddle.load( './model/model.pdparams' ))  <- this model is trained by the same config file
transform_fn = DownstreamTransformFn(is_inference=True )

collate_fn = DownstreamCollateFn(
        atom_names=compound_encoder_config['atom_names'], 
        bond_names=compound_encoder_config['bond_names'],
        bond_float_names=compound_encoder_config['bond_float_names'],
        bond_angle_float_names=compound_encoder_config['bond_angle_float_names'],
        is_inference=True,
        task_type=task_type)

SMILES="Cc1c(O)nc2ccccn2c1=O"

graph=collate_fn([transform_fn({'smiles':SMILES})])
preds=model(graph.tensor()).numpy()[0]

print('SMILES:%s' % SMILES)
print('Predictions:')
print(str(preds))
for name,prob in zip(task_names.preds):
     print(" %s:\t%s" % (name, prob))

Then I got the following errors:

Traceback (most recent call last):
File "inference_regr.py", line 206, in
main(args)
File "inference_regr.py", line 177, in main
preds=model(graph.tensor()).numpy()[0]
AttributeError: 'tuple' object has no attribute 'tensor'

I do not understand what differences between 'qm7' model and the tutorial model ? It looks the tutorial should be applicable to all the finetune models here, right? Could you give me some hints please？ Thanks in advance

from paddlehelix.