Update README.md
Browse files
README.md
CHANGED
|
@@ -48,8 +48,6 @@ Then, we'll load the model weights from HuggingFace Hub and the GO graph using `
|
|
| 48 |
```python
|
| 49 |
import torch
|
| 50 |
|
| 51 |
-
import obonet
|
| 52 |
-
|
| 53 |
from esm.tokenization import EsmSequenceTokenizer
|
| 54 |
|
| 55 |
from esmc_function_classifier.model import EsmcGoTermClassifier
|
|
@@ -57,28 +55,53 @@ from esmc_function_classifier.model import EsmcGoTermClassifier
|
|
| 57 |
|
| 58 |
model_name = "andrewdalpino/ESMC-300M-Protein-Function"
|
| 59 |
|
| 60 |
-
# Visit https://geneontology.org/docs/download-ontology/ to download.
|
| 61 |
-
go_db_path = "./dataset/go-basic.obo"
|
| 62 |
-
|
| 63 |
sequence = "MPPKGHKKTADGDFRPVNSAGNTIQAKQKYSIDDLLYPKSTIKNLAKETLPDDAIISKDALTAIQRAATLFVSYMASHGNASAEAGGRKKIT"
|
| 64 |
|
| 65 |
top_p = 0.5
|
| 66 |
|
| 67 |
-
graph = obonet.read_obo(go_db_path)
|
| 68 |
-
|
| 69 |
tokenizer = EsmSequenceTokenizer()
|
| 70 |
|
| 71 |
model = EsmcGoTermClassifier.from_pretrained(model_name)
|
| 72 |
|
| 73 |
-
model.load_gene_ontology(graph)
|
| 74 |
-
|
| 75 |
out = tokenizer(sequence, max_length=2048, truncation=True)
|
| 76 |
|
| 77 |
input_ids = torch.tensor(out["input_ids"], dtype=torch.int64)
|
| 78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
subgraph, go_term_probabilities = model.predict_subgraph(
|
| 80 |
input_ids, top_p=top_p
|
| 81 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
```
|
| 83 |
|
| 84 |
## Code Repository
|
|
|
|
| 48 |
```python
|
| 49 |
import torch
|
| 50 |
|
|
|
|
|
|
|
| 51 |
from esm.tokenization import EsmSequenceTokenizer
|
| 52 |
|
| 53 |
from esmc_function_classifier.model import EsmcGoTermClassifier
|
|
|
|
| 55 |
|
| 56 |
model_name = "andrewdalpino/ESMC-300M-Protein-Function"
|
| 57 |
|
|
|
|
|
|
|
|
|
|
| 58 |
sequence = "MPPKGHKKTADGDFRPVNSAGNTIQAKQKYSIDDLLYPKSTIKNLAKETLPDDAIISKDALTAIQRAATLFVSYMASHGNASAEAGGRKKIT"
|
| 59 |
|
| 60 |
top_p = 0.5
|
| 61 |
|
|
|
|
|
|
|
| 62 |
tokenizer = EsmSequenceTokenizer()
|
| 63 |
|
| 64 |
model = EsmcGoTermClassifier.from_pretrained(model_name)
|
| 65 |
|
|
|
|
|
|
|
| 66 |
out = tokenizer(sequence, max_length=2048, truncation=True)
|
| 67 |
|
| 68 |
input_ids = torch.tensor(out["input_ids"], dtype=torch.int64)
|
| 69 |
|
| 70 |
+
go_term_probabilities = model.predict_terms(
|
| 71 |
+
input_ids, top_p=top_p
|
| 72 |
+
)
|
| 73 |
+
```
|
| 74 |
+
|
| 75 |
+
You can also output the gene-ontology (GO) `networkx` subgraph for a given sequence like in the example below. You'll need an up-to-date gene ontology database that you can import using the `obonet` package.
|
| 76 |
+
|
| 77 |
+
```python
|
| 78 |
+
import networkx as nx
|
| 79 |
+
|
| 80 |
+
import obonet
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
# Visit https://geneontology.org/docs/download-ontology/ to download.
|
| 84 |
+
go_db_path = "./dataset/go-basic.obo"
|
| 85 |
+
|
| 86 |
+
graph = obonet.read_obo(go_db_path)
|
| 87 |
+
|
| 88 |
+
model.load_gene_ontology(graph)
|
| 89 |
+
|
| 90 |
subgraph, go_term_probabilities = model.predict_subgraph(
|
| 91 |
input_ids, top_p=top_p
|
| 92 |
)
|
| 93 |
+
|
| 94 |
+
json = nx.node_link_data(subgraph)
|
| 95 |
+
|
| 96 |
+
print(json)
|
| 97 |
+
```
|
| 98 |
+
|
| 99 |
+
### Quantized Model
|
| 100 |
+
|
| 101 |
+
To quantize the model weights using int8 call the `quantize_weights()` method. Any model can be quantized, but we recommend one that has been quantization-aware trained (QAT) for the best performance. The `group_size` argument controls the granularity at which quantization scales are computed.
|
| 102 |
+
|
| 103 |
+
```python
|
| 104 |
+
model.quantize_weights(group_size=64)
|
| 105 |
```
|
| 106 |
|
| 107 |
## Code Repository
|