Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
The *OpenAI* API allows to retrieve log-probabilities per token (including both prompt and completion tokens) through the ``logprobs`` return argument. Currently, the ``CausalLM`` only provide ``logits`` return values, which should are the prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
|
2 |
+
|
3 |
+
The following code provides an example of how to retrieve the log-probabilities per token of ``CausalLMs`` for the huggingface API:
|
4 |
+
|
5 |
+
```python
|
6 |
+
def logprobs_from_prompt(prompt, tokenizer, model):
|
7 |
+
encoded = tokenizer(prompt, return_tensors="pt").to("cpu")
|
8 |
+
input_ids = encoded["input_ids"]
|
9 |
+
output = model(input_ids=input_ids)
|
10 |
+
shift_labels = input_ids[..., 1:].contiguous()
|
11 |
+
shift_logits = output.logits[..., :-1, :].contiguous()
|
12 |
+
log_probs = []
|
13 |
+
log_probs.append((tokenizer.decode(input_ids[0].tolist()[0]), None))
|
14 |
+
for idx, (label_id, logit) in enumerate(zip(shift_labels[0].tolist(), shift_logits[0])):
|
15 |
+
logprob = F.log_softmax(logit, dim=0).tolist()[label_id]
|
16 |
+
log_probs.append((tokenizer.decode(label_id), float(logprob)))
|
17 |
+
return log_probs
|
18 |
+
```
|
19 |
+
|
20 |
+
An example call would be:
|
21 |
+
```python
|
22 |
+
tokenizer = GPT2Tokenizer.from_pretrained("facebook/opt")
|
23 |
+
model = OPTForCausalLM.from_pretrained("facebook/opt")
|
24 |
+
prompt = "The horse raced past the barn fell."
|
25 |
+
logprobs = logprobs_from_prompt(prompt, tokenizer, model)
|
26 |
+
|
27 |
+
```
|
28 |
+
|
29 |
+
For its derivation and explanation see this [discussion](https://huggingface.co/bigscience/bloom/discussions/89#6321dcc9b97c618f9a5e3dac).
|