Update app.py
Browse files
app.py
CHANGED
@@ -9,6 +9,40 @@ from transformers import AutoModelForCausalLM
|
|
9 |
from transformers import AutoTokenizer
|
10 |
# from next_word_prediction import GPT2
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
from share_btn import community_icon_html, loading_icon_html, share_js
|
13 |
|
14 |
# get gpt2 model
|
|
|
9 |
from transformers import AutoTokenizer
|
10 |
# from next_word_prediction import GPT2
|
11 |
|
12 |
+
### code
|
13 |
+
gpt2 = AutoModelForCausalLM.from_pretrained("gpt2", return_dict_in_generate=True)
|
14 |
+
tokenizer = AutoTokenizer.from_pretrained("gpt2")
|
15 |
+
|
16 |
+
input_ids = tokenizer("Today is a nice day", return_tensors="pt").input_ids
|
17 |
+
|
18 |
+
generated_outputs = gpt2.generate(input_ids, do_sample=True, num_return_sequences=3, output_scores=True)
|
19 |
+
|
20 |
+
# only use id's that were generated
|
21 |
+
# gen_sequences has shape [3, 15]
|
22 |
+
gen_sequences = generated_outputs.sequences[:, input_ids.shape[-1]:]
|
23 |
+
|
24 |
+
# let's stack the logits generated at each step to a tensor and transform
|
25 |
+
# logits to probs
|
26 |
+
probs = torch.stack(generated_outputs.scores, dim=1).softmax(-1) # -> shape [3, 15, vocab_size]
|
27 |
+
|
28 |
+
# now we need to collect the probability of the generated token
|
29 |
+
# we need to add a dummy dim in the end to make gather work
|
30 |
+
gen_probs = torch.gather(probs, 2, gen_sequences[:, :, None]).squeeze(-1)
|
31 |
+
|
32 |
+
# now we can do all kinds of things with the probs
|
33 |
+
|
34 |
+
# 1) the probs that exactly those sequences are generated again
|
35 |
+
# those are normally going to be very small
|
36 |
+
unique_prob_per_sequence = gen_probs.prod(-1)
|
37 |
+
|
38 |
+
# 2) normalize the probs over the three sequences
|
39 |
+
normed_gen_probs = gen_probs / gen_probs.sum(0)
|
40 |
+
assert normed_gen_probs[:, 0].sum() == 1.0, "probs should be normalized"
|
41 |
+
|
42 |
+
# 3) compare normalized probs to each other like in 1)
|
43 |
+
unique_normed_prob_per_sequence = normed_gen_probs.prod(-1)
|
44 |
+
|
45 |
+
### end code
|
46 |
from share_btn import community_icon_html, loading_icon_html, share_js
|
47 |
|
48 |
# get gpt2 model
|