refactor: Modularize code by creating functions for model loading and processing
Browse files
app.py
CHANGED
@@ -6,117 +6,68 @@ from pprint import pprint
|
|
6 |
|
7 |
#%%
|
8 |
|
9 |
-
model_name
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
model
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
with torch.no_grad():
|
32 |
-
outputs = model(**inputs, labels=labels)
|
33 |
-
|
34 |
-
#%%
|
35 |
-
|
36 |
-
# Get logits and shift them
|
37 |
-
logits = outputs.logits[0, :-1, :]
|
38 |
-
|
39 |
-
# Calculate log probabilities
|
40 |
-
log_probs = torch.log_softmax(logits, dim=-1)
|
41 |
-
|
42 |
-
# Get the log probability of each token in the sequence
|
43 |
-
token_log_probs = log_probs[range(log_probs.shape[0]), input_ids[0][1:]]
|
44 |
-
|
45 |
-
# Decode tokens and pair them with their log probabilities
|
46 |
-
tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
|
47 |
-
result = list(zip(tokens[1:], token_log_probs.tolist()))
|
48 |
-
|
49 |
-
#%%
|
50 |
-
for token, logprob in result:
|
51 |
-
print(f"Token: {token}, Log Probability: {logprob:.4f}")
|
52 |
-
|
53 |
-
# %%
|
54 |
-
words = []
|
55 |
-
current_word = []
|
56 |
-
current_log_probs = []
|
57 |
-
|
58 |
-
for token, logprob in result:
|
59 |
-
if not token.startswith(chr(9601)) and token.isalpha():
|
60 |
-
current_word.append(token)
|
61 |
-
current_log_probs.append(logprob)
|
62 |
-
else:
|
63 |
-
if current_word:
|
64 |
-
words.append(("".join(current_word), sum(current_log_probs)))
|
65 |
-
current_word = [token]
|
66 |
-
current_log_probs = [logprob]
|
67 |
-
|
68 |
-
if current_word:
|
69 |
-
words.append(("".join(current_word), sum(current_log_probs)))
|
70 |
-
|
71 |
-
for word, avg_logprob in words:
|
72 |
-
print(f"Word: {word}, Log Probability: {avg_logprob:.4f}")
|
73 |
-
|
74 |
-
# %%
|
75 |
-
|
76 |
-
|
77 |
-
words = split_into_words(tokens[1:], token_log_probs)
|
78 |
-
|
79 |
-
# Define a threshold for low probability words
|
80 |
-
log_prob_threshold = -5.0
|
81 |
-
|
82 |
-
# Filter words with log probability below the threshold
|
83 |
-
low_prob_words = [word for word in words if word.logprob < log_prob_threshold]
|
84 |
-
|
85 |
-
|
86 |
-
#%%
|
87 |
-
def generate_replacements(model, tokenizer, prefix, num_samples=5):
|
88 |
input_context = tokenizer(prefix, return_tensors="pt").to(device)
|
89 |
input_ids = input_context["input_ids"]
|
90 |
-
|
91 |
new_words = []
|
92 |
for _ in range(num_samples):
|
93 |
with torch.no_grad():
|
94 |
outputs = model.generate(
|
95 |
input_ids=input_ids,
|
96 |
-
max_length=input_ids.shape[-1] + 5,
|
97 |
num_return_sequences=1,
|
98 |
temperature=1.0,
|
99 |
-
top_k=50,
|
100 |
-
top_p=0.95,
|
101 |
do_sample=True
|
102 |
)
|
103 |
-
|
104 |
-
generated_ids = outputs[0][input_ids.shape[-1]:] # extract the newly generated part
|
105 |
new_word = tokenizer.decode(generated_ids, skip_special_tokens=True).split()[0]
|
106 |
new_words.append(new_word)
|
107 |
-
|
108 |
return new_words
|
109 |
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
|
|
|
|
|
|
115 |
|
116 |
-
|
|
|
|
|
117 |
|
118 |
-
|
119 |
-
|
120 |
-
|
|
|
|
|
|
|
|
|
|
|
121 |
|
122 |
-
|
|
|
|
6 |
|
7 |
#%%
|
8 |
|
9 |
+
def load_model_and_tokenizer(model_name):
|
10 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
11 |
+
model = AutoModelForCausalLM.from_pretrained(model_name)
|
12 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
13 |
+
model.to(device)
|
14 |
+
return model, tokenizer, device
|
15 |
+
|
16 |
+
def process_input_text(input_text, tokenizer, device):
|
17 |
+
inputs = tokenizer(input_text, return_tensors="pt").to(device)
|
18 |
+
input_ids = inputs["input_ids"]
|
19 |
+
return inputs, input_ids
|
20 |
+
|
21 |
+
def calculate_log_probabilities(model, inputs, input_ids):
|
22 |
+
with torch.no_grad():
|
23 |
+
outputs = model(**inputs, labels=input_ids)
|
24 |
+
logits = outputs.logits[0, :-1, :]
|
25 |
+
log_probs = torch.log_softmax(logits, dim=-1)
|
26 |
+
token_log_probs = log_probs[range(log_probs.shape[0]), input_ids[0][1:]]
|
27 |
+
tokens = tokenizer.convert_ids_to_tokens(input_ids[0])
|
28 |
+
return list(zip(tokens[1:], token_log_probs.tolist()))
|
29 |
+
|
30 |
+
def generate_replacements(model, tokenizer, prefix, device, num_samples=5):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
input_context = tokenizer(prefix, return_tensors="pt").to(device)
|
32 |
input_ids = input_context["input_ids"]
|
|
|
33 |
new_words = []
|
34 |
for _ in range(num_samples):
|
35 |
with torch.no_grad():
|
36 |
outputs = model.generate(
|
37 |
input_ids=input_ids,
|
38 |
+
max_length=input_ids.shape[-1] + 5,
|
39 |
num_return_sequences=1,
|
40 |
temperature=1.0,
|
41 |
+
top_k=50,
|
42 |
+
top_p=0.95,
|
43 |
do_sample=True
|
44 |
)
|
45 |
+
generated_ids = outputs[0][input_ids.shape[-1]:]
|
|
|
46 |
new_word = tokenizer.decode(generated_ids, skip_special_tokens=True).split()[0]
|
47 |
new_words.append(new_word)
|
|
|
48 |
return new_words
|
49 |
|
50 |
+
def main():
|
51 |
+
model_name = "mistralai/Mistral-7B-v0.1"
|
52 |
+
model, tokenizer, device = load_model_and_tokenizer(model_name)
|
53 |
+
|
54 |
+
input_text = "He asked me to prostrate myself before the king, but I rifused."
|
55 |
+
inputs, input_ids = process_input_text(input_text, tokenizer, device)
|
56 |
+
|
57 |
+
result = calculate_log_probabilities(model, inputs, input_ids)
|
58 |
|
59 |
+
words = split_into_words([token for token, _ in result], [logprob for _, logprob in result])
|
60 |
+
log_prob_threshold = -5.0
|
61 |
+
low_prob_words = [word for word in words if word.logprob < log_prob_threshold]
|
62 |
|
63 |
+
for word in low_prob_words:
|
64 |
+
prefix_index = word.first_token_index
|
65 |
+
prefix_tokens = [token for token, _ in result][:prefix_index + 1]
|
66 |
+
prefix = tokenizer.convert_tokens_to_string(prefix_tokens)
|
67 |
+
replacements = generate_replacements(model, tokenizer, prefix, device)
|
68 |
+
print(f"Original word: {word.text}, Log Probability: {word.logprob:.4f}")
|
69 |
+
print(f"Proposed replacements: {replacements}")
|
70 |
+
print()
|
71 |
|
72 |
+
if __name__ == "__main__":
|
73 |
+
main()
|