Snapshot
Browse files
app.py
CHANGED
@@ -55,8 +55,10 @@ def tokenize(input_text: str, tokenizer: Tokenizer, device: torch.device) -> Bat
|
|
55 |
return tokenizer(input_text, return_tensors="pt").to(device)
|
56 |
|
57 |
def calculate_log_probabilities(model: PreTrainedModel, tokenizer: Tokenizer, inputs: BatchEncoding) -> list[tuple[int, float]]:
|
|
|
|
|
58 |
with torch.no_grad():
|
59 |
-
outputs = model(input_ids=
|
60 |
# B x T x V
|
61 |
logits: torch.Tensor = outputs.logits[:, :-1, :]
|
62 |
# B x T x V
|
@@ -71,8 +73,7 @@ def prepare_inputs(contexts: list[list[int]], tokenizer: Tokenizer, device: torc
|
|
71 |
texts = [tokenizer.decode(context, skip_special_tokens=True) for context in contexts]
|
72 |
return tokenizer(texts, return_tensors="pt", padding=True).to(device)
|
73 |
|
74 |
-
def
|
75 |
-
device: torch.device, num_samples: int = 5) -> tuple[GenerateOutput | torch.LongTensor, list[list[str]]]:
|
76 |
input_ids = inputs["input_ids"]
|
77 |
attention_mask = inputs["attention_mask"]
|
78 |
with torch.no_grad():
|
@@ -86,16 +87,19 @@ def generate_replacements(model: PreTrainedModel, tokenizer: Tokenizer, inputs:
|
|
86 |
top_p=0.95,
|
87 |
do_sample=True
|
88 |
)
|
|
|
|
|
|
|
89 |
all_new_words = []
|
90 |
-
for i in range(
|
91 |
replacements = []
|
92 |
for j in range(num_samples):
|
93 |
-
generated_ids = outputs[i * num_samples + j][
|
94 |
new_word = tokenizer.convert_ids_to_tokens(generated_ids.tolist())[0]
|
95 |
if new_word.startswith(chr(9601)):
|
96 |
replacements.append(new_word)
|
97 |
all_new_words.append(replacements)
|
98 |
-
return
|
99 |
|
100 |
#%%
|
101 |
|
@@ -126,11 +130,15 @@ input_ids = inputs["input_ids"]
|
|
126 |
|
127 |
#%%
|
128 |
|
|
|
129 |
start_time = time.time()
|
130 |
-
outputs
|
131 |
end_time = time.time()
|
132 |
print(f"Total time taken for replacements: {end_time - start_time:.4f} seconds")
|
133 |
|
|
|
|
|
|
|
134 |
#%%
|
135 |
|
136 |
for word, replacements in zip(low_prob_words, replacements_batch):
|
|
|
55 |
return tokenizer(input_text, return_tensors="pt").to(device)
|
56 |
|
57 |
def calculate_log_probabilities(model: PreTrainedModel, tokenizer: Tokenizer, inputs: BatchEncoding) -> list[tuple[int, float]]:
|
58 |
+
input_ids = inputs["input_ids"]
|
59 |
+
attention_mask = inputs["attention_mask"]
|
60 |
with torch.no_grad():
|
61 |
+
outputs = model(input_ids=input_ids, attention_mask=attention_mask, labels=input_ids)
|
62 |
# B x T x V
|
63 |
logits: torch.Tensor = outputs.logits[:, :-1, :]
|
64 |
# B x T x V
|
|
|
73 |
texts = [tokenizer.decode(context, skip_special_tokens=True) for context in contexts]
|
74 |
return tokenizer(texts, return_tensors="pt", padding=True).to(device)
|
75 |
|
76 |
+
def generate_outputs(model: PreTrainedModel, inputs: BatchEncoding, num_samples: int = 5) -> GenerateOutput | torch.LongTensor:
|
|
|
77 |
input_ids = inputs["input_ids"]
|
78 |
attention_mask = inputs["attention_mask"]
|
79 |
with torch.no_grad():
|
|
|
87 |
top_p=0.95,
|
88 |
do_sample=True
|
89 |
)
|
90 |
+
return outputs
|
91 |
+
|
92 |
+
def extract_replacements(outputs: GenerateOutput | torch.LongTensor, tokenizer: Tokenizer, num_inputs: int, input_len: int, num_samples: int = 5) -> list[list[str]]:
|
93 |
all_new_words = []
|
94 |
+
for i in range(num_inputs):
|
95 |
replacements = []
|
96 |
for j in range(num_samples):
|
97 |
+
generated_ids = outputs[i * num_samples + j][input_len:]
|
98 |
new_word = tokenizer.convert_ids_to_tokens(generated_ids.tolist())[0]
|
99 |
if new_word.startswith(chr(9601)):
|
100 |
replacements.append(new_word)
|
101 |
all_new_words.append(replacements)
|
102 |
+
return all_new_words
|
103 |
|
104 |
#%%
|
105 |
|
|
|
130 |
|
131 |
#%%
|
132 |
|
133 |
+
num_samples = 5
|
134 |
start_time = time.time()
|
135 |
+
outputs = generate_outputs(model, inputs, num_samples)
|
136 |
end_time = time.time()
|
137 |
print(f"Total time taken for replacements: {end_time - start_time:.4f} seconds")
|
138 |
|
139 |
+
#%%
|
140 |
+
replacements_batch = extract_replacements(outputs, tokenizer, input_ids.shape[0], input_ids.shape[1], num_samples)
|
141 |
+
|
142 |
#%%
|
143 |
|
144 |
for word, replacements in zip(low_prob_words, replacements_batch):
|