Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -9,102 +9,58 @@ model_name = "./t5-finetuned-final"
|
|
| 9 |
tokenizer = T5Tokenizer.from_pretrained(model_name)
|
| 10 |
model = T5ForConditionalGeneration.from_pretrained(model_name)
|
| 11 |
|
| 12 |
-
# Move model to CPU
|
| 13 |
device = torch.device("cpu")
|
| 14 |
model.to(device)
|
| 15 |
|
| 16 |
def extract_amount(input_text):
|
| 17 |
"""
|
| 18 |
-
Extracts the
|
| 19 |
-
|
|
|
|
|
|
|
| 20 |
"""
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
input_text,
|
| 25 |
-
re.IGNORECASE
|
| 26 |
-
)
|
| 27 |
-
if not amount_match:
|
| 28 |
-
# Fallback: match a number that is immediately followed by a currency symbol/abbreviation.
|
| 29 |
-
amount_match = re.search(
|
| 30 |
-
r'\b(\d+(?:\.\d+)?)\s*(?:AUD|USD|USDT|ETH|BTC|EUR)\b',
|
| 31 |
-
input_text,
|
| 32 |
-
re.IGNORECASE
|
| 33 |
-
)
|
| 34 |
-
if amount_match:
|
| 35 |
-
return amount_match.group(1).strip()
|
| 36 |
return None
|
| 37 |
|
| 38 |
-
def fix_json_output(output):
|
| 39 |
-
"""
|
| 40 |
-
Fixes common JSON formatting issues in the model's output .
|
| 41 |
-
"""
|
| 42 |
-
# Remove trailing commas before closing braces/brackets
|
| 43 |
-
output = re.sub(r',\s*([}\]])', r'\1', output)
|
| 44 |
-
# Fix missing or extra quotes around keys
|
| 45 |
-
output = re.sub(r'([{,])\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*:', r'\1"\2":', output)
|
| 46 |
-
# Fix missing or extra quotes around string values
|
| 47 |
-
output = re.sub(r':\s*([a-zA-Z_][a-zA-Z0-9_]*)\s*([,}])', r':"\1"\2', output)
|
| 48 |
-
return output
|
| 49 |
-
|
| 50 |
def merge_json_with_amount(model_output, amount):
|
| 51 |
"""
|
| 52 |
-
|
| 53 |
-
|
| 54 |
"""
|
| 55 |
try:
|
| 56 |
data = json.loads(model_output)
|
| 57 |
-
except
|
| 58 |
-
|
| 59 |
-
fixed_output = fix_json_output(model_output)
|
| 60 |
-
try:
|
| 61 |
-
data = json.loads(fixed_output)
|
| 62 |
-
except json.JSONDecodeError:
|
| 63 |
-
# If it still fails, return the model output unmodified.
|
| 64 |
-
return model_output
|
| 65 |
-
|
| 66 |
if amount:
|
| 67 |
try:
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
except ValueError:
|
| 71 |
-
# In case conversion fails, keep the original string.
|
| 72 |
data["amount"] = amount
|
| 73 |
-
|
| 74 |
return json.dumps(data, ensure_ascii=False)
|
| 75 |
|
| 76 |
def generate_command(input_command):
|
| 77 |
-
#
|
| 78 |
amount = extract_amount(input_command)
|
| 79 |
|
| 80 |
-
# Generate the JSON output
|
| 81 |
prompt = "extract: " + input_command
|
| 82 |
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
| 83 |
-
|
| 84 |
-
output_ids = model.generate(
|
| 85 |
-
input_ids,
|
| 86 |
-
max_length=128, # Increased max_length to allow for complete JSON output
|
| 87 |
-
num_beams=2, # Using beam search for better output quality
|
| 88 |
-
early_stopping=True
|
| 89 |
-
)
|
| 90 |
-
|
| 91 |
model_output = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 92 |
|
| 93 |
-
# Merge the
|
| 94 |
-
|
| 95 |
-
if amount:
|
| 96 |
-
result = merge_json_with_amount(model_output, amount)
|
| 97 |
-
else:
|
| 98 |
-
result = model_output # Use the model's output as-is if no amount is found
|
| 99 |
-
|
| 100 |
return result
|
| 101 |
|
| 102 |
iface = gr.Interface(
|
| 103 |
fn=generate_command,
|
| 104 |
inputs=gr.Textbox(lines=2, placeholder="Enter a command..."),
|
| 105 |
outputs=gr.Textbox(label="Extracted JSON Output"),
|
| 106 |
-
title="T5
|
| 107 |
-
description="
|
| 108 |
)
|
| 109 |
|
| 110 |
if __name__ == "__main__":
|
|
|
|
| 9 |
tokenizer = T5Tokenizer.from_pretrained(model_name)
|
| 10 |
model = T5ForConditionalGeneration.from_pretrained(model_name)
|
| 11 |
|
| 12 |
+
# Move model to CPU explicitly
|
| 13 |
device = torch.device("cpu")
|
| 14 |
model.to(device)
|
| 15 |
|
| 16 |
def extract_amount(input_text):
|
| 17 |
"""
|
| 18 |
+
Extracts the first number (with optional decimals) from the input text.
|
| 19 |
+
For example, in:
|
| 20 |
+
"Should I send 2659.53464 EUR to my wife today?"
|
| 21 |
+
it returns the string "2659.53464".
|
| 22 |
"""
|
| 23 |
+
match = re.search(r'\b(\d+(?:\.\d+)?)\b', input_text)
|
| 24 |
+
if match:
|
| 25 |
+
return match.group(1)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
return None
|
| 27 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
def merge_json_with_amount(model_output, amount):
|
| 29 |
"""
|
| 30 |
+
Parses the model's JSON output and overrides the "amount" key
|
| 31 |
+
with the manually extracted value.
|
| 32 |
"""
|
| 33 |
try:
|
| 34 |
data = json.loads(model_output)
|
| 35 |
+
except Exception:
|
| 36 |
+
data = {}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
if amount:
|
| 38 |
try:
|
| 39 |
+
data["amount"] = float(amount)
|
| 40 |
+
except Exception:
|
|
|
|
|
|
|
| 41 |
data["amount"] = amount
|
|
|
|
| 42 |
return json.dumps(data, ensure_ascii=False)
|
| 43 |
|
| 44 |
def generate_command(input_command):
|
| 45 |
+
# Manually extract the amount from the input.
|
| 46 |
amount = extract_amount(input_command)
|
| 47 |
|
| 48 |
+
# Generate the JSON output from the model.
|
| 49 |
prompt = "extract: " + input_command
|
| 50 |
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
|
| 51 |
+
output_ids = model.generate(input_ids, max_length=128, num_beams=2, early_stopping=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
model_output = tokenizer.decode(output_ids[0], skip_special_tokens=True)
|
| 53 |
|
| 54 |
+
# Merge the manually extracted amount into the model output.
|
| 55 |
+
result = merge_json_with_amount(model_output, amount)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
return result
|
| 57 |
|
| 58 |
iface = gr.Interface(
|
| 59 |
fn=generate_command,
|
| 60 |
inputs=gr.Textbox(lines=2, placeholder="Enter a command..."),
|
| 61 |
outputs=gr.Textbox(label="Extracted JSON Output"),
|
| 62 |
+
title="T5 Command Extractor",
|
| 63 |
+
description="The model provides action, currency, and recipient. The amount is manually extracted from the input."
|
| 64 |
)
|
| 65 |
|
| 66 |
if __name__ == "__main__":
|