Update inference_fine_tune.py
Browse files- inference_fine_tune.py +9 -95
inference_fine_tune.py
CHANGED
@@ -7,6 +7,15 @@ from pathlib import Path
|
|
7 |
from config import get_config, get_weights_file_path
|
8 |
from train import get_model
|
9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
config = get_config("./openweb.config.json")
|
12 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
@@ -22,101 +31,6 @@ model.eval()
|
|
22 |
state = torch.load(model_path,map_location=torch.device('cpu'))
|
23 |
model.load_state_dict(state['model_state_dict'])
|
24 |
|
25 |
-
def generate_text(
|
26 |
-
model, text, tokenizer, max_len, device,
|
27 |
-
temperature=0.7, top_k=50
|
28 |
-
):
|
29 |
-
eos_idx = tokenizer.token_to_id('</s>')
|
30 |
-
pad_idx = tokenizer.token_to_id('<pad>')
|
31 |
-
|
32 |
-
# Start with the input text as initial decoder input
|
33 |
-
decoder_input = text.to(device)
|
34 |
-
if decoder_input.dim() == 1:
|
35 |
-
decoder_input = decoder_input.unsqueeze(0)
|
36 |
-
|
37 |
-
|
38 |
-
# Print the initial prompt
|
39 |
-
|
40 |
-
while decoder_input.shape[1] < 2000 :
|
41 |
-
# Apply causal mask based on current decoder_input length
|
42 |
-
# decoder_mask = (decoder_input != pad_idx).unsqueeze(0).int() & causal_mask(decoder_input.size(1)).type_as(mask).to(device)
|
43 |
-
|
44 |
-
# Get model output
|
45 |
-
out = model.decode(decoder_input)
|
46 |
-
logits = model.project(out[:, -1]) # Get logits for last token
|
47 |
-
|
48 |
-
# Sampling: temperature + top-k
|
49 |
-
logits = logits / temperature
|
50 |
-
top_k_logits, top_k_indices = torch.topk(logits, top_k)
|
51 |
-
probs = torch.softmax(top_k_logits, dim=-1)
|
52 |
-
next_token = torch.multinomial(probs, num_samples=1)
|
53 |
-
next_token = top_k_indices.gather(-1, next_token)
|
54 |
-
|
55 |
-
# Decode and print token
|
56 |
-
word = tokenizer.decode([next_token.item()])
|
57 |
-
print(word, end="", flush=True)
|
58 |
-
|
59 |
-
# Append next token
|
60 |
-
|
61 |
-
decoder_input = torch.cat([decoder_input, next_token], dim=1)
|
62 |
-
if decoder_input.shape[1] > max_len:
|
63 |
-
decoder_input = decoder_input[:,-max_len:]
|
64 |
-
|
65 |
-
|
66 |
-
if next_token.item() == eos_idx:
|
67 |
-
break
|
68 |
-
|
69 |
-
print()
|
70 |
-
return decoder_input.squeeze(0)
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
def get_tokenizer(config)->Tokenizer:
|
75 |
-
tokenizers_path = Path(config['tokenizer_file'])
|
76 |
-
if Path.exists(tokenizers_path):
|
77 |
-
print("Loading tokenizer from ", tokenizers_path)
|
78 |
-
tokenizer = Tokenizer.from_file(str(tokenizers_path))
|
79 |
-
return tokenizer
|
80 |
-
else:
|
81 |
-
raise FileNotFoundError("Cant find tokenizer file : ",tokenizers_path)
|
82 |
-
|
83 |
-
def run_model(config):
|
84 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
85 |
-
print(f"Using device : {device}")
|
86 |
-
tokenizer = get_tokenizer(config)
|
87 |
-
model = get_model(config, tokenizer.get_vocab_size()).to(device)
|
88 |
-
model_path = get_weights_file_path(config,config['preload'])
|
89 |
-
model.eval()
|
90 |
-
|
91 |
-
if Path.exists(Path(model_path)):
|
92 |
-
print("Loading Model from : ", model_path)
|
93 |
-
state = torch.load(model_path)
|
94 |
-
model.load_state_dict(state['model_state_dict'])
|
95 |
-
print("You : ",end="")
|
96 |
-
input_text = input()
|
97 |
-
pad_token_id = tokenizer.token_to_id("<pad>")
|
98 |
-
user_token_id = tokenizer.token_to_id("<user>")
|
99 |
-
ai_token_id = tokenizer.token_to_id("<ai>")
|
100 |
-
while input_text != "exit":
|
101 |
-
input_tokens = tokenizer.encode(input_text).ids[:-1]
|
102 |
-
input_tokens.extend([user_token_id] + input_tokens + [ai_token_id] )
|
103 |
-
|
104 |
-
if len(input_tokens) > config['seq_len']:
|
105 |
-
print(f"exceeding max length of input : {config['seq_len']}")
|
106 |
-
continue
|
107 |
-
# if len(input_tokens) < config['seq_len']:
|
108 |
-
# input_tokens += [pad_token_id] * (config['seq_len'] - len(input_tokens))
|
109 |
-
input_tokens = torch.tensor(input_tokens)
|
110 |
-
output_tokens = generate_text(model, input_tokens, tokenizer, config['seq_len'], device )
|
111 |
-
print("MODEL : ",output_tokens)
|
112 |
-
output_text = tokenizer.decode(output_tokens.detach().cpu().numpy())
|
113 |
-
# print("Model : "+output_text)
|
114 |
-
print("You : ",end="")
|
115 |
-
input_text = input()
|
116 |
-
|
117 |
-
else:
|
118 |
-
raise FileNotFoundError("Model File not found : "+ model_path)
|
119 |
-
|
120 |
def generate_response(prompt:str):
|
121 |
print("Prompt : ",prompt)
|
122 |
|
|
|
7 |
from config import get_config, get_weights_file_path
|
8 |
from train import get_model
|
9 |
|
10 |
+
def get_tokenizer(config)->Tokenizer:
|
11 |
+
tokenizers_path = Path(config['tokenizer_file'])
|
12 |
+
if Path.exists(tokenizers_path):
|
13 |
+
print("Loading tokenizer from ", tokenizers_path)
|
14 |
+
tokenizer = Tokenizer.from_file(str(tokenizers_path))
|
15 |
+
return tokenizer
|
16 |
+
else:
|
17 |
+
raise FileNotFoundError("Cant find tokenizer file : ",tokenizers_path)
|
18 |
+
|
19 |
|
20 |
config = get_config("./openweb.config.json")
|
21 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
31 |
state = torch.load(model_path,map_location=torch.device('cpu'))
|
32 |
model.load_state_dict(state['model_state_dict'])
|
33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
def generate_response(prompt:str):
|
35 |
print("Prompt : ",prompt)
|
36 |
|