Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -6,6 +6,7 @@ import gradio as gr
|
|
6 |
import torch
|
7 |
from utils import *
|
8 |
from presets import *
|
|
|
9 |
|
10 |
|
11 |
######################################################################
|
@@ -19,12 +20,9 @@ base_model = "project-baize/baize-v2-7b" #load_8bit = False (in load_tokenizer_
|
|
19 |
tokenizer,model,device = load_tokenizer_and_model(base_model,False)
|
20 |
dataset_neu = daten_laden("alexkueck/tis")
|
21 |
|
22 |
-
|
23 |
-
#Vorbereiten für das
|
24 |
-
|
25 |
-
def tokenize_function(examples):
|
26 |
-
return tokenizer(examples["text"])
|
27 |
-
|
28 |
|
29 |
#alles zusammen auf das neue datenset anwenden - batched = True und 4 Prozesse, um die Berechnung zu beschleunigen. Die "text" - Spalte braucht man anschließend nicht mehr, daher weglassen.
|
30 |
tokenized_datasets = dataset_neu.map(tokenize_function, batched=True, num_proc=4, remove_columns=["text"])
|
@@ -36,203 +34,84 @@ tokenized_datasets = dataset_neu.map(tokenize_function, batched=True, num_proc=4
|
|
36 |
# block_size = tokenizer.model_max_length
|
37 |
block_size = 128
|
38 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
|
|
|
|
|
40 |
|
|
|
|
|
41 |
|
42 |
-
########################################################################
|
43 |
-
#Chat KI nutzen, um Text zu generieren...
|
44 |
-
def predict(text,
|
45 |
-
chatbotGr,
|
46 |
-
history,
|
47 |
-
top_p,
|
48 |
-
temperature,
|
49 |
-
max_length_tokens,
|
50 |
-
max_context_length_tokens,):
|
51 |
-
if text=="":
|
52 |
-
yield chatbotGr,history,"Empty context."
|
53 |
-
return
|
54 |
-
try:
|
55 |
-
model
|
56 |
-
except:
|
57 |
-
yield [[text,"No Model Found"]],[],"No Model Found"
|
58 |
-
return
|
59 |
-
|
60 |
-
inputs = generate_prompt_with_history(text,history,tokenizer,max_length=max_context_length_tokens)
|
61 |
-
if inputs is None:
|
62 |
-
yield chatbotGr,history,"Input too long."
|
63 |
-
return
|
64 |
-
else:
|
65 |
-
prompt,inputs=inputs
|
66 |
-
begin_length = len(prompt)
|
67 |
-
|
68 |
-
input_ids = inputs["input_ids"][:,-max_context_length_tokens:].to(device)
|
69 |
-
torch.cuda.empty_cache()
|
70 |
-
|
71 |
-
#torch.no_grad() bedeutet, dass für die betreffenden tensoren keine Ableitungen berechnet werden bei der backpropagation
|
72 |
-
#hier soll das NN ja auch nicht geändert werden 8backprop ist nicht nötig), da es um interference-prompts geht!
|
73 |
-
with torch.no_grad():
|
74 |
-
#die vergangenen prompts werden alle als Tupel in history abgelegt sortiert nach 'Human' und 'AI'- dass sind daher auch die stop-words, die den jeweils nächsten Eintrag kennzeichnen
|
75 |
-
for x in greedy_search(input_ids,model,tokenizer,stop_words=["[|Human|]", "[|AI|]"],max_length=max_length_tokens,temperature=temperature,top_p=top_p):
|
76 |
-
if is_stop_word_or_prefix(x,["[|Human|]", "[|AI|]"]) is False:
|
77 |
-
if "[|Human|]" in x:
|
78 |
-
x = x[:x.index("[|Human|]")].strip()
|
79 |
-
if "[|AI|]" in x:
|
80 |
-
x = x[:x.index("[|AI|]")].strip()
|
81 |
-
x = x.strip()
|
82 |
-
a, b= [[y[0],convert_to_markdown(y[1])] for y in history]+[[text, convert_to_markdown(x)]],history + [[text,x]]
|
83 |
-
yield a, b, "Generating..."
|
84 |
-
if shared_state.interrupted:
|
85 |
-
shared_state.recover()
|
86 |
-
try:
|
87 |
-
yield a, b, "Stop: Success"
|
88 |
-
return
|
89 |
-
except:
|
90 |
-
pass
|
91 |
-
del input_ids
|
92 |
-
gc.collect()
|
93 |
-
torch.cuda.empty_cache()
|
94 |
-
|
95 |
-
try:
|
96 |
-
yield a,b,"Generate: Success"
|
97 |
-
except:
|
98 |
-
pass
|
99 |
|
100 |
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
#Übersetzungs Ki nutzen
|
109 |
-
def translate():
|
110 |
-
return "Kommt noch!"
|
111 |
|
112 |
-
#
|
113 |
-
def
|
114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
115 |
|
116 |
#######################################################################
|
117 |
#Darstellung mit Gradio
|
118 |
|
119 |
-
with
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
user_question = gr.State("")
|
125 |
-
gr.Markdown("KIs am LI - wähle aus, was du bzgl. KI-Bots ausprobieren möchtest!")
|
126 |
-
with gr.Tabs():
|
127 |
-
with gr.TabItem("LI-Chat"):
|
128 |
-
with gr.Row():
|
129 |
-
gr.HTML(title)
|
130 |
-
status_display = gr.Markdown("Erfolg", elem_id="status_display")
|
131 |
-
gr.Markdown(description_top)
|
132 |
-
with gr.Row(scale=1).style(equal_height=True):
|
133 |
-
with gr.Column(scale=5):
|
134 |
-
with gr.Row(scale=1):
|
135 |
-
chatbotGr = gr.Chatbot(elem_id="LI_chatbot").style(height="100%")
|
136 |
-
with gr.Row(scale=1):
|
137 |
-
with gr.Column(scale=12):
|
138 |
-
user_input = gr.Textbox(
|
139 |
-
show_label=False, placeholder="Gib deinen Text / Frage ein."
|
140 |
-
).style(container=False)
|
141 |
-
with gr.Column(min_width=100, scale=1):
|
142 |
-
submitBtn = gr.Button("Absenden")
|
143 |
-
with gr.Column(min_width=100, scale=1):
|
144 |
-
cancelBtn = gr.Button("Stoppen")
|
145 |
-
with gr.Row(scale=1):
|
146 |
-
emptyBtn = gr.Button(
|
147 |
-
"🧹 Neuer Chat",
|
148 |
-
)
|
149 |
-
with gr.Column():
|
150 |
-
with gr.Column(min_width=50, scale=1):
|
151 |
-
with gr.Tab(label="Parameter zum Model"):
|
152 |
-
gr.Markdown("# Parameters")
|
153 |
-
top_p = gr.Slider(
|
154 |
-
minimum=-0,
|
155 |
-
maximum=1.0,
|
156 |
-
value=0.95,
|
157 |
-
step=0.05,
|
158 |
-
interactive=True,
|
159 |
-
label="Top-p",
|
160 |
-
)
|
161 |
-
temperature = gr.Slider(
|
162 |
-
minimum=0.1,
|
163 |
-
maximum=2.0,
|
164 |
-
value=1,
|
165 |
-
step=0.1,
|
166 |
-
interactive=True,
|
167 |
-
label="Temperature",
|
168 |
-
)
|
169 |
-
max_length_tokens = gr.Slider(
|
170 |
-
minimum=0,
|
171 |
-
maximum=512,
|
172 |
-
value=512,
|
173 |
-
step=8,
|
174 |
-
interactive=True,
|
175 |
-
label="Max Generation Tokens",
|
176 |
-
)
|
177 |
-
max_context_length_tokens = gr.Slider(
|
178 |
-
minimum=0,
|
179 |
-
maximum=4096,
|
180 |
-
value=2048,
|
181 |
-
step=128,
|
182 |
-
interactive=True,
|
183 |
-
label="Max History Tokens",
|
184 |
-
)
|
185 |
-
gr.Markdown(description)
|
186 |
-
|
187 |
-
with gr.TabItem("Übersetzungen"):
|
188 |
-
with gr.Row():
|
189 |
-
gr.Textbox(
|
190 |
-
show_label=False, placeholder="Ist noch in Arbeit..."
|
191 |
-
).style(container=False)
|
192 |
-
with gr.TabItem("Code-Generierungen"):
|
193 |
-
with gr.Row():
|
194 |
-
gr.Textbox(
|
195 |
-
show_label=False, placeholder="Ist noch in Arbeit..."
|
196 |
-
).style(container=False)
|
197 |
-
|
198 |
-
predict_args = dict(
|
199 |
-
fn=predict,
|
200 |
-
inputs=[
|
201 |
-
user_question,
|
202 |
-
chatbotGr,
|
203 |
-
history,
|
204 |
-
top_p,
|
205 |
-
temperature,
|
206 |
-
max_length_tokens,
|
207 |
-
max_context_length_tokens,
|
208 |
-
],
|
209 |
-
outputs=[chatbotGr, history, status_display],
|
210 |
-
show_progress=True,
|
211 |
-
)
|
212 |
-
|
213 |
-
#neuer Chat
|
214 |
-
reset_args = dict(
|
215 |
-
#fn=reset_chat, inputs=[], outputs=[user_input, status_display]
|
216 |
-
fn=reset_textbox, inputs=[], outputs=[user_input, status_display]
|
217 |
-
)
|
218 |
-
|
219 |
-
# Chatbot
|
220 |
-
transfer_input_args = dict(
|
221 |
-
fn=transfer_input, inputs=[user_input], outputs=[user_question, user_input, submitBtn], show_progress=True
|
222 |
-
)
|
223 |
-
|
224 |
-
#Listener auf Start-Click auf Button oder Return
|
225 |
-
predict_event1 = user_input.submit(**transfer_input_args).then(**predict_args)
|
226 |
-
predict_event2 = submitBtn.click(**transfer_input_args).then(**predict_args)
|
227 |
-
|
228 |
-
#Listener, Wenn reset...
|
229 |
-
emptyBtn.click(
|
230 |
-
reset_state,
|
231 |
-
outputs=[chatbotGr, history, status_display],
|
232 |
-
show_progress=True,
|
233 |
-
)
|
234 |
-
emptyBtn.click(**reset_args)
|
235 |
|
236 |
-
demo.
|
237 |
-
#demo.queue(concurrency_count=1).launch(share=True)
|
238 |
-
demo.queue(concurrency_count=1).launch(debug=True)
|
|
|
6 |
import torch
|
7 |
from utils import *
|
8 |
from presets import *
|
9 |
+
from transformers import Trainer, TrainingArguments
|
10 |
|
11 |
|
12 |
######################################################################
|
|
|
20 |
tokenizer,model,device = load_tokenizer_and_model(base_model,False)
|
21 |
dataset_neu = daten_laden("alexkueck/tis")
|
22 |
|
23 |
+
#############################################
|
24 |
+
#Vorbereiten für das Training der neuen Daten
|
25 |
+
#############################################
|
|
|
|
|
|
|
26 |
|
27 |
#alles zusammen auf das neue datenset anwenden - batched = True und 4 Prozesse, um die Berechnung zu beschleunigen. Die "text" - Spalte braucht man anschließend nicht mehr, daher weglassen.
|
28 |
tokenized_datasets = dataset_neu.map(tokenize_function, batched=True, num_proc=4, remove_columns=["text"])
|
|
|
34 |
# block_size = tokenizer.model_max_length
|
35 |
block_size = 128
|
36 |
|
37 |
+
#nochmal die map-Funktion auf das bereits tokenisierte Datenset anwenden
|
38 |
+
#die bereits tokenisierten Datensatze ändern sich dadurch: die samples enthalten nun Mengen aus block_size Tokens
|
39 |
+
lm_datasets = tokenized_datasets.map(
|
40 |
+
group_texts,
|
41 |
+
batched=True,
|
42 |
+
batch_size=1000,
|
43 |
+
num_proc=4,
|
44 |
+
)
|
45 |
+
|
46 |
+
#die Daten wurden nun "gereinigt" und für das Model vorbereitet.
|
47 |
+
#z.B. anschauen mit: tokenizer.decode(lm_datasets["train"][1]["input_ids"])
|
48 |
+
|
49 |
+
####################################################
|
50 |
+
#Training
|
51 |
+
####################################################
|
52 |
+
|
53 |
+
#Training Args
|
54 |
+
model_name = base_model.split("/")[-1]
|
55 |
+
training_args = TrainingArguments(
|
56 |
+
f"{model_name}-finetuned-tis",
|
57 |
+
evaluation_strategy = "epoch",
|
58 |
+
learning_rate=2e-5,
|
59 |
+
weight_decay=0.01,
|
60 |
+
push_to_hub=True,
|
61 |
+
)
|
62 |
+
|
63 |
+
############################################
|
64 |
+
def trainieren_neu():
|
65 |
+
#Trainer zusammenstellen
|
66 |
+
trainer = Trainer(
|
67 |
+
model=model,
|
68 |
+
args=training_args,
|
69 |
+
train_dataset=lm_datasets["train"],
|
70 |
+
eval_dataset=lm_datasets["validation"],
|
71 |
+
)
|
72 |
|
73 |
+
#trainer ausführen
|
74 |
+
trainer.train()
|
75 |
|
76 |
+
#in den Hub laden
|
77 |
+
trainer.push_to_hub()
|
78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
|
80 |
|
81 |
+
#####################################################
|
82 |
+
#Hilfsfunktionen für das training
|
83 |
+
#####################################################
|
84 |
+
#Datensets in den Tokenizer schieben...
|
85 |
+
def tokenize_function(examples):
|
86 |
+
return tokenizer(examples["text"])
|
87 |
+
|
|
|
|
|
|
|
88 |
|
89 |
+
#Funktion, die den gegebenen Text aus dem Datenset gruppiert
|
90 |
+
def group_texts(examples):
|
91 |
+
# Concatenate all texts.
|
92 |
+
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
|
93 |
+
total_length = len(concatenated_examples[list(examples.keys())[0]])
|
94 |
+
# We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
|
95 |
+
# customize this part to your needs.
|
96 |
+
total_length = (total_length // block_size) * block_size
|
97 |
+
# Split by chunks of max_len.
|
98 |
+
result = {
|
99 |
+
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
|
100 |
+
for k, t in concatenated_examples.items()
|
101 |
+
}
|
102 |
+
result["labels"] = result["input_ids"].copy()
|
103 |
+
return result
|
104 |
+
|
105 |
+
|
106 |
+
|
107 |
|
108 |
#######################################################################
|
109 |
#Darstellung mit Gradio
|
110 |
|
111 |
+
with gr.Blocks() as demo:
|
112 |
+
output = gr.Textbox(label="Output Box")
|
113 |
+
start_btn = gr.Button("Start")
|
114 |
+
start_btn.click(fn=greet, inputs, outputs=output, api_name="trainieren_neu")
|
115 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
|
117 |
+
demo.launch()
|
|
|
|