File size: 10,575 Bytes
77df4ca a5a59c1 77df4ca a5a59c1 77df4ca a5a59c1 5ea4cc2 77df4ca a5a59c1 5ea4cc2 77df4ca a5a59c1 77df4ca a5a59c1 77df4ca 68f056c 77df4ca 68f056c 77df4ca 68f056c 77df4ca 68f056c 77df4ca 68f056c 77df4ca a5a59c1 77df4ca a5a59c1 77df4ca a5a59c1 de92875 77df4ca 5ea4cc2 68f056c 77df4ca 5ea4cc2 77df4ca de92875 77df4ca 1c7cd02 77df4ca 1c7cd02 77df4ca a5a59c1 68f056c 77df4ca a5a59c1 77df4ca a5a59c1 de92875 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 |
import time
import io
import gradio as gr
import cv2
import base64
import openai
from langchain.prompts import PromptTemplate
from langchain.chat_models import ChatOpenAI
from langchain.schema import StrOutputParser
from PIL import Image
global_dict = {}
def failure():
raise gr.Error("This should fail!")
def validate_api_key(api_key):
client = openai.OpenAI(api_key=api_key)
try:
# Make your OpenAI API request here
response = client.completions.create(
prompt="Hello world",
model="gpt-3.5-turbo-instruct"
)
except openai.RateLimitError as e:
# Handle rate limit error (we recommend using exponential backoff)
print(f"OpenAI API request exceeded rate limit: {e}")
response = None
pass
except openai.APIConnectionError as e:
# Handle connection error here
print(f"Failed to connect to OpenAI API: {e}")
response = None
pass
except openai.APIError as e:
# Handle API error here, e.g. retry or log
print(f"OpenAI API returned an API Error: {e}")
response = None
pass
if response:
return True
else:
raise gr.Error(f"OpenAI API returned an API Error")
def _process_video(image_file):
# Read and process the video file
video = cv2.VideoCapture(image_file.name)
base64Frames = []
while video.isOpened():
success, frame = video.read()
if not success:
break
_, buffer = cv2.imencode(".jpg", frame)
base64Frames.append(base64.b64encode(buffer).decode("utf-8"))
video.release()
if len(base64Frames) > 700:
raise gr.Warning(f"Video's play time is too long. (>20s)")
print(len(base64Frames), "frames read.")
if not base64Frames:
raise gr.Error(f"Cannot open the video.")
return base64Frames
def _make_video_batch(image_file, batch_size, total_batch_percent):
frames = _process_video(image_file)
TOTAL_FRAME_COUNT = len(frames)
BATCH_SIZE = int(batch_size)
TOTAL_BATCH_SIZE = int(TOTAL_FRAME_COUNT * total_batch_percent / 100)
BATCH_STEP = int(TOTAL_FRAME_COUNT / TOTAL_BATCH_SIZE)
base64FramesBatch = []
for idx in range(0, TOTAL_FRAME_COUNT, BATCH_STEP * BATCH_SIZE):
# print(f'## {idx}')
temp = []
for i in range(BATCH_SIZE):
# print(f'# {idx + BATCH_STEP * i}')
if (idx + BATCH_STEP * i) < TOTAL_FRAME_COUNT:
temp.append(frames[idx + BATCH_STEP * i])
else:
continue
base64FramesBatch.append(temp)
for idx, batch in enumerate(base64FramesBatch):
# assert len(batch) <= BATCH_SIZE
print(f'##{idx} - batch_size: {len(batch)}')
global_dict.setdefault('batched_frames', base64FramesBatch)
return base64FramesBatch
def show_batches(image_file, batch_size, total_batch_percent):
batched_frames = _make_video_batch(image_file, batch_size, total_batch_percent)
images = []
for i, l in enumerate(batched_frames):
print(f"#### Batch_{i+1}")
for j, img in enumerate(l):
print(f'## Image_{j+1}')
image_bytes = base64.b64decode(img.encode("utf-8"))
# Convert the bytes to a stream (file-like object)
image_stream = io.BytesIO(image_bytes)
# Open the image as a PIL image
image = Image.open(image_stream)
images.append((image, f"batch {i+1}"))
print("-"*100)
return images
def call_gpt_vision(api_key, instruction):
frames = global_dict.get('batched_frames')
openai.api_key = api_key
full_result = []
for idx, batch in enumerate(frames):
PROMPT_MESSAGES = [
{
"role": "system",
"content": "You will evaluate the behavior of the person in the sequences of images. They show discrete parts of the whole continuous behavior. You should only evaluate the parts you can rate based on the given images. Remember, you're evaluating the given parts to evaluate the whole continuous behavior, and you'll connect them later to evaluate the whole. Never add your own judgment. Evlaute only in the contents of images themselves. If you can't evaluate it, just answer '(Unevaluable)'"
},
{
"role": "user",
"content": [
"Evaluate the behavior's actions based on the <CRITERIA> provided.\n\n" + instruction,
*map(lambda x: {"image": x, "resize": 300}, batch),
],
},
]
params = {
"model": "gpt-4-vision-preview",
"messages": PROMPT_MESSAGES,
"max_tokens": 1024,
}
try:
result = openai.chat.completions.create(**params)
print(result.choices[0].message.content)
full_result.append(result)
except Exception as e:
print(f"Error: {e}")
yield f'### BATCH_{idx+1}\n' + "-"*50 + "\n" + f"Error: {e}" + "\n" + "-"*50
if 'full_result' not in global_dict:
global_dict.setdefault('full_result', full_result)
else:
global_dict['full_result'] = full_result
print(f'### BATCH_{idx+1}')
print('-'*100)
time.sleep(2)
yield f'### BATCH_{idx+1}\n' + "-"*50 + "\n" + result.choices[0].message.content + "\n" + "-"*50
def get_full_result():
full_result = global_dict.get('full_result')
result_text = ""
for idx, res in enumerate(full_result):
result_text += f'<Evaluation_{idx+1}>\n'
result_text += res.choices[0].message.content
result_text += "\n"
result_text += "-"*5
result_text += "\n"
global_dict.setdefault('result_text', result_text)
return result_text
def get_final_anser(api_key, result_text):
chain = ChatOpenAI(model="gpt-4", max_tokens=1024, temperature=0, api_key=api_key)
prompt = PromptTemplate.from_template(
"""
You see the following list of texts that evaluate forward roll:
{evals}
Write an full text that synthesizes and summarizes the contents of all the text above.
Each evaluates a specific part, and you should combine them based on what was evaluated in each part.
The way to combine them is 'or', not 'and', which means you only need to evaluate the parts of a post that are rated based on that.
Concatenate based on what was evaluated, if anything.
Example:
an overview of evaluations
1. Specific assessments for each item
2.
3.
....
Overall opinion
Total score : 1~10 / 10
Output:
"""
)
runnable = prompt | chain | StrOutputParser()
final_eval = runnable.invoke({"evals": result_text})
return final_eval
# Define the Gradio app
def main():
with gr.Blocks() as demo:
gr.Markdown("# GPT-4 Vision for Evaluation")
gr.Markdown("## 1st STEP. Make Batched Snapshots")
with gr.Row():
with gr.Column(scale=1):
api_key_input = gr.Textbox(
label="Enter your OpenAI API Key",
info="Your API Key must be allowed to use GPT-4 Vision",
placeholder="sk-*********...",
lines=1
)
video_upload = gr.File(
label="Upload your video (under 10 second video is the best..!)",
file_types=["video"],
)
batch_size = gr.Number(
label="Number of images in one batch",
value=2,
minimum=2,
maximum=5
)
total_batch_percent = gr.Number(
label="Percentage(%) of batched image frames to total frames",
value=5,
minimum=5,
maximum=20,
step=5
)
process_button = gr.Button("Process")
with gr.Column(scale=1):
gallery = gr.Gallery(
label="Batched Snapshots of Video",
columns=[5],
rows=[1],
object_fit="contain",
height="auto"
)
gr.Markdown("## 2nd STEP. Set Evaluation Criteria")
with gr.Row():
with gr.Column(scale=1):
instruction_input = gr.Textbox(
label="Evaluation Criteria",
info="Enter your evaluation criteria here...",
placeholder="<CRITERIA>\nThe correct way to do a forward roll is as follows:\n1. From standing, bend your knees and straighten your arms in front of you.\n2. Place your hands on the floor, shoulder width apart with fingers pointing forward and your chin on your chest.\n3. Rock forward, straighten legs and transfer body weight onto shoulders.\n4. Rock forward on a rounded back placing both feet on the floor.\n5. Stand using arms for balance, without hands touching the floor.",
lines=7)
submit_button = gr.Button("Evaluate")
with gr.Column(scale=1):
output_box = gr.Textbox(
label="Batched Generated Response...(Streaming)",
lines=10,
interactive=False
)
gr.Markdown("## 3rd STEP. Summarize and Get Result")
with gr.Row():
with gr.Column(scale=1):
output_box_fin = gr.Textbox(
label="FULL Response",
info="You can edit partial evaluation in here...",
lines=10,
interactive=True)
submit_button_2 = gr.Button("Summarize")
with gr.Column(scale=1):
output_box_fin_fin = gr.Textbox(label="FINAL EVALUATION", lines=10, interactive=True)
process_button.click(fn=validate_api_key, inputs=api_key_input, outputs=None).success(fn=show_batches, inputs=[video_upload, batch_size, total_batch_percent], outputs=gallery)
submit_button.click(fn=call_gpt_vision, inputs=[api_key_input, instruction_input], outputs=output_box).then(get_full_result, None, output_box_fin)
submit_button_2.click(fn=get_final_anser, inputs=[api_key_input, output_box_fin], outputs=output_box_fin_fin)
demo.launch()
if __name__ == "__main__":
main() |