Commit @fc1d4dbbd0dc65b2d522b17d81fd01455ad2cbf8

윤영준 2023-10-24

made ai analysis api and tested it. 25TPS using gunicorn app:app -w 12 -b 127.0.0.1:8080

@fc1d4dbbd0dc65b2d522b17d81fd01455ad2cbf8

a350b71

fc1d4db

action.py

--- action.py

+++ action.py


 from flask_restx import Resource, Namespace
-from flask import request
-from werkzeug.utils import secure_filename
+from flask import request, jsonify
 import os
+import json
 from database.database import DB
 import torch
 from torchvision.transforms import ToTensor

 from model.AttentiveRNN import AttentiveRNN
 from model.Classifier import Resnet as Classifier
 from subfuction.image_crop import crop_image
+import numpy as np
+import cv2
 
+db = DB()
 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
-paths = os.getcwd()
+# pre-loading models
+arnn = AttentiveRNN(6, 3, 2)
+arnn.eval()
+arnn.load_state_dict(torch.load("weights/ARNN_trained_weight_6_3_2.pt"))
+arnn.to(device=device)
+
+classifier = Classifier()
+classifier.eval()
+classifier.load_state_dict(torch.load("weights/Classifier_512.pt"))
+classifier.to(device=device)
+
+tf_toTensor = ToTensor()
+crop_size = (512, 512)
+start_point = (750, 450)
+root_dir = os.getcwd()
 
 Action = Namespace(
     name="Action",

 )
 
 
[email protected]('/image_summit')
-class fileUpload(Resource):
-    @Action.doc(responses={200: 'Success'})
-    @Action.doc(responses={500: 'Register Failed'})
-    def post(self):
-        if request.method == 'POST':
-            f = request.files['file']
-            f.save(secure_filename(f.filename))
-            return {
-                'save': 'done'  # str으로 반환하여 return
-            }, 200
+# @Action.route('/image_summit')
+# class fileUpload(Resource):
+#     @Action.doc(responses={200: 'Success'})
+#     @Action.doc(responses={500: 'Register Failed'})
+#     def post(self):
+#         if request.method == 'POST':
+#             f = request.files['file']
+#             f.save(secure_filename(f.filename))
+#             return {
+#                 'save': 'done'  # str으로 반환하여 return
+#             }, 200
 
 
 @Action.route('/image_anal')

     @Action.doc(responses={200: 'Success'})
     @Action.doc(responses={500: 'Register Failed'})
     def post(self):
-        if request.method == 'POST':
-            db = DB()
-            arnn = AttentiveRNN(6, 3, 2)
-            arnn.load_state_dict(torch.load("weights/ARNN_trained_weight_6_3_2.pt"))
-            arnn.to(device=device)
-            crop_size = (512, 512)
-            start_point = (750, 450)
-            tf_toTensor = ToTensor()
-            classifier = Classifier()
-            classifier.load_state_dict(torch.load("weights/Classifier_512.pt"))
-            classifier.to(device=device)
-            dir = os.getcwd()
-            lat = float(request.json['gps_x'])
-            lon = float(request.json['gps_y'])
-            filename = request.json['filename']
-            file_type = request.json['file_type']
-            total_path = dir + os.path.sep + filename + file_type
-            image = crop_image(total_path, crop_size, start_point)
-            if not image:
-                return {
-                    'node': (lat, lon),
-                    'rain': None,
-                }, 500
-            image_tensor = tf_toTensor(image)
-            image_tensor = image_tensor.unsqueeze(0)
-            image_tensor = image_tensor.to(device)
+        # Extracting JSON data
+        json_data = request.form.get('data')
+        if not json_data:
+            return jsonify({"message": "Missing JSON data"}), 400
+        data = json.loads(json_data)
+
+        lat = float(data['gps_x'])
+        lon = float(data['gps_y'])
+        filename = data['filename']
+        file_type = data['file_type']
+
+        uploaded_file = request.files.get('file')
+        file_bytes = np.asarray(bytearray(uploaded_file.read()), dtype=np.uint8)
+        image = cv2.imdecode(file_bytes, cv2.IMREAD_COLOR)
+
+        image = crop_image(image, crop_size, start_point)
+
+        image_tensor = tf_toTensor(image)
+        image_tensor = image_tensor.unsqueeze(0)
+        image_tensor = image_tensor.to(device)
+        with torch.no_grad():
             image_arnn = arnn(image_tensor)
+            image_tensor.cpu()
+            del image_tensor
             result = classifier(image_arnn['x'])
-            result = result.to("cpu")
-            _, predicted = torch.max(result.data, 1)
-            if predicted == 0:
-                rain = False
-            else:  # elif result == 1
-                rain = True
-            user_id = 'test'
-            action_success = True
-            action_id = 'test'
-            db.db_add_action(action_id, lat, lon, user_id, action_success)
-            return {
-                'node': (lat, lon),
-                'rain': rain,
-            }, 200
+            image_arnn['x'].cpu()
+            del image_arnn
+
+        result = result.to("cpu")
+        _, predicted = torch.max(result.data, 1)
+        del result
+        if predicted == 0:
+            rain = False
+        else:  # elif result == 1
+            rain = True
+        user_id = 'test'
+        action_success = True
+        action_id = 'test'
+        db.db_add_action(action_id, lat, lon, user_id, action_success)
+        return {
+            'node': (lat, lon),
+            'rain': rain,
+        }, 200
 
 
 @Action.route('/action_display')

a350b71

fc1d4db

app.py

--- app.py

+++ app.py


-from flask import Flask
+import flask
 from flask_restx import Api
 from auth import Auth
 from action import Action
 
-app = Flask(__name__)
+app = flask.Flask(__name__)
 
-
-print("Api Start")
 api = Api(    app,
     version='0.1',
     title="RDS",

     contact="[email protected]",
     license="MIT")
 
-
-
 api.add_namespace(Auth, '/auth')
-print("Api Add Auth")
 
 api.add_namespace(Action, '/action')
 

fc1d4db

asset/rain-svgrepo-com.png (Binary) (added)

+++ asset/rain-svgrepo-com.png

Binary file is not shown

fc1d4db

asset/sun-svgrepo-com.png (Binary) (added)

+++ asset/sun-svgrepo-com.png

Binary file is not shown

a350b71

fc1d4db

auth.py

--- auth.py

+++ auth.py


 import hashlib
-from flask import request, jsonify, render_template,redirect,url_for
+from flask import request, jsonify
 from flask_restx import Resource, Api, Namespace, fields
 from database.database import DB
 import datetime

fc1d4db

demonstration.py (added)

+++ demonstration.py

...	...	@@ -0,0 +1,107 @@
	1	+import random
	2	+import time
	3	+import visdom
	4	+import glob
	5	+import torch
	6	+import cv2
	7	+from torchvision.transforms import ToTensor, Compose, Normalize
	8	+from flask import request
	9	+from model.AttentiveRNN import AttentiveRNN
	10	+from model.Classifier import Resnet as Classifier
	11	+from subfuction.image_crop import crop_image
	12	+
	13	+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
	14	+
	15	+# execute visdom instance first
	16	+# to do that, install visdom via pip and execute in terminal
	17	+
	18	+def process_image():
	19	+ vis = visdom.Visdom()
	20	+ arnn = AttentiveRNN(6, 3, 2)
	21	+ arnn.load_state_dict(torch.load("weights/ARNN_trained_weight_6_3_2.pt"))
	22	+ arnn.to(device=device)
	23	+ arnn.eval()
	24	+ crop_size = (512, 512)
	25	+ start_point = (750, 450)
	26	+ tf_toTensor = ToTensor()
	27	+ classifier = Classifier(in_ch=1)
	28	+ classifier.load_state_dict(torch.load("weights/classifier_e19_weight_1080p_512512_fixed_wrong_resolution_and_ch.pt"))
	29	+ classifier.to(device=device)
	30	+ classifier.eval()
	31	+ rainy_data_path = glob.glob("/home/takensoft/Pictures/화창한날, 비오는날 프레임2000장/SUNNY///*.png")
	32	+ # rainy_data_path = glob.glob("/home/takensoft/Pictures/폭우 빗방울 (475개)/*.png")
	33	+ img_path = rainy_data_path
	34	+ # clean_data_path = glob.glob("/home/takensoft/Documents/AttentiveRNNClassifier/output/original/*.png")
	35	+
	36	+ # img_path = rainy_data_path + clean_data_path
	37	+ # normalize = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
	38	+ random.shuffle(img_path)
	39	+
	40	+ for i in iter(range(len(img_path))):
	41	+ image = crop_image(img_path[i], crop_size, start_point)
	42	+ if not image.any():
	43	+ continue
	44	+ image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
	45	+ image_tensor = tf_toTensor(image)
	46	+ image_tensor = image_tensor.unsqueeze(0)
	47	+ image_tensor = image_tensor.to(device)
	48	+ image_arnn = arnn(image_tensor)
	49	+
	50	+ input_win = 'input_window'
	51	+ attention_map_wins = [f'attention_map_{i}' for i in range(6)]
	52	+ prediction_win = 'prediction_window'
	53	+
	54	+ # Visualize attention maps using visdom
	55	+ vis.images(
	56	+ image_tensor,
	57	+ opts=dict(title=f"input"),
	58	+ win=input_win
	59	+ )
	60	+ for idx, attention_map in enumerate(image_arnn['attention_map_list']):
	61	+ if idx == 0 or idx == 5:
	62	+ vis.images(
	63	+ attention_map.cpu(), # Expected shape: (batch_size, C, H, W)
	64	+ opts=dict(title=f'Attention Map {idx + 1}'),
	65	+ win=attention_map_wins[idx]
	66	+ )
	67	+ # arnn_result = normalize(image_arnn['x'])
	68	+ result = classifier(image_arnn['attention_map_list'][-1])
	69	+ result = result.to("cpu")
	70	+ _, predicted = torch.max(result.data, 1)
	71	+ print(result.data)
	72	+ print(_)
	73	+ print(predicted)
	74	+ # Load and display the corresponding icon
	75	+ if predicted == 0:
	76	+ icon_path = 'asset/sun-svgrepo-com.png'
	77	+ else: # elif result == 1
	78	+ icon_path = 'asset/rain-svgrepo-com.png'
	79	+
	80	+ # Load icon and convert to tensor
	81	+ icon_image = cv2.imread(icon_path, cv2.IMREAD_UNCHANGED)
	82	+ transform = Compose([
	83	+ ToTensor()
	84	+ ])
	85	+ icon_tensor = transform(icon_image).unsqueeze(0) # Add batch dimension
	86	+
	87	+ # Visualize icon using visdom
	88	+ vis.images(
	89	+ icon_tensor,
	90	+ opts=dict(title='Weather Prediction'),
	91	+ win=prediction_win
	92	+ )
	93	+ time.sleep(1)
	94	+
	95	+ # result = classifier(image_arnn['x'])
	96	+ # result = result.to("cpu")
	97	+ # _, predicted = torch.max(result.data, 1)
	98	+ # if predicted == 0:
	99	+ # rain = False
	100	+ # else: # elif result == 1
	101	+ # rain = True
	102	+ # return {
	103	+ # 'rain': rain,
	104	+ # }, 200
	105	+
	106	+if __name__ == "__main__":
	107	+ process_image()(파일 끝에 줄바꿈 문자 없음)

a350b71

fc1d4db

model/AttentiveRNN.py

--- model/AttentiveRNN.py

+++ model/AttentiveRNN.py


             lstm_feats.append(lstm_feats_i)
         ret = {
             'x' : x,
-            'attention_map_list' : attention_map,
-            'lstm_feats' : lstm_feats
+            # 'attention_map_list' : attention_map,
+            # 'lstm_feats' : lstm_feats
         }
         return ret
 

a350b71

requirements.txt (deleted)

--- requirements.txt

...	...	@@ -1,11 +0,0 @@
	1	-torch~=2.0.1+cu118
	2	-flask~=2.3.3
	3	-jwt~=1.3.1
	4	-werkzeug~=2.3.7
	5	-torchvision~=0.15.2+cu118
	6	-networkx~=3.0
	7	-geojson~=3.0.1
	8	-haversine~=2.8.0
	9	-opencv-python~=4.8.0.76
	10	-joblib~=1.3.2
	11	-pandas~=2.0.3(파일 끝에 줄바꿈 문자 없음)

a350b71

fc1d4db

subfuction/image_crop.py

--- subfuction/image_crop.py

+++ subfuction/image_crop.py


 import cv2
 import os
 
-def crop_image(image_path, crop_size, start_point):
-    if image_path.endswith(".jpg") or image_path.endswith(".png"):
-        image = cv2.imread(image_path)
-        height, width = image.shape[:2]
+def crop_image(image, crop_size, start_point):
+    height, width = image.shape[:2]
 
-        if width > start_point[0] + crop_size[0] and height > start_point[1] + crop_size[1]:
-            cropped_image = image[start_point[1]:start_point[1]+crop_size[1], start_point[0]:start_point[0]+crop_size[0]]
-            return cropped_image
-        else:
-            print(f"Image {os.path.basename(image_path)} is too small to be cropped with the current settings.")
-            return False
+    if width > start_point[0] + crop_size[0] and height > start_point[1] + crop_size[1]:
+        cropped_image = image[start_point[1]:start_point[1]+crop_size[1], start_point[0]:start_point[0]+crop_size[0]]
+        return cropped_image
+    else:
+        print(f"Image is too small to be cropped with the current settings.")
+        return False

Add a comment

Open 0
Closed 0

List

...	...	@@ -217,8 +217,8 @@
217	217	lstm_feats.append(lstm_feats_i)
218	218	ret = {
219	219	'x' : x,
220		- 'attention_map_list' : attention_map,
221		- 'lstm_feats' : lstm_feats
	220	+ # 'attention_map_list' : attention_map,
	221	+ # 'lstm_feats' : lstm_feats
222	222	}
223	223	return ret
224	224

...	...	@@ -1,7 +1,7 @@
1	1	from flask_restx import Resource, Namespace
2		-from flask import request
3		-from werkzeug.utils import secure_filename
	2	+from flask import request, jsonify
4	3	import os
	4	+import json
5	5	from database.database import DB
6	6	import torch
7	7	from torchvision.transforms import ToTensor
...	...	@@ -9,10 +9,27 @@
9	9	from model.AttentiveRNN import AttentiveRNN
10	10	from model.Classifier import Resnet as Classifier
11	11	from subfuction.image_crop import crop_image
	12	+import numpy as np
	13	+import cv2
12	14
	15	+db = DB()
13	16	device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
14	17
15		-paths = os.getcwd()
	18	+# pre-loading models
	19	+arnn = AttentiveRNN(6, 3, 2)
	20	+arnn.eval()
	21	+arnn.load_state_dict(torch.load("weights/ARNN_trained_weight_6_3_2.pt"))
	22	+arnn.to(device=device)
	23	+
	24	+classifier = Classifier()
	25	+classifier.eval()
	26	+classifier.load_state_dict(torch.load("weights/Classifier_512.pt"))
	27	+classifier.to(device=device)
	28	+
	29	+tf_toTensor = ToTensor()
	30	+crop_size = (512, 512)
	31	+start_point = (750, 450)
	32	+root_dir = os.getcwd()
16	33
17	34	Action = Namespace(
18	35	name="Action",
...	...	@@ -20,17 +37,17 @@
20	37	)
21	38
22	39
23		[email protected]('/image_summit')
24		-class fileUpload(Resource):
25		- @Action.doc(responses={200: 'Success'})
26		- @Action.doc(responses={500: 'Register Failed'})
27		- def post(self):
28		- if request.method == 'POST':
29		- f = request.files['file']
30		- f.save(secure_filename(f.filename))
31		- return {
32		- 'save': 'done' # str으로 반환하여 return
33		- }, 200
	40	+# @Action.route('/image_summit')
	41	+# class fileUpload(Resource):
	42	+# @Action.doc(responses={200: 'Success'})
	43	+# @Action.doc(responses={500: 'Register Failed'})
	44	+# def post(self):
	45	+# if request.method == 'POST':
	46	+# f = request.files['file']
	47	+# f.save(secure_filename(f.filename))
	48	+# return {
	49	+# 'save': 'done' # str으로 반환하여 return
	50	+# }, 200
34	51
35	52
36	53	@Action.route('/image_anal')
...	...	@@ -38,48 +55,49 @@
38	55	@Action.doc(responses={200: 'Success'})
39	56	@Action.doc(responses={500: 'Register Failed'})
40	57	def post(self):
41		- if request.method == 'POST':
42		- db = DB()
43		- arnn = AttentiveRNN(6, 3, 2)
44		- arnn.load_state_dict(torch.load("weights/ARNN_trained_weight_6_3_2.pt"))
45		- arnn.to(device=device)
46		- crop_size = (512, 512)
47		- start_point = (750, 450)
48		- tf_toTensor = ToTensor()
49		- classifier = Classifier()
50		- classifier.load_state_dict(torch.load("weights/Classifier_512.pt"))
51		- classifier.to(device=device)
52		- dir = os.getcwd()
53		- lat = float(request.json['gps_x'])
54		- lon = float(request.json['gps_y'])
55		- filename = request.json['filename']
56		- file_type = request.json['file_type']
57		- total_path = dir + os.path.sep + filename + file_type
58		- image = crop_image(total_path, crop_size, start_point)
59		- if not image:
60		- return {
61		- 'node': (lat, lon),
62		- 'rain': None,
63		- }, 500
64		- image_tensor = tf_toTensor(image)
65		- image_tensor = image_tensor.unsqueeze(0)
66		- image_tensor = image_tensor.to(device)
	58	+ # Extracting JSON data
	59	+ json_data = request.form.get('data')
	60	+ if not json_data:
	61	+ return jsonify({"message": "Missing JSON data"}), 400
	62	+ data = json.loads(json_data)
	63	+
	64	+ lat = float(data['gps_x'])
	65	+ lon = float(data['gps_y'])
	66	+ filename = data['filename']
	67	+ file_type = data['file_type']
	68	+
	69	+ uploaded_file = request.files.get('file')
	70	+ file_bytes = np.asarray(bytearray(uploaded_file.read()), dtype=np.uint8)
	71	+ image = cv2.imdecode(file_bytes, cv2.IMREAD_COLOR)
	72	+
	73	+ image = crop_image(image, crop_size, start_point)
	74	+
	75	+ image_tensor = tf_toTensor(image)
	76	+ image_tensor = image_tensor.unsqueeze(0)
	77	+ image_tensor = image_tensor.to(device)
	78	+ with torch.no_grad():
67	79	image_arnn = arnn(image_tensor)
	80	+ image_tensor.cpu()
	81	+ del image_tensor
68	82	result = classifier(image_arnn['x'])
69		- result = result.to("cpu")
70		- _, predicted = torch.max(result.data, 1)
71		- if predicted == 0:
72		- rain = False
73		- else: # elif result == 1
74		- rain = True
75		- user_id = 'test'
76		- action_success = True
77		- action_id = 'test'
78		- db.db_add_action(action_id, lat, lon, user_id, action_success)
79		- return {
80		- 'node': (lat, lon),
81		- 'rain': rain,
82		- }, 200
	83	+ image_arnn['x'].cpu()
	84	+ del image_arnn
	85	+
	86	+ result = result.to("cpu")
	87	+ _, predicted = torch.max(result.data, 1)
	88	+ del result
	89	+ if predicted == 0:
	90	+ rain = False
	91	+ else: # elif result == 1
	92	+ rain = True
	93	+ user_id = 'test'
	94	+ action_success = True
	95	+ action_id = 'test'
	96	+ db.db_add_action(action_id, lat, lon, user_id, action_success)
	97	+ return {
	98	+ 'node': (lat, lon),
	99	+ 'rain': rain,
	100	+ }, 200
83	101
84	102
85	103	@Action.route('/action_display')

...	...	@@ -1,12 +1,10 @@
1		-from flask import Flask
	1	+import flask
2	2	from flask_restx import Api
3	3	from auth import Auth
4	4	from action import Action
5	5
6		-app = Flask(__name__)
	6	+app = flask.Flask(__name__)
7	7
8		-
9		-print("Api Start")
10	8	api = Api( app,
11	9	version='0.1',
12	10	title="RDS",
...	...	@@ -15,10 +13,7 @@
15	13	contact="[email protected]",
16	14	license="MIT")
17	15
18		-
19		-
20	16	api.add_namespace(Auth, '/auth')
21		-print("Api Add Auth")
22	17
23	18	api.add_namespace(Action, '/action')
24	19

...	...	@@ -1,5 +1,5 @@
1	1	import hashlib
2		-from flask import request, jsonify, render_template,redirect,url_for
	2	+from flask import request, jsonify
3	3	from flask_restx import Resource, Api, Namespace, fields
4	4	from database.database import DB
5	5	import datetime

...	...	@@ -1,14 +1,12 @@
1	1	import cv2
2	2	import os
3	3
4		-def crop_image(image_path, crop_size, start_point):
5		- if image_path.endswith(".jpg") or image_path.endswith(".png"):
6		- image = cv2.imread(image_path)
7		- height, width = image.shape[:2]
	4	+def crop_image(image, crop_size, start_point):
	5	+ height, width = image.shape[:2]
8	6
9		- if width > start_point[0] + crop_size[0] and height > start_point[1] + crop_size[1]:
10		- cropped_image = image[start_point[1]:start_point[1]+crop_size[1], start_point[0]:start_point[0]+crop_size[0]]
11		- return cropped_image
12		- else:
13		- print(f"Image {os.path.basename(image_path)} is too small to be cropped with the current settings.")
14		- return False (파일 끝에 줄바꿈 문자 없음)
	7	+ if width > start_point[0] + crop_size[0] and height > start_point[1] + crop_size[1]:
	8	+ cropped_image = image[start_point[1]:start_point[1]+crop_size[1], start_point[0]:start_point[0]+crop_size[0]]
	9	+ return cropped_image
	10	+ else:
	11	+ print(f"Image is too small to be cropped with the current settings.")
	12	+ return False (파일 끝에 줄바꿈 문자 없음)

Delete comment