swastikbansal commited on
Commit
fab7d4f
·
verified ·
1 Parent(s): c7265b1

Added CustomTrainer module

Browse files
Files changed (1) hide show
  1. CustomTrainer.py +288 -0
CustomTrainer.py ADDED
@@ -0,0 +1,288 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import json
3
+ import cv2
4
+ import pickle
5
+ from sklearn.ensemble import RandomForestClassifier
6
+ from sklearn.model_selection import train_test_split
7
+ import os
8
+ from supabase import create_client, Client
9
+ from dotenv import load_dotenv
10
+ from utils import Utils
11
+
12
+ class Trainer:
13
+ def __init__(self,hands,pose):
14
+ load_dotenv()
15
+ self.utils = Utils()
16
+
17
+ self.data: list = []
18
+ self.labels: list = []
19
+
20
+ os.makedirs('./Custom_Dataset/MP_DATA', exist_ok=True)
21
+ os.makedirs('./Custom_Dataset/Models', exist_ok=True)
22
+ os.makedirs('./Custom_Dataset/Data', exist_ok=True)
23
+
24
+ # Base paths
25
+ self.ROOT_FOLDER: str = "./Custom_Dataset"
26
+ self.DATA_FOLDER: str = f"{self.ROOT_FOLDER}/Data"
27
+ self.JSON_PATH: str = f'{self.ROOT_FOLDER}/MP_DATA'
28
+ self.MODEL_PATH: str = f'{self.ROOT_FOLDER}/Models'
29
+
30
+ # Org Json paths
31
+ self.LEFT_JSON: str = r'./MP_Data/left.json'
32
+ self.RIGHT_JSON: str = r'./MP_Data/right.json'
33
+ self.POSE_JSON: str = r'./MP_Data/pose.json'
34
+
35
+ # New JSON PATHS
36
+ self.NEW_LEFT_JSON: str = rf'{self.JSON_PATH}/new_left.json'
37
+ self.NEW_RIGHT_JSON: str = rf'{self.JSON_PATH}/new_right.json'
38
+ self.NEW_POSE_JSON: str = rf'{self.JSON_PATH}/new_pose.json'
39
+
40
+ # Mediapipe
41
+ self.hands = hands
42
+ self.pose = pose
43
+
44
+ #Supabase
45
+ self.SUPABASE_URL: str = os.environ.get("SUPABASE_URL")
46
+ self.SUPABASE_KEY: str = os.environ.get("SUPABASE_KEY")
47
+ self.supabase: Client = create_client(self.SUPABASE_URL, self.SUPABASE_KEY)
48
+ self.SUPABASE_BUCKET:str = "Custom_Dataset"
49
+
50
+ def download_files(self, path: str = ""):
51
+ """Recursively list all files in a Supabase storage bucket."""
52
+ items = self.supabase.storage.from_(self.SUPABASE_BUCKET).list(path)
53
+ all_files = []
54
+
55
+ for item in items:
56
+ # Skipping models folder
57
+ if item["name"].lower() == "models":
58
+ continue
59
+
60
+ # Folder
61
+ if item["metadata"] is None:
62
+ sub_path = f"{path}/{item['name']}" if path else item["name"]
63
+
64
+ all_files.extend(self.download_files(sub_path))
65
+
66
+ # File
67
+ else:
68
+ file_path = f"{path}/{item['name']}" if path else item["name"]
69
+ down_path = f"./{self.DATA_FOLDER}/{file_path}"
70
+ os.makedirs(os.path.dirname(down_path), exist_ok=True)
71
+ with open(rf"{down_path}", "wb+") as f:
72
+ response = (
73
+ self.supabase.storage
74
+ .from_(self.SUPABASE_BUCKET)
75
+ .download(rf"{file_path}")
76
+ )
77
+ f.write(response)
78
+
79
+ all_files.append(file_path)
80
+
81
+ return all_files
82
+
83
+ def download_images(self):
84
+ # Request files from the root of the bucket (empty path)
85
+ files = self.download_files("")
86
+ print(f"Downloaded files: {files}")
87
+ if not files:
88
+ print("No files downloaded.")
89
+ else:
90
+ print(f"Downloaded images: {files}")
91
+ return files
92
+
93
+ def process_hand(self, hand_label:str) -> tuple:
94
+ """Iterate through each class folder in the data directory"""
95
+ # Reset data and labels for each hand type
96
+ data = []
97
+ labels = []
98
+
99
+ for label in os.listdir(self.DATA_FOLDER):
100
+ class_folder = os.path.join(self.DATA_FOLDER, label)
101
+
102
+ if os.path.isdir(class_folder):
103
+ # Iterate through each image in the class folder
104
+ for image_name in os.listdir(class_folder):
105
+ image_path = os.path.join(class_folder, image_name)
106
+
107
+ image = cv2.imread(image_path)
108
+ if image is None:
109
+ print(f"Warning: Could not read image {image_path}")
110
+ continue
111
+
112
+ image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
113
+ results = self.hands.process(image_rgb)
114
+
115
+ pose_results = self.pose.process(image_rgb)
116
+
117
+ # If hands are detected, extract features for the specified hand
118
+ if results.multi_hand_landmarks:
119
+ for hand_landmarks, handedness in zip(results.multi_hand_landmarks, results.multi_handedness):
120
+ if handedness.classification[0].label == hand_label:
121
+ if pose_results.pose_landmarks:
122
+ features = self.utils.extract_hand_features(hand_landmarks.landmark, pose_results.pose_landmarks.landmark)
123
+ data.append(features)
124
+ labels.append(label)
125
+
126
+ return data, labels
127
+
128
+ def process_pose(self):
129
+ """
130
+ Processes images in the data folder to extract pose features and labels.
131
+ """
132
+ data = []
133
+ labels = []
134
+
135
+ for label in os.listdir(self.DATA_FOLDER):
136
+ class_folder = os.path.join(self.DATA_FOLDER, label)
137
+
138
+ if os.path.isdir(class_folder):
139
+ for image_name in os.listdir(class_folder):
140
+ image_path = os.path.join(class_folder, image_name)
141
+
142
+ image = cv2.imread(image_path)
143
+ image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
144
+ results = self.pose.process(image_rgb)
145
+
146
+ if results.pose_landmarks:
147
+ landmarks = results.pose_landmarks.landmark
148
+ features = self.utils.extract_pose_features(landmarks)
149
+ data.append(features)
150
+ labels.append(label)
151
+
152
+ return data, labels
153
+
154
+ def load_json(self,json_path):
155
+ with open(json_path, 'r') as f:
156
+ data_dict = json.load(f)
157
+ X = np.array(data_dict['data'])
158
+ y = np.array(data_dict['labels'])
159
+ return X, y
160
+
161
+ def merge_json(self,new_data, new_labels, json_path, new_path):
162
+ merged_data = new_data
163
+ merged_labels = new_labels
164
+
165
+ if os.path.exists(json_path):
166
+ with open(json_path, 'r') as f:
167
+ old = json.load(f)
168
+ merged_data = old.get('data', []) + new_data
169
+ merged_labels = old.get('labels', []) + new_labels
170
+
171
+ with open(new_path, 'w') as f:
172
+ json.dump({'data': merged_data, 'labels': merged_labels}, f)
173
+ print(f"JSON file merged and saved at {new_path}")
174
+
175
+ def train_model(self,new_path, model_path):
176
+ os.makedirs(os.path.dirname(model_path), exist_ok=True)
177
+ X, y = self.load_json(new_path)
178
+
179
+ if len(X) == 0:
180
+ raise ValueError(f"No data found for training model at {new_path}")
181
+
182
+ # Check if we have enough samples for stratification
183
+ unique_labels = np.unique(y)
184
+ if len(unique_labels) < 2:
185
+ raise ValueError(f"Need at least 2 classes for training, found {len(unique_labels)}")
186
+
187
+ # Check if each class has at least 2 samples for stratification
188
+ label_counts = np.bincount(np.array([np.where(unique_labels == label)[0][0] for label in y]))
189
+ if np.any(label_counts < 2):
190
+ print("Warning: Some classes have less than 2 samples, using shuffle without stratify")
191
+ X_train, X_test, y_train, y_test = train_test_split(
192
+ X, y, test_size=0.2, shuffle=True
193
+ )
194
+ else:
195
+ X_train, X_test, y_train, y_test = train_test_split(
196
+ X, y, test_size=0.2, shuffle=True, stratify=y
197
+ )
198
+
199
+ model = RandomForestClassifier(
200
+ n_estimators=100,
201
+ max_depth=6,
202
+ min_samples_split=3,
203
+ min_samples_leaf=2,
204
+ bootstrap=True,
205
+ criterion='entropy',
206
+ oob_score=True
207
+ )
208
+
209
+ model.fit(X_train, y_train)
210
+
211
+ y_pred = model.predict(X_test)
212
+ acc = np.mean(y_pred == y_test)
213
+ print(f"Accuracy for {model_path}: {acc *100 :.2f}")
214
+
215
+ with open(model_path, 'wb') as f:
216
+ pickle.dump({'model': model}, f)
217
+
218
+ def compute_accuracy(self,y_true, y_pred):
219
+ correct_predictions = sum(1 for true_label, predicted in zip(y_true, y_pred) if true_label == predicted)
220
+ return correct_predictions / len(y_true)
221
+
222
+ def upload_model(self,newModel_file):
223
+ with open(newModel_file, "rb") as f:
224
+ self.response = (
225
+ self.supabase.storage
226
+ .from_(self.SUPABASE_BUCKET)
227
+ .upload(
228
+ file=f,
229
+ path=f"models/{os.path.basename(newModel_file)}",
230
+ file_options={"cache-control": "3600", "upsert": "true"}
231
+ )
232
+ )
233
+ print(f"Uploaded model to Supabase: {newModel_file}")
234
+
235
+ def delete_images(self, files):
236
+ """
237
+ Delete local copies and remove files from Supabase bucket.
238
+ 'files' is expected to be a list of file paths relative to the bucket root
239
+ (the same strings produced by download_files).
240
+ """
241
+ for file in files:
242
+ local_path = os.path.normpath(os.path.join(self.DATA_FOLDER, file))
243
+ try:
244
+ if os.path.exists(local_path):
245
+ os.remove(local_path)
246
+ else:
247
+ print(f"Local file not found (skipping): {local_path}")
248
+ except Exception as e:
249
+ print(f"Failed to delete local file {local_path}: {e}")
250
+
251
+ try:
252
+ self.supabase.storage.from_(self.SUPABASE_BUCKET).remove([file])
253
+ except Exception as e:
254
+ print(f"Failed to delete remote file {file} from Supabase: {e}")
255
+
256
+ print("Deleting images")
257
+
258
+ def run_training(self) -> bool:
259
+ try:
260
+ files = self.download_images()
261
+
262
+ # Extracting data
263
+ left_data, left_labels = self.process_hand('Left')
264
+ right_data, right_labels = self.process_hand('Right')
265
+ pose_data, pose_labels = self.process_pose()
266
+
267
+ # Save the data and labels in JSON file
268
+ self.merge_json(left_data, left_labels, self.LEFT_JSON, self.NEW_LEFT_JSON)
269
+ self.merge_json(right_data, right_labels, self.RIGHT_JSON, self.NEW_RIGHT_JSON)
270
+ self.merge_json(pose_data, pose_labels, self.POSE_JSON, self.NEW_POSE_JSON)
271
+
272
+ # Train and save all three models using JSON data files
273
+ self.train_model(self.NEW_POSE_JSON, f'{self.MODEL_PATH}/pose_model_new.p')
274
+ self.train_model(self.NEW_LEFT_JSON, f'{self.MODEL_PATH}/left_model_new.p')
275
+ self.train_model(self.NEW_RIGHT_JSON, f'{self.MODEL_PATH}/right_model_new.p')
276
+
277
+ #Upload Model
278
+ self.upload_model(f'{self.MODEL_PATH}/pose_model_new.p')
279
+ self.upload_model(f'{self.MODEL_PATH}/left_model_new.p')
280
+ self.upload_model(f'{self.MODEL_PATH}/right_model_new.p')
281
+
282
+ self.delete_images(files)
283
+
284
+ return True
285
+
286
+ except Exception as e:
287
+ print(f"Training failed: {e}")
288
+ return False