upoad load_data.py to replace in Flair#2 Git to use HF dataset

load_data.py

@@ -0,0 +1,111 @@
+def load_data (config: dict): 
+    
+    def get_data_paths(config: dict, path_domains: str, paths_data: dict, matching_dict: dict) -> dict: 
+        
+        #### return data paths 
+        def list_items(path, filter): 
+            for path in Path(path).rglob(filter): 
+                yield path.resolve().as_posix() 
+
+        ## data paths dict
+        data = {'PATH_IMG':[], 'PATH_SP_DATA':[], 'SP_COORDS':[], 'PATH_SP_DATES':[],  'PATH_SP_MASKS':[], 'PATH_LABELS':[], 'MTD_AERIAL':[]} 
+        for domain in path_domains: 
+            aerial = sorted(list(list_items(Path(path_data)/domain, 'IMG*.tif')), key=lambda x: int(x.split('_')[-1][:-4])) 
+            sen2sp = sorted(list(list_items(Path(path_data)/domain, '*data.npy'))) 
+            sprods = sorted(list(list_items(Path(path_data)/domain, '*products.txt')))
+            smasks = sorted(list(list_items(Path(path_data)/domain, '*masks.npy')))
+            labels = sorted(list(list_items(Path(path_data)/domain, 'MSK*.tif')), key=lambda x: int(x.split('_')[-1][:-4])) 
+            coords = [] 
+            for k in aerial: 
+                coords.append(matching_dict[k.split('/')[-1]]) 
+            
+            data['PATH_IMG'] += aerial 
+            data['PATH_SP_DATA'] += sen2sp*len(aerial) 
+            data['PATH_SP_DATES'] += sprods*len(aerial)
+            data['PATH_SP_MASKS'] += smasks*len(aerial) 
+            data['SP_COORDS'] += coords 
+            data['PATH_LABELS'] += labels
+ 
+        if config['aerial_metadata'] == True:
+            data = adding_encoded_metadata(config['data']['path_metadata_aerial'], data) 
+
+        return data  
+
+    ###### READING CONFIG AND GETTING DATA PATHS
+    path_data = config['data']['HF_data_path'] 
+
+    train_domains, val_domains, test_domains = config['data']['domains_train'], config['data']['domains_val'], config['data']['domains_test']
+
+    with open(config['data']['path_sp_centroids'], 'r') as file: 
+        matching_dict = json.load(file) 
+
+    dict_train = get_data_paths(config, train_domains, path_data, matching_dict) 
+    dict_val   = get_data_paths(config, val_domains,   path_data, matching_dict) 
+    dict_test  = get_data_paths(config, test_domains,  path_data, matching_dict) 
+    
+    return dict_train, dict_val, dict_test    
+    
+    
+    
+    
+    
+    
+    
+    
+ def adding_encoded_metadata(path_metadata_file: str, dict_paths: dict, loc_enc_size: int = 32):
+    """
+    For every aerial image in the dataset, get metadata, encode and add to data dict.
+    """
+    #### encode metadata
+    def coordenc_opt(coords, enc_size=32) -> np.array:
+        d = int(enc_size/2)
+        d_i = np.arange(0, d / 2)
+        freq = 1 / (10e7 ** (2 * d_i / d))
+
+        x,y = coords[0]/10e7, coords[1]/10e7
+        enc = np.zeros(d * 2)
+        enc[0:d:2]    = np.sin(x * freq)
+        enc[1:d:2]    = np.cos(x * freq)
+        enc[d::2]     = np.sin(y * freq)
+        enc[d + 1::2] = np.cos(y * freq)
+        return list(enc)           
+
+    def norm_alti(alti: int) -> float:
+        min_alti = 0
+        max_alti = 3164.9099121094  ### MAX DATASET
+        return [(alti-min_alti) / (max_alti-min_alti)]        
+
+    def format_cam(cam: str) -> np.array:
+        return [[1,0] if 'UCE' in cam else [0,1]][0]
+
+    def cyclical_enc_datetime(date: str, time: str) -> list:
+        def norm(num: float) -> float:
+            return (num-(-1))/(1-(-1))
+        year, month, day = date.split('-')
+        if year == '2018':   enc_y = [1,0,0,0]
+        elif year == '2019': enc_y = [0,1,0,0]
+        elif year == '2020': enc_y = [0,0,1,0]
+        elif year == '2021': enc_y = [0,0,0,1]    
+        sin_month = np.sin(2*np.pi*(int(month)-1/12)) ## months of year
+        cos_month = np.cos(2*np.pi*(int(month)-1/12))    
+        sin_day = np.sin(2*np.pi*(int(day)/31)) ## max days
+        cos_day = np.cos(2*np.pi*(int(day)/31))     
+        h,m=time.split('h')
+        sec_day = int(h) * 3600 + int(m) * 60
+        sin_time = np.sin(2*np.pi*(sec_day/86400)) ## total sec in day
+        cos_time = np.cos(2*np.pi*(sec_day/86400))
+        return enc_y+[norm(sin_month),norm(cos_month),norm(sin_day),norm(cos_day),norm(sin_time),norm(cos_time)] 
+    
+    
+    with open(path_metadata_file, 'r') as f:
+        metadata_dict = json.load(f)              
+    for img in dict_paths['PATH_IMG']:
+        curr_img     = img.split('/')[-1][:-4]
+        enc_coords   = coordenc_opt([metadata_dict[curr_img]["patch_centroid_x"], metadata_dict[curr_img]["patch_centroid_y"]], enc_size=loc_enc_size)
+        enc_alti     = norm_alti(metadata_dict[curr_img]["patch_centroid_z"])
+        enc_camera   = format_cam(metadata_dict[curr_img]['camera'])
+        enc_temporal = cyclical_enc_datetime(metadata_dict[curr_img]['date'], metadata_dict[curr_img]['time'])
+        mtd_enc      = enc_coords+enc_alti+enc_camera+enc_temporal 
+        dict_paths['MTD_AERIAL'].append(mtd_enc)    
+        
+    return dict_paths