fix a rotation pred err

MaverickPeter · Jul 27, 2021 · 7f2ab04 · 7f2ab04
1 parent 25c0f27
commit 7f2ab04
Show file tree

Hide file tree

Showing 4 changed files with 19 additions and 17 deletions.
diff --git a/config.py b/config.py
@@ -2,7 +2,7 @@
 EXPERIMENT_NAME = 'occ'
 
 NUM_POINTS = 4096
-FEATURE_OUTPUT_DIM = 256
+FEATURE_OUTPUT_DIM = 1024
 RESULTS_FOLDER = "results/"
 OUTPUT_FILE = "results/results.txt"
 
@@ -17,7 +17,7 @@
 TRAIN_NEGATIVES_PER_QUERY = 4
 DECAY_STEP = 100000
 DECAY_RATE = 0.7
-BASE_LEARNING_RATE = 0.0001
+BASE_LEARNING_RATE = 0.00001
 MOMENTUM = 0.9
 OPTIMIZER = 'ADAM'
 MAX_EPOCH = 20

diff --git a/evaluate.py b/evaluate.py
@@ -160,6 +160,7 @@ def get_latent_vectors(model, dict_to_process):
         queries = load_pc_files(file_names)
         queries = np.array(queries, dtype=np.float32)
         heading = np.array(heading)
+        heading = heading / np.pi * 180.
 
         with torch.no_grad():
             feed_tensor = torch.from_numpy(queries).float()
@@ -203,14 +204,7 @@ def get_latent_vectors(model, dict_to_process):
         queries = np.array(queries, dtype=np.float32)
         heading_edge = np.array(heading)
         heading_edge = heading_edge.reshape(-1,1)
-
-        randomYaw = (np.random.rand(heading_edge.shape[0],1) - 0.5) * 90.
         heading_edge = heading_edge / np.pi * 180.
-        heading_edge += randomYaw
-
-        for b in range(queries.shape[0]):
-            for dims in range(queries.shape[1]):
-                queries[b,dims,...] = rotation_on_SCI(queries[b,dims,...], randomYaw[b,0])
 
         with torch.no_grad():
             feed_tensor = torch.from_numpy(queries).float()
@@ -287,7 +281,8 @@ def get_recall(m, n, DATABASE_VECTORS, QUERY_VECTORS, QUERY_SETS, FFT_DATABASE_V
             if indices[0][j] in true_neighbors:
                 if(j == 0):
                     count += 1
-                    gt_angle = GT_sc_angle_convert(-gt_yaw_query[i] + gt_yaw_database[indices[0][j]], cfg.num_sector)
+
+                    gt_angle = GT_sc_angle_convert(gt_yaw_database[indices[0][j]] - gt_yaw_query[i], cfg.num_sector)
                     angle, _ = phase_corr(database_fft[indices[0][j]], query_fft[i], device, corr2soft)
                     angle = angle.detach().cpu().numpy()
 
@@ -476,7 +471,7 @@ def rotation_on_SCI(sc, rotation):
     cfg.DECAY_RATE = FLAGS.decay_rate
     cfg.RESULTS_FOLDER = FLAGS.results_dir
 
-    cfg.LOG_DIR = 'log/'
+    cfg.LOG_DIR = 'log/100celoss/'
     cfg.OUTPUT_FILE = cfg.RESULTS_FOLDER + 'results.txt'
     cfg.MODEL_FILENAME = "model.ckpt"
     cfg.DATASET_FOLDER = FLAGS.dataset_folder

diff --git a/generating_queries/nclt/config.py b/generating_queries/nclt/config.py
@@ -2,7 +2,7 @@
 EXPERIMENT_NAME = 'occ'
 
 NUM_POINTS = 4096
-FEATURE_OUTPUT_DIM = 256
+FEATURE_OUTPUT_DIM = 1024
 RESULTS_FOLDER = "results/"
 OUTPUT_FILE = "results/results.txt"
 
@@ -17,7 +17,7 @@
 TRAIN_NEGATIVES_PER_QUERY = 4
 DECAY_STEP = 100000
 DECAY_RATE = 0.7
-BASE_LEARNING_RATE = 0.0001
+BASE_LEARNING_RATE = 0.00001
 MOMENTUM = 0.9
 OPTIMIZER = 'ADAM'
 MAX_EPOCH = 20

diff --git a/train_DiSCO.py b/train_DiSCO.py
@@ -41,8 +41,8 @@
                     help='Epoch to run [default: 20]')
 parser.add_argument('--batch_num_queries', type=int, default=2,
                     help='Batch Size during training [default: 2]')
-parser.add_argument('--learning_rate', type=float, default=0.0001,
-                    help='Initial learning rate [default: 0.0001]')
+parser.add_argument('--learning_rate', type=float, default=0.00001,
+                    help='Initial learning rate [default: 0.00001]')
 parser.add_argument('--momentum', type=float, default=0.9,
                     help='Initial learning rate [default: 0.9]')
 parser.add_argument('--optimizer', default='adam',
@@ -57,7 +57,7 @@
                     help='Margin for hinge loss [default: 0.2]')
 parser.add_argument('--loss_function', default='quadruplet', choices=[
                     'triplet', 'quadruplet'], help='triplet or quadruplet [default: quadruplet]')
-parser.add_argument('--loss_not_lazy', action='store_false',
+parser.add_argument('--loss_not_lazy', action='store_true',
                     help='If present, do not use lazy variant of loss')
 parser.add_argument('--loss_ignore_zero_batch', action='store_true',
                     help='If present, mean only batches with loss > 0.0')
@@ -314,6 +314,7 @@ def train_one_epoch(model, optimizer, corr2soft, optimizer_c2s, train_writer, lo
         for dim in range(queries.shape[2]):
             queries[0,0,dim,...] = rotation_on_SCI(queries[0,0,dim,...], randomYaw)
 
+        # rotate positives to original query so we only use random yaw to gernerate yaw diff
         for b in range(positives.shape[1]):
             for dims in range(positives.shape[2]):
                 positives[0,b,dims,...] = rotation_on_SCI(positives[0,b,dims,...], (heading[0]-heading[1+b])/np.pi*180)
@@ -336,12 +337,18 @@ def train_one_epoch(model, optimizer, corr2soft, optimizer_c2s, train_writer, lo
         # visualization
         train_writer.add_scalar("Loss", (loss - yaw_loss).cpu().item(), TOTAL_ITERATIONS)
         train_writer.add_scalar("Yaw_Loss", yaw_loss.cpu().item(), TOTAL_ITERATIONS)
-        train_writer.add_scalar("Yaw_Loss_L1", yaw_loss_l1.cpu().item(), TOTAL_ITERATIONS)
+        train_writer.add_scalar("Yaw_Loss_L1", yaw_loss_l1.item(), TOTAL_ITERATIONS)
 
         train_writer.add_image("query", outfft[0,0,...].unsqueeze(0).detach().cpu(), TOTAL_ITERATIONS)
         train_writer.add_image("positive", outfft[1,0,...].unsqueeze(0).detach().cpu(), TOTAL_ITERATIONS)
         train_writer.add_image("corr", corr[0,...].unsqueeze(0).detach().cpu(), TOTAL_ITERATIONS)
 
+        # if(yaw_loss_l1 > 10):
+        #     print(positives.shape)
+        #     train_writer.add_image("query image", torch.from_numpy(queries[0,0,...].sum(axis=0)).unsqueeze(0).float(), TOTAL_ITERATIONS)
+        #     train_writer.add_image("positive image1", torch.from_numpy(positives[0,0,...].sum(axis=0)).unsqueeze(0).float(), TOTAL_ITERATIONS)
+        #     train_writer.add_image("positive image2", torch.from_numpy(positives[0,1,...].sum(axis=0)).unsqueeze(0).float(), TOTAL_ITERATIONS)
+
         TOTAL_ITERATIONS += cfg.BATCH_NUM_QUERIES
 
         # -------------------- Evaluating ---------------------------