remove redundant comments

examples/openlrm/configs/train-sample-base.yaml

@@ -1,6 +1,5 @@
 experiment:
     type: lrm
-    # seed: 42 # args.seed
     parent: lrm-objaverse
     child: base
 
@@ -40,47 +39,14 @@ dataset:
     pin_mem: true
 
 train:
-    # mixed_precision: bf16  # REPLACE THIS BASED ON GPU TYPE
     find_unused_parameters: false
     loss:
         pixel_weight: 1.0
         perceptual_weight: 1.0
         tv_weight: 5e-4
-    optim:
-        # lr: 4e-4 # args.start_learning_rate
-        # weight_decay: 0.05  #args.weight_decay
-        # beta1: 0.9 #args.betas
-        # beta2: 0.95 #args.betas
-        # clip_grad_norm: 1.0 # args.max_grad_norm
-    scheduler:
-        # type: cosine_annealing_warm_restarts_lr # args.scheduler
-        # warmup_real_iters: 300  # args.warmup_steps
-    batch_size: 1 # REPLACE THIS
-    accum_steps: 1  # REPLACE THIS
-    epochs: 10000 # 60  # REPLACE THIS
-    debug_global_steps: null
+    batch_size: 1
 
 val:
-    batch_size: 1 #4
-    global_step_period: 100 #1000
+    batch_size: 1
+    global_step_period: 1000
     debug_batches: null
-
-saver:
-    auto_resume: true
-    load_model: null
-    # checkpoint_root: ./exps/checkpoints # args.output_path
-    # checkpoint_global_steps: 100 #1000 # args.ckpt_save_steps(unit:step) / args.ckpt_save_interval (unit: epoch)
-    checkpoint_keep_level: 5
-
-# Not to use
-logger:
-    stream_level: WARNING
-    log_level: INFO
-    log_root: ./exps/logs
-    tracker_root: ./exps/trackers
-    enable_profiler: false
-    trackers:
-        - tensorboard
-    image_monitor:
-        train_global_steps: 100
-        samples_per_log: 4

examples/openlrm/configs/train-sample-large.yaml

@@ -1,15 +1,14 @@
 experiment:
     type: lrm
-    # seed: 42 # args.seed
     parent: lrm-objaverse
     child: large
 
 model:
     camera_embed_dim: 1024
-    rendering_samples_per_ray: 96 # 128
-    transformer_dim: 768 # 1024
-    transformer_layers: 12 # 16
-    transformer_heads: 12 # 16
+    rendering_samples_per_ray: 96   # w/ larger memory use: 128
+    transformer_dim: 768            # w/ larger memory use: 1024
+    transformer_layers: 12          # w/ larger net use: 16
+    transformer_heads: 12           # w/ larger memory use: 16
     triplane_low_res: 32
     triplane_high_res: 64
     triplane_dim: 80
@@ -40,47 +39,14 @@ dataset:
     pin_mem: true
 
 train:
-    # mixed_precision: bf16  # REPLACE THIS BASED ON GPU TYPE
     find_unused_parameters: false
     loss:
         pixel_weight: 1.0
         perceptual_weight: 1.0
         tv_weight: 5e-4
-    optim:
-        # lr: 4e-4 # args.start_learning_rate
-        # weight_decay: 0.05  #args.weight_decay
-        # beta1: 0.9 #args.betas
-        # beta2: 0.95 #args.betas
-        # clip_grad_norm: 1.0 # args.max_grad_norm
-    scheduler:
-        # type: cosine_annealing_warm_restarts_lr # args.scheduler
-        # warmup_real_iters: 300  # args.warmup_steps
-    batch_size: 1 # REPLACE THIS
-    accum_steps: 1  # REPLACE THIS
-    epochs: 10000 # 60  # REPLACE THIS
-    debug_global_steps: null
+    batch_size: 1
 
 val:
-    batch_size: 1 #4
-    global_step_period: 100 #1000
+    batch_size: 1
+    global_step_period: 1000
     debug_batches: null
-
-saver:
-    auto_resume: true
-    load_model: null
-    # checkpoint_root: ./exps/checkpoints # args.output_path
-    # checkpoint_global_steps: 100 #1000 # args.ckpt_save_steps(unit:step) / args.ckpt_save_interval (unit: epoch)
-    checkpoint_keep_level: 5
-
-# Not to use
-logger:
-    stream_level: WARNING
-    log_level: INFO
-    log_root: ./exps/logs
-    tracker_root: ./exps/trackers
-    enable_profiler: false
-    trackers:
-        - tensorboard
-    image_monitor:
-        train_global_steps: 100
-        samples_per_log: 4

examples/openlrm/configs/train-sample-small.yaml

@@ -1,6 +1,5 @@
 experiment:
     type: lrm
-    # seed: 42 # args.seed
     parent: lrm-objaverse
     child: small
 
@@ -40,47 +39,14 @@ dataset:
     pin_mem: true
 
 train:
-    # mixed_precision: bf16  # REPLACE THIS BASED ON GPU TYPE
     find_unused_parameters: false
     loss:
         pixel_weight: 1.0
         perceptual_weight: 1.0
         tv_weight: 5e-4
-    optim:
-        # lr: 4e-4 # args.start_learning_rate
-        # weight_decay: 0.05  #args.weight_decay
-        # beta1: 0.9 #args.betas
-        # beta2: 0.95 #args.betas
-        # clip_grad_norm: 1.0 # args.max_grad_norm
-    scheduler:
-        # type: cosine_annealing_warm_restarts_lr # args.scheduler
-        # warmup_real_iters: 300  # args.warmup_steps
-    batch_size: 1 # REPLACE THIS
-    accum_steps: 1  # REPLACE THIS
-    epochs: 10000 # 60  # REPLACE THIS
-    debug_global_steps: null
+    batch_size: 1
 
 val:
-    batch_size: 1 #4
-    global_step_period: 100 #1000
+    batch_size: 1
+    global_step_period: 1000
     debug_batches: null
-
-saver:
-    auto_resume: true
-    load_model: null
-    # checkpoint_root: ./exps/checkpoints # args.output_path
-    # checkpoint_global_steps: 100 #1000 # args.ckpt_save_steps(unit:step) / args.ckpt_save_interval (unit: epoch)
-    checkpoint_keep_level: 5
-
-# Not to use
-logger:
-    stream_level: WARNING
-    log_level: INFO
-    log_root: ./exps/logs
-    tracker_root: ./exps/trackers
-    enable_profiler: false
-    trackers:
-        - tensorboard
-    image_monitor:
-        train_global_steps: 100
-        samples_per_log: 4

examples/openlrm/eval.py

@@ -19,10 +19,6 @@
 
 from mindone.utils.logger import set_logger
 
-# __dir__ = os.path.dirname(os.path.abspath(__file__))
-# sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "../..")))  # for loading mindone
-
-
 logger = logging.getLogger(__name__)
 
 
@@ -99,7 +95,6 @@ def load_data(root_dirs, uid, render_image_res, num_all_views=32, normalize_came
 def evaluate(args):
     save_dir = args.output_path
 
-    # device_num = 1
     rank_id = 0
     ms.set_context(
         mode=args.mode,
@@ -202,13 +197,12 @@ def evaluate(args):
         )[
             0
         ]  # only batch=1  [M, C, H, W]
-        # print(f"render_images {render_images.shape}")
+
         batch_time = time.time() - start_time
         batches_time.append(batch_time)
         logger.info("Batch time cost: %.3fs.", batch_time)
 
         target_images = data_batch["target_images"]  # [M, C, H, W]
-        # print(f"target_images {target_images.shape}")
 
         for view_id, (target, pred) in enumerate(zip(target_images, render_images)):
             target = target.asnumpy()
@@ -219,6 +213,8 @@ def evaluate(args):
             # save mviews outputs
             im = Image.fromarray(np.clip(pred.transpose(1, 2, 0) * 255, 0, 255).astype(np.uint8))
             im.save(os.path.join(image_path, f"{uids[index]}_{view_id:03d}.jpg"))
+
+            # if save target image:
             # im = Image.fromarray(np.clip(target.transpose(1, 2, 0) * 255, 0, 255).astype(np.uint8))
             # im.save(os.path.join(image_path, f"{uids[index]}_{view_id:03d}_target.jpg"))
 

examples/openlrm/openlrm/datasets/base.py

@@ -30,8 +30,6 @@ def __init__(self, root_dirs: list[str], meta_path: str):
         super().__init__()
         self.root_dirs = root_dirs
         self.uids = self._load_uids(meta_path)
-        # self.uids = self.uids[:1] #TODO: debug use, delete later
-        print(f"uids: {self.uids}")
 
     def __len__(self):
         return len(self.uids)
@@ -66,9 +64,7 @@ def _load_rgba_image(file_path, bg_color: float = 1.0, resize=None, crop_pos=Non
             rgba = vision.Resize([resize, resize], Inter.BICUBIC)(rgba)
         if (crop_pos is not None) and (crop_size is not None):  # rand crop
             assert (crop_pos[0] + crop_size <= rgba.shape[0]) and (crop_pos[1] + crop_size <= rgba.shape[1])
-            # print(f"crop rgba: {rgba.shape}")
             rgba = vision.Crop(crop_pos, crop_size)(rgba)
-            # print(f"crop_pos {crop_pos}, crop_size {crop_size} cropped rgba {rgba.shape}")
 
         # convert to Tensor, in shape [B, C, H, W]
         rgba = ms.Tensor(rgba).float() / 255.0

examples/openlrm/openlrm/launch.py

@@ -12,9 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# import os, sys
-# sys.path.insert(0, os.path.abspath("../.."))  # for mindone
-
 import argparse
 
 from openlrm.runners import REGISTRY_RUNNERS

examples/openlrm/openlrm/losses/lpips.py

@@ -49,11 +49,9 @@ def construct(self, input, target, normalize=False):
         else:
             in0_input, in1_input = input, target
         outs0, outs1 = self.net(in0_input), self.net(in1_input)
-        val = 0  # ms.Tensor(0, dtype=input.dtype)
+        val = 0
         for kk in range(len(self.chns)):
             diff = (normalize_tensor(outs0[kk]) - normalize_tensor(outs1[kk])) ** 2
-            # res += spatial_average(lins[kk](diff), keepdim=True)
-            # lin_layer = lins[kk]
             val += ops.mean(self.lins[kk](diff), axis=[2, 3], keep_dims=True)
         return val
 

examples/openlrm/openlrm/models/encoders/dino.py

@@ -423,34 +423,19 @@ def _init_weights(self, module: Union[nn.Dense, nn.Conv2d, nn.LayerNorm]) -> Non
                     TruncatedNormal(sigma=self.config.initializer_range, mean=0.0), module.weight.data.shape, ms.float32
                 )
             )
-            # module.weight.data = nn.init.trunc_normal_(
-            #     module.weight.data.to(ms.float32), mean=0.0, std=self.config.initializer_range
-            # ).to(module.weight.dtype)
             if module.bias is not None:
                 module.bias.set_data(initializer("Zero", module.bias.shape, ms.float32))
-                # module.bias.data.zero_()
         elif isinstance(module, nn.LayerNorm):
             module.beta.data.set_data(initializer("Zero", module.beta.data.shape, ms.float32))
             module.gamma.data.set_data(initializer("One", module.gamma.data.shape, ms.float32))
         elif isinstance(module, ViTEmbeddings):
-            # module.position_embeddings.data = nn.init.trunc_normal_(
-            #     module.position_embeddings.data.to(ms.float32),
-            #     mean=0.0,
-            #     std=self.config.initializer_range,
-            # ).to(module.position_embeddings.dtype)
             module.position_embeddings.data.set_data(
                 initializer(
                     TruncatedNormal(sigma=self.config.initializer_range, mean=0.0),
                     module.position_embeddings.data.shape,
                     ms.float32,
                 )
             )
-
-            # module.cls_token.data = nn.init.trunc_normal_(
-            #     module.cls_token.data.to(ms.float32),
-            #     mean=0.0,
-            #     std=self.config.initializer_range,
-            # ).to(module.cls_token.dtype)
             module.cls_token.data.set_data(
                 initializer(
                     TruncatedNormal(sigma=self.config.initializer_range, mean=0.0),

examples/openlrm/openlrm/models/encoders/dinov2/layers/__init__.py

@@ -16,7 +16,7 @@
 # Avoid using nested tensor for now, deprecating usage of NestedTensorBlock
 from .block import Block, BlockWithModulation
 
-# from .dino_head import DINOHead #not use
+# from .dino_head import DINOHead # not use
 from .mlp import Mlp
 from .patch_embed import PatchEmbed
 from .swiglu_ffn import SwiGLUFFN, SwiGLUFFNFused

examples/openlrm/openlrm/models/encoders/dinov2/layers/attention.py

@@ -82,7 +82,7 @@ def construct(self, x: Tensor, attn_bias=None) -> Tensor:
         v = v.swapaxes(1, 2)
         # 'b n h d' -> (b, h=num_head, n, d) == BNSD
 
-        if attn_bias is not None:  # TODO
+        if attn_bias is not None:  # TODO & FIXME: add this function
             raise AssertionError(f"Do not support MemEffAttention in Dinov2 with attn_bias {attn_bias} yet!")
             # q_bias, k_bias, v_bias = attn_bias.chunk(3)
             # q = q + q_bias

examples/openlrm/openlrm/models/encoders/dinov2/layers/block.py

@@ -47,7 +47,7 @@ def __init__(
         ffn_layer: Callable[..., nn.Cell] = Mlp,
     ) -> None:
         super().__init__()
-        # print(f"biases: qkv: {qkv_bias}, proj: {proj_bias}, ffn: {ffn_bias}")
+
         self.norm1 = norm_layer((dim,))
         self.attn = attn_class(
             dim,
@@ -95,7 +95,7 @@ def ffn_residual_func(x: Tensor) -> Tensor:
             )
         elif self.training and self.sample_drop_ratio > 0.0:
             x = x + self.drop_path1(attn_residual_func(x))
-            x = x + self.drop_path1(ffn_residual_func(x))  # FIXME: drop_path2
+            x = x + self.drop_path1(ffn_residual_func(x))
         else:
             x = x + attn_residual_func(x)
             x = x + ffn_residual_func(x)

examples/openlrm/openlrm/models/encoders/dinov2/layers/dino_head.py

@@ -24,9 +24,6 @@ def __init__(
         nlayers = max(nlayers, 1)
         self.mlp = _build_mlp(nlayers, in_dim, bottleneck_dim, hidden_dim=hidden_dim, use_bn=use_bn, bias=mlp_bias)
         self.apply(self._init_weights)
-        # from torch.nn.utils import weight_norm # Deprecated
-        # self.last_layer = weight_norm(nn.Linear(bottleneck_dim, out_dim, has_bias=False))
-        # self.last_layer.weight_g.set_data(initializer("ones"))
         self.last_layer = nn.Dense(bottleneck_dim, out_dim, has_bias=False)
         self.last_layer.weight.set_data(initializer("ones"))
 

examples/openlrm/openlrm/models/encoders/dinov2/layers/patch_embed.py

@@ -75,7 +75,7 @@ def construct(self, x: Tensor) -> Tensor:
 
         x = self.proj(x)  # B C H W
         H, W = x.shape[2], x.shape[3]
-        x = x.flatten(start_dim=2).swapaxes(1, 2)  # B HW C
+        x = x.flatten(start_dim=2).swapaxes(1, 2)  # B H W C
         x = self.norm(x)
         if not self.flatten_embedding:
             x = x.reshape(-1, H, W, self.embed_dim)  # B H W C

examples/openlrm/openlrm/models/modeling_lrm.py

@@ -245,4 +245,7 @@ def construct(
             render_results["images_rgb"].shape[1] == M
         ), "Number of rendered views should be consistent with render_cameras"
 
-        return (planes, render_results["images_rgb"]) # keys also includes 'images_depth', 'images_weight' which are never used.
+        return (
+            planes,
+            render_results["images_rgb"],
+        )  # keys also include 'images_depth', 'images_weight' which are never used.

examples/openlrm/openlrm/runners/train/lrm.py

@@ -152,7 +152,7 @@ def _build_model(self, cfg):
 
     # Build training utils: lr, optim, callbacks, trainer
     def _build_utils(self, args, cfg, dataset_size):
-        total_train_steps = cfg.train.epochs
+        total_train_steps = args.epochs
         # build learning rate scheduler
         if not args.decay_steps:
             args.decay_steps = total_train_steps - args.warmup_steps  # fix lr scheduling

examples/openlrm/openlrm/utils/hf_hub.py

@@ -24,8 +24,6 @@
 
 from mindone.safetensors.mindspore import load_file
 
-# from huggingface_hub.utils import validate_hf_hub_args
-
 
 def wrap_model_hub(model_cls: nn.Cell):
     """
@@ -136,9 +134,6 @@ def _from_pretrained(
             )
 
             print(f"state_dict.dtype {state_dict[loaded_keys[0]].dtype}")  # float32
-            # print(f"model.dtype {model.dtype}")
-            # if state_dict[loaded_keys[0]].dtype != model.dtype:
-            #     model = model.to(state_dict[loaded_keys[0]].dtype)
 
             # Instantiate the model
             param_not_load, ckpt_not_load = ms.load_param_into_net(model, state_dict, strict_load=strict)