Upload 8 files

add pre-trained checkpoints for full and partial point clouds

checkpoints/generation/fpc_1a_latentc3_z4_pc64/ddm/checkpoints/last.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30d5224635de058c47919ceb81ba57e2a4b311b063660e6e9ea914e216dbcbc8
+size 47899897

checkpoints/generation/fpc_1a_latentc3_z4_pc64/ddm/ppc_1a_partial_63cat8k_filtered_latentc3_z16_pc256_180k.py

@@ -0,0 +1,289 @@
+import os
+
+## --------------------  Most frequently changed params here  --------------------
+
+resume_training_from_last = True
+
+max_steps = 180000
+batch_size = 10
+
+num_gpus = 1
+num_workers_per_gpu = 7
+
+# During training, if a ckpt is provided here, it overrides resume_training_from_last and instead resumes training from this ckpt
+vae_ckpt_path = None  # "output/boilerplate_kldanneal_c0.1/vae/checkpoints/last.ckpt"
+ddm_ckpt_path = None
+
+max_scenes = None
+
+## -------------------- Inputs/Shapes ------------------------
+# Input/Output: grasp representation [mrp(3), t(3), cls_success(1), qualities(4)]
+
+pc_num_points = 1024
+pc_latent_dims = 64
+pc_latent_channels = 3
+
+grasp_pose_dims = 6
+num_output_qualities = 0
+grasp_latent_dims = 4
+
+grasp_representation_dims = (
+    grasp_pose_dims + num_output_qualities + 1
+    if num_output_qualities is not None
+    else grasp_pose_dims + 1
+)
+
+## ----------------------- Model -----------------------
+
+dropout = 0.1  # or None
+
+pc_encoder_config = dict(
+    type="PVCNNEncoder",
+    args=dict(
+        in_features=3,
+        n_points=pc_num_points,
+        scale_channels=0.75,
+        scale_voxel_resolution=0.75,
+        num_blocks=(1, 1, 1, 1),
+        out_channels=pc_latent_channels,
+        use_global_attention=False,
+    ),
+)
+
+grasp_encoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        in_features=grasp_representation_dims,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+decoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        block_channels=(32, 64, 128, 256),
+        # out_dim=grasp_pose_dims,
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+loss_config = dict(
+    reconstruction_loss=dict(
+        type="GraspReconstructionLoss",
+        name="reconstruction_loss",
+        args=dict(translation_weight=1, rotation_weight=1),
+    ),
+    latent_loss=dict(
+        type="VAELatentLoss",
+        args=dict(
+            name="grasp_latent",
+            cyclical_annealing=True,
+            num_steps=max_steps,
+            num_cycles=1,
+            ratio=0.5,
+            start=1e-7,
+            stop=0.1,
+        ),
+    ),
+    classification_loss=dict(type="ClassificationLoss", args=dict(weight=0.1)),
+    # quality_loss=dict(type="QualityLoss", args=dict(weight=0.1)),
+)
+
+denoiser_model = dict(
+    type="TimeConditionedResNet1D",
+    args=dict(
+        dim=grasp_latent_dims,
+        channels=1,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+        is_time_conditioned=True,
+        learned_variance=False,
+        learned_sinusoidal_cond=False,
+        random_fourier_features=True,
+        # learned_sinusoidal_dim=16,
+    ),
+)
+# Use `model` for single module to be built. If a list of modules are required to be built, use `models` to make sure the outer
+# See models/builder.py for more info.
+model = dict(
+    vae=dict(
+        model=dict(
+            type="GraspCVAE",
+            args=dict(
+                grasp_latent_size=grasp_latent_dims,
+                pc_latent_size=pc_latent_dims,
+                pc_encoder_config=pc_encoder_config,
+                grasp_encoder_config=grasp_encoder_config,
+                decoder_config=decoder_config,
+                loss_config=loss_config,
+                num_output_qualities=num_output_qualities,
+                intermediate_feature_resolution=16,
+            ),
+        ),
+        ckpt_path=vae_ckpt_path,
+    ),
+    ddm=dict(
+        model=dict(
+            type="GraspLatentDDM",
+            args=dict(
+                model=denoiser_model,
+                latent_in_features=grasp_latent_dims,
+                diffusion_timesteps=1000,
+                noise_scheduler_type="ddpm",
+                diffusion_loss="l2",
+                beta_schedule="linear",
+                is_conditioned=True,
+                joint_training=False,
+                denoising_loss_weight=1,
+                variance_type="fixed_large",
+                elucidated_diffusion=False,
+                beta_start=0.00005,
+                beta_end=0.001,
+            ),
+        ),
+        ckpt_path=ddm_ckpt_path,
+        use_vae_ema_model=True,
+    ),
+)
+## -- Data --
+augs_config = [
+    dict(type="RandomRotation", args=dict(p=0.5, max_angle=180, is_degree=True)),
+    dict(type="PointcloudJitter", args=dict(p=1, sigma=0.005, clip=0.005)),
+    dict(type="RandomPointcloudDropout", args=dict(p=0.5, max_dropout_ratio=0.4)),
+]
+
+root_data_dir = "/mnt/irisgpfs/projects/mis-urso/grasp/data/acronym"
+object_categories = [
+    "Cup",
+    "Mug",
+    "Fork",
+    "Hat",
+    "Bottle",
+    "Bowl",
+    "Car",
+    "Donut",
+    "Laptop",
+    "MousePad",
+    "Pencil",
+    "Plate",
+    "ScrewDriver",
+    "WineBottle",
+    "Backpack",
+    "Bag",
+    "Banana",
+    "Battery",
+    "BeanBag",
+    "Bear",
+    "Book",
+    "Books",
+    "Camera",
+    "CerealBox",
+    "Cookie",
+    "Hammer",
+    "Hanger",
+    "Knife",
+    "MilkCarton",
+    "Painting",
+    "PillBottle",
+    "Plant",
+    "PowerSocket",
+    "PowerStrip",
+    "PS3",
+    "PSP",
+    "Ring",
+    "Scissors",
+    "Shampoo",
+    "Shoes",
+    "Sheep",
+    "Shower",
+    "Sink",
+    "SoapBottle",
+    "SodaCan",
+    "Spoon",
+    "Statue",
+    "Teacup",
+    "Teapot",
+    "ToiletPaper",
+    "ToyFigure",
+    "Wallet",
+    "WineGlass",
+    "Cow",
+    "Sheep",
+    "Cat",
+    "Dog",
+    "Pizza",
+    "Elephant",
+    "Donkey",
+    "RubiksCube",
+    "Tank",
+    "Truck",
+    "USBStick",
+]
+
+train_data = dict(
+    type="AcronymShapenetPointclouds",
+    args=dict(
+        data_root_dir=root_data_dir,
+        batch_num_points_per_pc=pc_num_points,
+        batch_num_grasps_per_pc=100,
+        rotation_repr="mrp",
+        augs_config=augs_config,
+        split="train",
+        batch_failed_grasps_ratio=0,
+        use_dataset_statistics_for_norm=False,
+        filter_categories=object_categories,
+        load_fixed_subset_grasps_per_obj=None,
+        num_repeat_dataset=10,
+    ),
+)
+
+data = dict(
+    train=train_data,
+)
+
+# Patch: Mesh Categories. Used for simulation
+mesh_root = "/home/kuldeep/phd/data/ACRONYM/"
+mesh_root = (
+    mesh_root
+    if os.path.exists(mesh_root)
+    else "/mnt/irisgpfs/users/kbarad/grasp/data/acronym"
+)
+mesh_categories = object_categories
+
+## --------------------  Trainer  --------------------
+## Logger
+logger = dict(type="WandbLogger", project="full-pc-ema-63c")
+
+optimizer = dict(
+    initial_lr=0.001,
+    scheduler=dict(
+        type="MultiStepLR",
+        args=dict(milestones=[int(max_steps / 3), int(2 * max_steps / 3)], gamma=0.1),
+    ),
+)
+
+trainer = dict(
+    max_steps=max_steps,
+    batch_size=batch_size,
+    num_workers=num_workers_per_gpu * num_gpus,
+    accelerator="gpu",
+    devices=num_gpus,
+    strategy="ddp",
+    logger=logger,
+    log_every_n_steps=100,
+    optimizer=optimizer,
+    resume_training_from_last=resume_training_from_last,
+    check_val_every_n_epoch=1,
+    ema=dict(
+        beta=0.990,
+        update_after_step=1000,
+    ),
+    deterministic=True,
+)

checkpoints/generation/fpc_1a_latentc3_z4_pc64/vae/checkpoints/last.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7c6759f52c2fe0895bd54fee24425f720fa4d67fd8454e25cb6e338a03b05d7
+size 40291309

checkpoints/generation/fpc_1a_latentc3_z4_pc64/vae/ppc_1a_partial_63cat8k_filtered_latentc3_z16_pc256_180k.py

@@ -0,0 +1,289 @@
+import os
+
+## --------------------  Most frequently changed params here  --------------------
+
+resume_training_from_last = True
+
+max_steps = 180000
+batch_size = 10
+
+num_gpus = 1
+num_workers_per_gpu = 7
+
+# During training, if a ckpt is provided here, it overrides resume_training_from_last and instead resumes training from this ckpt
+vae_ckpt_path = None  # "output/boilerplate_kldanneal_c0.1/vae/checkpoints/last.ckpt"
+ddm_ckpt_path = None
+
+max_scenes = None
+
+## -------------------- Inputs/Shapes ------------------------
+# Input/Output: grasp representation [mrp(3), t(3), cls_success(1), qualities(4)]
+
+pc_num_points = 1024
+pc_latent_dims = 64
+pc_latent_channels = 3
+
+grasp_pose_dims = 6
+num_output_qualities = 0
+grasp_latent_dims = 4
+
+grasp_representation_dims = (
+    grasp_pose_dims + num_output_qualities + 1
+    if num_output_qualities is not None
+    else grasp_pose_dims + 1
+)
+
+## ----------------------- Model -----------------------
+
+dropout = 0.1  # or None
+
+pc_encoder_config = dict(
+    type="PVCNNEncoder",
+    args=dict(
+        in_features=3,
+        n_points=pc_num_points,
+        scale_channels=0.75,
+        scale_voxel_resolution=0.75,
+        num_blocks=(1, 1, 1, 1),
+        out_channels=pc_latent_channels,
+        use_global_attention=False,
+    ),
+)
+
+grasp_encoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        in_features=grasp_representation_dims,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+decoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        block_channels=(32, 64, 128, 256),
+        # out_dim=grasp_pose_dims,
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+loss_config = dict(
+    reconstruction_loss=dict(
+        type="GraspReconstructionLoss",
+        name="reconstruction_loss",
+        args=dict(translation_weight=1, rotation_weight=1),
+    ),
+    latent_loss=dict(
+        type="VAELatentLoss",
+        args=dict(
+            name="grasp_latent",
+            cyclical_annealing=True,
+            num_steps=max_steps,
+            num_cycles=1,
+            ratio=0.5,
+            start=1e-7,
+            stop=0.1,
+        ),
+    ),
+    classification_loss=dict(type="ClassificationLoss", args=dict(weight=0.1)),
+    # quality_loss=dict(type="QualityLoss", args=dict(weight=0.1)),
+)
+
+denoiser_model = dict(
+    type="TimeConditionedResNet1D",
+    args=dict(
+        dim=grasp_latent_dims,
+        channels=1,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+        is_time_conditioned=True,
+        learned_variance=False,
+        learned_sinusoidal_cond=False,
+        random_fourier_features=True,
+        # learned_sinusoidal_dim=16,
+    ),
+)
+# Use `model` for single module to be built. If a list of modules are required to be built, use `models` to make sure the outer
+# See models/builder.py for more info.
+model = dict(
+    vae=dict(
+        model=dict(
+            type="GraspCVAE",
+            args=dict(
+                grasp_latent_size=grasp_latent_dims,
+                pc_latent_size=pc_latent_dims,
+                pc_encoder_config=pc_encoder_config,
+                grasp_encoder_config=grasp_encoder_config,
+                decoder_config=decoder_config,
+                loss_config=loss_config,
+                num_output_qualities=num_output_qualities,
+                intermediate_feature_resolution=16,
+            ),
+        ),
+        ckpt_path=vae_ckpt_path,
+    ),
+    ddm=dict(
+        model=dict(
+            type="GraspLatentDDM",
+            args=dict(
+                model=denoiser_model,
+                latent_in_features=grasp_latent_dims,
+                diffusion_timesteps=1000,
+                noise_scheduler_type="ddpm",
+                diffusion_loss="l2",
+                beta_schedule="linear",
+                is_conditioned=True,
+                joint_training=False,
+                denoising_loss_weight=1,
+                variance_type="fixed_large",
+                elucidated_diffusion=False,
+                beta_start=0.00005,
+                beta_end=0.001,
+            ),
+        ),
+        ckpt_path=ddm_ckpt_path,
+        use_vae_ema_model=True,
+    ),
+)
+## -- Data --
+augs_config = [
+    dict(type="RandomRotation", args=dict(p=0.5, max_angle=180, is_degree=True)),
+    dict(type="PointcloudJitter", args=dict(p=1, sigma=0.005, clip=0.005)),
+    dict(type="RandomPointcloudDropout", args=dict(p=0.5, max_dropout_ratio=0.4)),
+]
+
+root_data_dir = "/mnt/irisgpfs/projects/mis-urso/grasp/data/acronym"
+object_categories = [
+    "Cup",
+    "Mug",
+    "Fork",
+    "Hat",
+    "Bottle",
+    "Bowl",
+    "Car",
+    "Donut",
+    "Laptop",
+    "MousePad",
+    "Pencil",
+    "Plate",
+    "ScrewDriver",
+    "WineBottle",
+    "Backpack",
+    "Bag",
+    "Banana",
+    "Battery",
+    "BeanBag",
+    "Bear",
+    "Book",
+    "Books",
+    "Camera",
+    "CerealBox",
+    "Cookie",
+    "Hammer",
+    "Hanger",
+    "Knife",
+    "MilkCarton",
+    "Painting",
+    "PillBottle",
+    "Plant",
+    "PowerSocket",
+    "PowerStrip",
+    "PS3",
+    "PSP",
+    "Ring",
+    "Scissors",
+    "Shampoo",
+    "Shoes",
+    "Sheep",
+    "Shower",
+    "Sink",
+    "SoapBottle",
+    "SodaCan",
+    "Spoon",
+    "Statue",
+    "Teacup",
+    "Teapot",
+    "ToiletPaper",
+    "ToyFigure",
+    "Wallet",
+    "WineGlass",
+    "Cow",
+    "Sheep",
+    "Cat",
+    "Dog",
+    "Pizza",
+    "Elephant",
+    "Donkey",
+    "RubiksCube",
+    "Tank",
+    "Truck",
+    "USBStick",
+]
+
+train_data = dict(
+    type="AcronymShapenetPointclouds",
+    args=dict(
+        data_root_dir=root_data_dir,
+        batch_num_points_per_pc=pc_num_points,
+        batch_num_grasps_per_pc=100,
+        rotation_repr="mrp",
+        augs_config=augs_config,
+        split="train",
+        batch_failed_grasps_ratio=0,
+        use_dataset_statistics_for_norm=False,
+        filter_categories=object_categories,
+        load_fixed_subset_grasps_per_obj=None,
+        num_repeat_dataset=10,
+    ),
+)
+
+data = dict(
+    train=train_data,
+)
+
+# Patch: Mesh Categories. Used for simulation
+mesh_root = "/home/kuldeep/phd/data/ACRONYM/"
+mesh_root = (
+    mesh_root
+    if os.path.exists(mesh_root)
+    else "/mnt/irisgpfs/users/kbarad/grasp/data/acronym"
+)
+mesh_categories = object_categories
+
+## --------------------  Trainer  --------------------
+## Logger
+logger = dict(type="WandbLogger", project="full-pc-ema-63c")
+
+optimizer = dict(
+    initial_lr=0.001,
+    scheduler=dict(
+        type="MultiStepLR",
+        args=dict(milestones=[int(max_steps / 3), int(2 * max_steps / 3)], gamma=0.1),
+    ),
+)
+
+trainer = dict(
+    max_steps=max_steps,
+    batch_size=batch_size,
+    num_workers=num_workers_per_gpu * num_gpus,
+    accelerator="gpu",
+    devices=num_gpus,
+    strategy="ddp",
+    logger=logger,
+    log_every_n_steps=100,
+    optimizer=optimizer,
+    resume_training_from_last=resume_training_from_last,
+    check_val_every_n_epoch=1,
+    ema=dict(
+        beta=0.990,
+        update_after_step=1000,
+    ),
+    deterministic=True,
+)

checkpoints/generation/ppc_1a_partial_63cat8k_filtered_latentc3_z16_pc256_180k/ddm/checkpoints/last.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1cb9d2683e0d750df0c163b4ad4d5d9288ec9d7bb2b4246d11f90250dd856f28
+size 25175359

checkpoints/generation/ppc_1a_partial_63cat8k_filtered_latentc3_z16_pc256_180k/ddm/exp16e3_partial_63cat8k_filtered_latentc3_z16_pc256_simple_180k.py

@@ -0,0 +1,225 @@
+import os
+
+# Input/Output: grasp representation [mrp(3), t(3), cls_success(1), qualities(4)]
+grasp_pose_dims = 6
+num_output_qualities = 0
+
+grasp_representation_dims = (
+    grasp_pose_dims + num_output_qualities + 1
+    if num_output_qualities is not None
+    else grasp_pose_dims + 1
+)
+
+grasp_latent_dims = 16
+pc_latent_dims = 256
+pc_latent_channels = 3
+
+pc_num_points = 1024
+batch_num_scenes = 10
+# max_scenes = 10
+
+# Max batch steps or epochs. Only one of them should be defined. If both, steps will considered.
+max_steps = 180000
+max_epochs = None
+
+## Checkpoints:
+# If to auto check the exp directory and resume from last saved checkpoints
+resume_training_from_last = True
+
+# TODO: Not passed in config.
+save_ckpt_every_n_epochs = 50
+# During training, if a ckpt is provided here, it overrides resume_training_from_last and instead resumes training from this ckpt
+vae_ckpt_path = None  # "output/boilerplate_kldanneal_c0.1/vae/checkpoints/last.ckpt"
+ddm_ckpt_path = None
+
+
+## -- Model --
+dropout = 0.1  # or None
+
+pc_encoder_config = dict(
+    type="PVCNNEncoder",
+    args=dict(
+        in_features=3,
+        n_points=pc_num_points,
+        scale_channels=0.75,
+        scale_voxel_resolution=0.75,
+        num_blocks=(1, 1, 1, 1),
+        out_channels=pc_latent_channels,
+        use_global_attention=False,
+    ),
+)
+
+grasp_encoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        in_features=grasp_representation_dims,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+decoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        block_channels=(32, 64, 128, 256),
+        # out_dim=grasp_pose_dims,
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+loss_config = dict(
+    reconstruction_loss=dict(
+        type="GraspReconstructionLoss",
+        name="reconstruction_loss",
+        args=dict(translation_weight=1, rotation_weight=1),
+    ),
+    latent_loss=dict(
+        type="VAELatentLoss",
+        args=dict(
+            name="grasp_latent",
+            cyclical_annealing=True,
+            num_steps=max_steps,
+            num_cycles=1,
+            ratio=0.5,
+            start=1e-7,
+            stop=0.1,
+        ),
+    ),
+    classification_loss=dict(type="ClassificationLoss", args=dict(weight=0.1)),
+    # quality_loss=dict(type="QualityLoss", args=dict(weight=0.1)),
+)
+
+denoiser_model = dict(
+    type="TimeConditionedResNet1D",
+    args=dict(
+        dim=grasp_latent_dims,
+        channels=1,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+        is_time_conditioned=True,
+        learned_variance=False,
+        learned_sinusoidal_cond=False,
+        random_fourier_features=True,
+        # learned_sinusoidal_dim=16,
+    ),
+)
+# Use `model` for single module to be built. If a list of modules are required to be built, use `models` to make sure the outer
+# See models/builder.py for more info.
+models = dict(
+    vae=dict(
+        model=dict(
+            type="GraspCVAE",
+            args=dict(
+                grasp_latent_size=grasp_latent_dims,
+                pc_latent_size=pc_latent_dims,
+                pc_encoder_config=pc_encoder_config,
+                grasp_encoder_config=grasp_encoder_config,
+                decoder_config=decoder_config,
+                loss_config=loss_config,
+                num_output_qualities=num_output_qualities,
+                intermediate_feature_resolution=16,
+            ),
+        ),
+        ckpt_path=vae_ckpt_path,
+    ),
+    ddm=dict(
+        model=dict(
+            type="GraspLatentDDM",
+            args=dict(
+                model=denoiser_model,
+                latent_in_features=grasp_latent_dims,
+                diffusion_timesteps=1000,
+                noise_scheduler_type="ddpm",
+                diffusion_loss="l2",
+                beta_schedule="linear",
+                is_conditioned=True,
+                joint_training=False,
+                denoising_loss_weight=1,
+                variance_type="fixed_large",
+                elucidated_diffusion=False,
+                beta_start=0.00005,
+                beta_end=0.001,
+            ),
+        ),
+        ckpt_path=ddm_ckpt_path,
+    ),
+)
+## -- Data --
+augs_config = [
+    dict(type="RandomRotation", args=dict(p=0.5, max_angle=180, is_degree=True)),
+    dict(type="PointcloudJitter", args=dict(p=1, sigma=0.005, clip=0.005)),
+    dict(type="RandomPointcloudDropout", args=dict(p=0.5, max_dropout_ratio=0.4)),
+]
+
+root_data_dir = (
+    "/mnt/irisgpfs/projects/mis-urso/grasp/data/acronym/renders/objects_filtered_grasps_63cat_8k/"
+)
+camera_json = "data/cameras/camera_d435i_dummy.json"
+max_scenes = None
+train_data = dict(
+    type="AcronymPartialPointclouds",
+    args=dict(
+        data_root_dir=root_data_dir,
+        max_scenes=max_scenes,
+        camera_json=camera_json,
+        num_points_per_pc=pc_num_points,
+        num_grasps_per_obj=100,
+        rotation_repr="mrp",
+        augs_config=augs_config,
+        split="train",
+        depth_px_scale=10000,
+        scene_prefix="scene_",
+        min_usable_pc_points=1024,
+        preempt_load_data=True,
+        use_failed_grasps=False,
+        failed_grasp_ratio=0.3,
+        load_fixed_grasp_transforms=None,
+        is_input_dataset_normalized=False,
+    ),
+)
+
+data = dict(
+    train=train_data,
+)
+
+# Patch: Mesh Categories. Used for simulation
+mesh_root = "/home/kuldeep/phd/data/ACRONYM/"
+mesh_root = (
+    mesh_root
+    if os.path.exists(mesh_root)
+    else "/mnt/irisgpfs/users/kbarad/grasp/data/acronym"
+)
+mesh_categories = ["Cup", "Mug", "Fork", "Hat", "Bottle", "Bowl", "Car", "Donut", "Laptop", "MousePad", "Pencil", "Plate", "ScrewDriver", "WineBottle", "Backpack", "Bag", "Banana", "Battery", "BeanBag", "Bear", "Book", "Books", "Camera", "CerealBox", "Cookie", "Hammer", "Hanger", "Knife", "MilkCarton", "Painting", "PillBottle", "Plant", "PowerSocket", "PowerStrip", "PS3", "PSP", "Ring", "Scissors", "Shampoo", "Shoes", "Sheep", "Shower", "Sink", "SoapBottle", "SodaCan", "Spoon", "Statue", "Teacup", "Teapot", "ToiletPaper", "ToyFigure", "Wallet", "WineGlass", "Cow", "Sheep", "Cat", "Dog", "Pizza", "Elephant", "Donkey", "RubiksCube", "Tank", "Truck", "USBStick"]
+
+## Logger
+logger = dict(type="WandbLogger", project="partial-pc-baseline")
+
+optimizer = dict(
+    initial_lr=0.001,
+    scheduler=dict(
+        type="MultiStepLR",
+        args=dict(milestones=[int(max_steps / 3), int(2 * max_steps / 3)], gamma=0.1),
+    ),
+)
+
+num_gpus = 1
+
+
+steps_or_epochs = (
+    dict(max_steps=max_steps) if max_steps is not None else dict(max_epochs=max_epochs)
+)
+
+train = dict(
+    **steps_or_epochs,
+    batch_size=batch_num_scenes,
+    num_workers=7 * num_gpus,
+    accelerator="gpu",
+    devices=num_gpus,
+    strategy="ddp",
+)

checkpoints/generation/ppc_1a_partial_63cat8k_filtered_latentc3_z16_pc256_180k/vae/checkpoints/last.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f0e017b0790fcb742797dbd6c3a31cb9c48ba62020ebffd08408c3f395d44e7
+size 20985977

checkpoints/generation/ppc_1a_partial_63cat8k_filtered_latentc3_z16_pc256_180k/vae/exp16e3_partial_63cat8k_filtered_latentc3_z16_pc256_simple_180k.py

@@ -0,0 +1,225 @@
+import os
+
+# Input/Output: grasp representation [mrp(3), t(3), cls_success(1), qualities(4)]
+grasp_pose_dims = 6
+num_output_qualities = 0
+
+grasp_representation_dims = (
+    grasp_pose_dims + num_output_qualities + 1
+    if num_output_qualities is not None
+    else grasp_pose_dims + 1
+)
+
+grasp_latent_dims = 16
+pc_latent_dims = 256
+pc_latent_channels = 3
+
+pc_num_points = 1024
+batch_num_scenes = 10
+# max_scenes = 10
+
+# Max batch steps or epochs. Only one of them should be defined. If both, steps will considered.
+max_steps = 180000
+max_epochs = None
+
+## Checkpoints:
+# If to auto check the exp directory and resume from last saved checkpoints
+resume_training_from_last = True
+
+# TODO: Not passed in config.
+save_ckpt_every_n_epochs = 50
+# During training, if a ckpt is provided here, it overrides resume_training_from_last and instead resumes training from this ckpt
+vae_ckpt_path = None  # "output/boilerplate_kldanneal_c0.1/vae/checkpoints/last.ckpt"
+ddm_ckpt_path = None
+
+
+## -- Model --
+dropout = 0.1  # or None
+
+pc_encoder_config = dict(
+    type="PVCNNEncoder",
+    args=dict(
+        in_features=3,
+        n_points=pc_num_points,
+        scale_channels=0.75,
+        scale_voxel_resolution=0.75,
+        num_blocks=(1, 1, 1, 1),
+        out_channels=pc_latent_channels,
+        use_global_attention=False,
+    ),
+)
+
+grasp_encoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        in_features=grasp_representation_dims,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+decoder_config = dict(
+    type="ResNet1D",
+    args=dict(
+        block_channels=(32, 64, 128, 256),
+        # out_dim=grasp_pose_dims,
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+    ),
+)
+
+loss_config = dict(
+    reconstruction_loss=dict(
+        type="GraspReconstructionLoss",
+        name="reconstruction_loss",
+        args=dict(translation_weight=1, rotation_weight=1),
+    ),
+    latent_loss=dict(
+        type="VAELatentLoss",
+        args=dict(
+            name="grasp_latent",
+            cyclical_annealing=True,
+            num_steps=max_steps,
+            num_cycles=1,
+            ratio=0.5,
+            start=1e-7,
+            stop=0.1,
+        ),
+    ),
+    classification_loss=dict(type="ClassificationLoss", args=dict(weight=0.1)),
+    # quality_loss=dict(type="QualityLoss", args=dict(weight=0.1)),
+)
+
+denoiser_model = dict(
+    type="TimeConditionedResNet1D",
+    args=dict(
+        dim=grasp_latent_dims,
+        channels=1,
+        block_channels=(32, 64, 128, 256),
+        input_conditioning_dims=pc_latent_dims,
+        resnet_block_groups=4,
+        dropout=dropout,
+        is_time_conditioned=True,
+        learned_variance=False,
+        learned_sinusoidal_cond=False,
+        random_fourier_features=True,
+        # learned_sinusoidal_dim=16,
+    ),
+)
+# Use `model` for single module to be built. If a list of modules are required to be built, use `models` to make sure the outer
+# See models/builder.py for more info.
+models = dict(
+    vae=dict(
+        model=dict(
+            type="GraspCVAE",
+            args=dict(
+                grasp_latent_size=grasp_latent_dims,
+                pc_latent_size=pc_latent_dims,
+                pc_encoder_config=pc_encoder_config,
+                grasp_encoder_config=grasp_encoder_config,
+                decoder_config=decoder_config,
+                loss_config=loss_config,
+                num_output_qualities=num_output_qualities,
+                intermediate_feature_resolution=16,
+            ),
+        ),
+        ckpt_path=vae_ckpt_path,
+    ),
+    ddm=dict(
+        model=dict(
+            type="GraspLatentDDM",
+            args=dict(
+                model=denoiser_model,
+                latent_in_features=grasp_latent_dims,
+                diffusion_timesteps=1000,
+                noise_scheduler_type="ddpm",
+                diffusion_loss="l2",
+                beta_schedule="linear",
+                is_conditioned=True,
+                joint_training=False,
+                denoising_loss_weight=1,
+                variance_type="fixed_large",
+                elucidated_diffusion=False,
+                beta_start=0.00005,
+                beta_end=0.001,
+            ),
+        ),
+        ckpt_path=ddm_ckpt_path,
+    ),
+)
+## -- Data --
+augs_config = [
+    dict(type="RandomRotation", args=dict(p=0.5, max_angle=180, is_degree=True)),
+    dict(type="PointcloudJitter", args=dict(p=1, sigma=0.005, clip=0.005)),
+    dict(type="RandomPointcloudDropout", args=dict(p=0.5, max_dropout_ratio=0.4)),
+]
+
+root_data_dir = (
+    "/mnt/irisgpfs/projects/mis-urso/grasp/data/acronym/renders/objects_filtered_grasps_63cat_8k/"
+)
+camera_json = "data/cameras/camera_d435i_dummy.json"
+max_scenes = None
+train_data = dict(
+    type="AcronymPartialPointclouds",
+    args=dict(
+        data_root_dir=root_data_dir,
+        max_scenes=max_scenes,
+        camera_json=camera_json,
+        num_points_per_pc=pc_num_points,
+        num_grasps_per_obj=100,
+        rotation_repr="mrp",
+        augs_config=augs_config,
+        split="train",
+        depth_px_scale=10000,
+        scene_prefix="scene_",
+        min_usable_pc_points=1024,
+        preempt_load_data=True,
+        use_failed_grasps=False,
+        failed_grasp_ratio=0.3,
+        load_fixed_grasp_transforms=None,
+        is_input_dataset_normalized=False,
+    ),
+)
+
+data = dict(
+    train=train_data,
+)
+
+# Patch: Mesh Categories. Used for simulation
+mesh_root = "/home/kuldeep/phd/data/ACRONYM/"
+mesh_root = (
+    mesh_root
+    if os.path.exists(mesh_root)
+    else "/mnt/irisgpfs/users/kbarad/grasp/data/acronym"
+)
+mesh_categories = ["Cup", "Mug", "Fork", "Hat", "Bottle", "Bowl", "Car", "Donut", "Laptop", "MousePad", "Pencil", "Plate", "ScrewDriver", "WineBottle", "Backpack", "Bag", "Banana", "Battery", "BeanBag", "Bear", "Book", "Books", "Camera", "CerealBox", "Cookie", "Hammer", "Hanger", "Knife", "MilkCarton", "Painting", "PillBottle", "Plant", "PowerSocket", "PowerStrip", "PS3", "PSP", "Ring", "Scissors", "Shampoo", "Shoes", "Sheep", "Shower", "Sink", "SoapBottle", "SodaCan", "Spoon", "Statue", "Teacup", "Teapot", "ToiletPaper", "ToyFigure", "Wallet", "WineGlass", "Cow", "Sheep", "Cat", "Dog", "Pizza", "Elephant", "Donkey", "RubiksCube", "Tank", "Truck", "USBStick"]
+
+## Logger
+logger = dict(type="WandbLogger", project="partial-pc-baseline")
+
+optimizer = dict(
+    initial_lr=0.001,
+    scheduler=dict(
+        type="MultiStepLR",
+        args=dict(milestones=[int(max_steps / 3), int(2 * max_steps / 3)], gamma=0.1),
+    ),
+)
+
+num_gpus = 1
+
+
+steps_or_epochs = (
+    dict(max_steps=max_steps) if max_steps is not None else dict(max_epochs=max_epochs)
+)
+
+train = dict(
+    **steps_or_epochs,
+    batch_size=batch_num_scenes,
+    num_workers=7 * num_gpus,
+    accelerator="gpu",
+    devices=num_gpus,
+    strategy="ddp",
+)