add placeholder huggingface space app, gradio version == 4.15.0

.gitattributes

@@ -1,2 +1,3 @@
 *.tar filter=lfs diff=lfs merge=lfs -text
 *.npy filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,13 +1,14 @@
 ---
 title: Text2EMotionDiffuse
-emoji: 🏢
+emoji: 🧠
 colorFrom: blue
 colorTo: red
 sdk: gradio
 sdk_version: 3.44.1
-app_file: app.py
+app_file: text2motion/app.py
 pinned: false
 license: mit
+tags: diffusion, motiondiffuse, text2motion, smplx, smpl, smpl-x, smplify-x
 ---
 <div align="center">
 

text2motion/Makefile

@@ -12,6 +12,11 @@ GT_FILE=s2/cubesmall_lift # ground-truth
 FRAMES=60
 MEAN_EMOTION=surprise
 
+# to push changes to space, run 'git push space main'
+# make sure to do 'git remote add space https://huggingface.co/spaces/ellemac/Text2EMotionDiffuse'
+huggingface:
+	cd ${ROOT_DIR} && ${PYTHON_BIN} -m app
+
 mean-mesh:
 	cd $(ROOT_DIR) && vglrun ${PYTHON_BIN} -m datasets.mean_mesh \
 	--emotion ${MEAN_EMOTION} \
@@ -42,9 +47,8 @@ train: w_stats
 	--seed ${SEED} \
 	--use_wandb \
 
-# get makes model generate seq according to text and writes result to npy file
+# get-npy makes model generate seq according to text and writes result to npy file
 gen-npy:
-# checkpoints/t2m/t2m_motiondiffuse/opt.txt
 	cd ${ROOT_DIR} && ${PYTHON_BIN} -m tools.inference \
 	--opt_path ${MODEL_DIR}/opt.txt \
 	--which_epoch ${EPOCH} \

text2motion/app.py

@@ -0,0 +1,27 @@
+import os
+import sys
+import gradio as gr
+import logging
+
+os.makedirs("outputs", exist_ok=True)
+sys.path.insert(0, ".")
+
+def generate(prompt, length):
+    logging.warning("NOT generating per the prompt [TODO], just returning a denoising gif.")
+    result_path = "outputs/denoising_grab_model.gif"
+    return result_path
+
+
+demo = gr.Interface(
+    fn=generate,
+    inputs=["text", gr.Slider(5, 30, value=10)],
+    examples=[
+        ["happily flying airplane", 10],
+    ],
+    outputs="image",
+    title="COMING SOON: Text2EMotionDiffuse Demo. Currently: shows denoising gif for any prompt.",
+    description="COMING SOON, SPACE NOT CURRENTLY CONFIGURED TO HANDLE PROMPTS, but please Github: https://github.com/ellemcfarlane/Text2EMotionDiffuse",
+)
+
+if __name__ == "__main__":
+    demo.launch()

text2motion/datasets/motionx_explorer.py

@@ -1,19 +1,21 @@
 import argparse
 import logging as log
 import os
-import time
 from collections import defaultdict
 from os.path import join as pjoin
 from typing import Dict, Optional, Tuple
-
+import pyrender
+from tqdm import tqdm
+import trimesh
+import numpy as np
+import os
+import imageio
 import numpy as np
 import smplx
 import torch
 from numpy.typing import ArrayLike
 from torch import Tensor
 
-from .rendering import render_meshes
-
 log.basicConfig(
     level=log.INFO,
     format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
@@ -23,7 +25,9 @@ log.basicConfig(
 MOCAP_DATASETS = {"egobody", "grab", "humanml", "grab_motion"}
 DATA_DIR = "data"
 MODELS_DIR = "models"
-MOCAP_FACE_DIR = f"{DATA_DIR}/face_motion_data/smplx_322" # contains face motion data only
+MOCAP_FACE_DIR = (
+    f"{DATA_DIR}/face_motion_data/smplx_322"  # contains face motion data only
+)
 MOTION_DIR = f"{DATA_DIR}/motion_data/smplx_322"
 ACTION_LABEL_DIR = f"{DATA_DIR}/semantic_labels"
 EMOTION_LABEL_DIR = f"{DATA_DIR}/face_texts"
@@ -40,20 +44,24 @@ NOTE: I think they are wrong about n_body_joints though, data indicates it's act
 
 MY_REPO = os.path.abspath("")
 log.info(f"MY_REPO: {MY_REPO}")
-NUM_BODY_JOINTS = 23 - 2  # SMPL has hand joints but we're replacing them with more detailed ones by SMLP-X, paper: 22x3 total body dims * not sure why paper says 22
-NUM_JAW_JOINTS = 1 # 1x3 total jaw dims
+NUM_BODY_JOINTS = (
+    23 - 2
+)  # SMPL has hand joints but we're replacing them with more detailed ones by SMLP-X, paper: 22x3 total body dims * not sure why paper says 22
+NUM_JAW_JOINTS = 1  # 1x3 total jaw dims
 # Motion-X paper says there
-NUM_HAND_JOINTS = 15 # x2 for each hand -> 30x3 total hand dims
-NUM_JOINTS = NUM_BODY_JOINTS + NUM_HAND_JOINTS * 2 + NUM_JAW_JOINTS # 21 + 30 + 1 = 52
-NUM_FACIAL_EXPRESSION_DIMS = 50  # as per Motion-X paper, but why is default 10 in smplx code then?
+NUM_HAND_JOINTS = 15  # x2 for each hand -> 30x3 total hand dims
+NUM_JOINTS = NUM_BODY_JOINTS + NUM_HAND_JOINTS * 2 + NUM_JAW_JOINTS  # 21 + 30 + 1 = 52
+NUM_FACIAL_EXPRESSION_DIMS = (
+    50  # as per Motion-X paper, but why is default 10 in smplx code then?
+)
 FACE_SHAPE_DIMS = 100
-BODY_SHAPE_DIMS = 10 # betas
+BODY_SHAPE_DIMS = 10  # betas
 ROOT_DIMS = 3
-TRANS_DIMS = 3 # same as root, no?
+TRANS_DIMS = 3  # same as root, no?
 
 pose_type_to_dims = {
     "pose_body": NUM_BODY_JOINTS * 3,
-    "pose_hand": NUM_HAND_JOINTS * 2 * 3, # both hands
+    "pose_hand": NUM_HAND_JOINTS * 2 * 3,  # both hands
     "pose_jaw": NUM_JAW_JOINTS * 3,
     "face_expr": NUM_FACIAL_EXPRESSION_DIMS * 1,  # double check
     "face_shape": FACE_SHAPE_DIMS * 1,  # double check
@@ -62,6 +70,7 @@ pose_type_to_dims = {
     "trans": TRANS_DIMS * 1,
 }
 
+
 def names_to_arrays(root_dir, names, drop_shapes=True):
     all_arrays = []
     for name in names:
@@ -73,12 +82,14 @@ def names_to_arrays(root_dir, names, drop_shapes=True):
         all_arrays.append(array)
     return all_arrays
 
+
 def get_seq_names(file_path):
     with open(file_path, "r") as f:
         names = f.readlines()
     names = [name.strip() for name in names]
     return names
 
+
 def get_data_path(dataset_dir: str, seq: str, file: str) -> str:
     # MY_REPO/face_motion_data/smplx_322/GRAB/s1/airplane_fly_1.npy
     top_dir = MOCAP_FACE_DIR if dataset_dir.lower() in MOCAP_DATASETS else MOTION_DIR
@@ -88,11 +99,16 @@ def get_data_path(dataset_dir: str, seq: str, file: str) -> str:
 
 def get_label_paths(dataset_dir: str, seq: str, file: str) -> Dict[str, str]:
     # MY_REPO/MotionDiffuse/face_texts/GRAB/s1/airplane_fly_1.txt
-    action_path = f"{os.path.join(MY_REPO, ACTION_LABEL_DIR, dataset_dir, seq, file)}.txt"
-    emotion_path = f"{os.path.join(MY_REPO, EMOTION_LABEL_DIR, dataset_dir, seq, file)}.txt"
+    action_path = (
+        f"{os.path.join(MY_REPO, ACTION_LABEL_DIR, dataset_dir, seq, file)}.txt"
+    )
+    emotion_path = (
+        f"{os.path.join(MY_REPO, EMOTION_LABEL_DIR, dataset_dir, seq, file)}.txt"
+    )
     paths = {"action": action_path, "emotion": emotion_path}
     return paths
 
+
 def load_data_as_dict(dataset_dir: str, seq: str, file: str) -> Dict[str, Tensor]:
     path = get_data_path(dataset_dir, seq, file)
     motion = np.load(path)
@@ -108,7 +124,10 @@ def load_data_as_dict(dataset_dir: str, seq: str, file: str) -> Dict[str, Tensor
         "betas": motion[:, 312:],  # controls the body shape. Body shape is static
     }
 
-def motion_arr_to_dict(motion_arr: ArrayLike, shapes_dropped=False) -> Dict[str, Tensor]:
+
+def motion_arr_to_dict(
+    motion_arr: ArrayLike, shapes_dropped=False
+) -> Dict[str, Tensor]:
     # TODO (elmc): why did I need to convert to tensor again???
     motion_arr = torch.tensor(motion_arr).float()
     motion_dict = {
@@ -119,30 +138,40 @@ def motion_arr_to_dict(motion_arr: ArrayLike, shapes_dropped=False) -> Dict[str,
         "face_expr": motion_arr[:, 159 : 159 + 50],  # controls the face expression
     }
     if not shapes_dropped:
-        motion_dict["face_shape"] = motion_arr[:, 209 : 209 + 100] # controls the face shape
-        motion_dict["trans"] = motion_arr[:, 309 : 309 + 3] # controls the global body position
-        motion_dict["betas"] = motion_arr[:, 312:] # controls the body shape. Body shape is static
+        motion_dict["face_shape"] = motion_arr[
+            :, 209 : 209 + 100
+        ]  # controls the face shape
+        motion_dict["trans"] = motion_arr[
+            :, 309 : 309 + 3
+        ]  # controls the global body position
+        motion_dict["betas"] = motion_arr[
+            :, 312:
+        ]  # controls the body shape. Body shape is static
     else:
-        motion_dict["trans"] = motion_arr[:, 209:] # controls the global body position
-    
+        motion_dict["trans"] = motion_arr[:, 209:]  # controls the global body position
+
     return motion_dict
-        
+
 
 def drop_shapes_from_motion_arr(motion_arr: ArrayLike) -> ArrayLike:
     if isinstance(motion_arr, torch.Tensor):
         new_motion_arr = motion_arr.numpy()
-    
+
     # Slice the array to exclude 'face_shape' and 'betas'
-    new_motion_arr = np.concatenate((motion_arr[:, :209], motion_arr[:, 309:312]), axis=1)
-    
+    new_motion_arr = np.concatenate(
+        (motion_arr[:, :209], motion_arr[:, 309:312]), axis=1
+    )
+
     return new_motion_arr
 
+
 def load_label_from_file(file_path: str) -> str:
     with open(file_path, "r") as file:
         # Read the contents of the file into a string
         label = file.read()
     return label
 
+
 def load_label(dataset_dir: str, seq: str, file_path: str) -> Dict[str, str]:
     paths = get_label_paths(dataset_dir, seq, file_path)
     action_path, emotion_path = paths["action"], paths["emotion"]
@@ -163,28 +192,31 @@ def label_code(full_label):
     # airplane -> air
     return full_label[:3]
 
+
 def get_seq_type(motion_label_dir, file_name):
     # e.g. s5/airplane_fly_1 -> airplane fly (motion label)
     seq_type_path = pjoin(motion_label_dir, f"{file_name}.txt")
-    with open(seq_type_path, 'r') as f:
+    with open(seq_type_path, "r") as f:
         seq_type = f.readline().strip()
     return seq_type
 
+
 def calc_mean_stddev_pose(arrays):
     # all_arrays = []
     # for file_path in file_list:
     #     # Load each NumPy array and add it to the list
     #     array = np.load(file_path)
     #     all_arrays.append(array)
-    
+
     # Concatenate all arrays along the first axis (stacking them on top of each other)
     concatenated_arrays = np.concatenate(arrays, axis=0)
     # Calculate the mean and standard deviation across all arrays
     mean = np.mean(concatenated_arrays, axis=0)
     stddev = np.std(concatenated_arrays, axis=0)
-    
+
     return mean, stddev
 
+
 def get_info_from_file(file_path, emotions_label_dir, motion_label_dir):
     # train_names = get_seq_names(pjoin(data_dir, "train.txt"))
     names = get_seq_names(file_path)
@@ -220,9 +252,12 @@ def get_info_from_file(file_path, emotions_label_dir, motion_label_dir):
         "n_seq": n_seq,
         "code_to_label": code_to_label,
     }
-    return info_dict 
+    return info_dict
 
-def to_smplx_dict(motion_dict: Dict[str, Tensor], timestep_range: Optional[Tuple[int, int]] = None) -> Dict[str, Tensor]:
+
+def to_smplx_dict(
+    motion_dict: Dict[str, Tensor], timestep_range: Optional[Tuple[int, int]] = None
+) -> Dict[str, Tensor]:
     if timestep_range is None:
         # get all timesteps
         timestep_range = (0, len(motion_dict["pose_body"]))
@@ -230,31 +265,51 @@ def to_smplx_dict(motion_dict: Dict[str, Tensor], timestep_range: Optional[Tuple
         "global_orient": motion_dict["root_orient"][
             timestep_range[0] : timestep_range[1]
         ],  # controls the global root orientation
-        "body_pose": motion_dict["pose_body"][timestep_range[0] : timestep_range[1]],  # controls the body
-        "left_hand_pose": motion_dict["pose_hand"][timestep_range[0] : timestep_range[1]][
-            :, : NUM_HAND_JOINTS * 3
-        ],  # controls the finger articulation
-        "right_hand_pose": motion_dict["pose_hand"][timestep_range[0] : timestep_range[1]][:, NUM_HAND_JOINTS * 3 :],
-        "expression": motion_dict["face_expr"][timestep_range[0] : timestep_range[1]],  # controls the face expression
-        "jaw_pose": motion_dict["pose_jaw"][timestep_range[0] : timestep_range[1]],  #  controls the jaw pose
+        "body_pose": motion_dict["pose_body"][
+            timestep_range[0] : timestep_range[1]
+        ],  # controls the body
+        "left_hand_pose": motion_dict["pose_hand"][
+            timestep_range[0] : timestep_range[1]
+        ][:, : NUM_HAND_JOINTS * 3],  # controls the finger articulation
+        "right_hand_pose": motion_dict["pose_hand"][
+            timestep_range[0] : timestep_range[1]
+        ][:, NUM_HAND_JOINTS * 3 :],
+        "expression": motion_dict["face_expr"][
+            timestep_range[0] : timestep_range[1]
+        ],  # controls the face expression
+        "jaw_pose": motion_dict["pose_jaw"][
+            timestep_range[0] : timestep_range[1]
+        ],  #  controls the jaw pose
         # 'face_shape': motion_dict['face_shape'][timestep],  # controls the face shape, drop since we don't care to train on this
-        "transl": motion_dict["trans"][timestep_range[0] : timestep_range[1]],  # controls the global body position
+        "transl": motion_dict["trans"][
+            timestep_range[0] : timestep_range[1]
+        ],  # controls the global body position
         # "betas": motion["betas"][
         #     timestep_range[0] : timestep_range[1]
         # ],  # controls the body shape. Body shape is static, drop since we don't care to train on this
     }
     return smplx_params
 
+
 def smplx_dict_to_array(smplx_dict):
     # convert smplx dict to array
     # list keys to ensure known order when iterating over dict
-    keys = ["global_orient", "body_pose", "left_hand_pose", "right_hand_pose", "expression", "jaw_pose", "transl"]
+    keys = [
+        "global_orient",
+        "body_pose",
+        "left_hand_pose",
+        "right_hand_pose",
+        "expression",
+        "jaw_pose",
+        "transl",
+    ]
     smplx_array = []
     for key in keys:
         smplx_array.append(smplx_dict[key])
     smplx_array = torch.cat(smplx_array, dim=1)
     return smplx_array
 
+
 def save_gif(gif_path, gif_frames, duration=0.01):
     if gif_frames:
         print(f"Saving GIF with {len(gif_frames)} frames to {gif_path}")
@@ -262,6 +317,7 @@ def save_gif(gif_path, gif_frames, duration=0.01):
     else:
         print("No frames to save.")
 
+
 # based on https://github.com/vchoutas/smplx/blob/main/examples/demo.py
 def render_meshes(output, should_save_gif=False, gif_path=None):
     should_display = not should_save_gif
@@ -294,7 +350,9 @@ def render_meshes(output, should_save_gif=False, gif_path=None):
             plot_joints = False
             if plotting_module == "pyrender":
                 vertex_colors = np.ones([vertices.shape[0], 4]) * [0.3, 0.3, 0.3, 0.8]
-                tri_mesh = trimesh.Trimesh(vertices, model.faces, vertex_colors=vertex_colors)
+                tri_mesh = trimesh.Trimesh(
+                    vertices, model.faces, vertex_colors=vertex_colors
+                )
 
                 # Apply rotation
                 tri_mesh.apply_transform(rot)
@@ -322,7 +380,7 @@ def render_meshes(output, should_save_gif=False, gif_path=None):
                 cam_pose = np.array(
                     [
                         [1.0, 0, 0, center[0]],
-                        [0, 1.0, 0, center[1]-1.0],
+                        [0, 1.0, 0, center[1] - 1.0],
                         [0, 0, 1.0, center[2] + distance + 0.5],
                         [0, 0, 0, 1],
                     ]
@@ -332,12 +390,14 @@ def render_meshes(output, should_save_gif=False, gif_path=None):
                 angle = np.radians(90)
                 cos_angle = np.cos(angle)
                 sin_angle = np.sin(angle)
-                rot_x = np.array([
-                    [1, 0,        0,         0],
-                    [0, cos_angle, -sin_angle, 0],
-                    [0, sin_angle, cos_angle,  0],
-                    [0, 0,        0,         1]
-                ])
+                rot_x = np.array(
+                    [
+                        [1, 0, 0, 0],
+                        [0, cos_angle, -sin_angle, 0],
+                        [0, sin_angle, cos_angle, 0],
+                        [0, 0, 0, 1],
+                    ]
+                )
                 cam_pose = np.matmul(cam_pose, rot_x)
                 cam_pose[:3, 3] += np.array([0, -2.5, -3.5])
 
@@ -361,7 +421,9 @@ def render_meshes(output, should_save_gif=False, gif_path=None):
                         scene.remove_node(joints_node)
                     joints_node = scene.add(joints_pcl)
                 if should_save_gif:
-                    r = pyrender.OffscreenRenderer(viewport_width=640, viewport_height=480)
+                    r = pyrender.OffscreenRenderer(
+                        viewport_width=640, viewport_height=480
+                    )
                     color, _ = r.render(scene)
                     gif_frames.append(color)
                     r.delete()  # Free up the resources
@@ -375,11 +437,13 @@ def render_meshes(output, should_save_gif=False, gif_path=None):
     finally:
         save_gif(gif_path, gif_frames)
 
+
 def get_numpy_file_path(prompt, epoch, n_frames):
     # e.g. "airplane_fly_1_1000_60f.npy"
-    prompt_no_spaces = prompt.replace(' ', '_')
+    prompt_no_spaces = prompt.replace(" ", "_")
     return f"{prompt_no_spaces}_{epoch}_{n_frames}f"
 
+
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
 
@@ -401,10 +465,10 @@ if __name__ == "__main__":
     parser.add_argument(
         "-dm",
         "--display_mesh",
-        action='store_true',
+        action="store_true",
         required=False,
         default=False,
-        help="Display mesh if this flag is present"
+        help="Display mesh if this flag is present",
     )
     # for now just specifies file name (with spaces) made by inference
     parser.add_argument(
@@ -435,10 +499,10 @@ if __name__ == "__main__":
     parser.add_argument(
         "-sg",
         "--save_gif",
-        action='store_true',
+        action="store_true",
         required=False,
         default=False,
-        help="Save gif if this flag is present"
+        help="Save gif if this flag is present",
     )
     # add which_epoch
     parser.add_argument(
@@ -453,26 +517,30 @@ if __name__ == "__main__":
     prompt = args.prompt
     is_inference = len(prompt) > 0
     if args.seq_file != "" and args.prompt != "":
-        log.error("cannot provide both prompt and seq_file; if trying to verify model inference, use --prompt, otherwise specify numpy --seq_file name to display")
+        log.error(
+            "cannot provide both prompt and seq_file; if trying to verify model inference, use --prompt, otherwise specify numpy --seq_file name to display"
+        )
         exit(1)
     elif args.seq_file == "" and args.prompt == "":
-        log.error("must provide either prompt or seq_file; if trying to verify model inference, use --prompt, otherwise specify numpy --seq_file name to display")
+        log.error(
+            "must provide either prompt or seq_file; if trying to verify model inference, use --prompt, otherwise specify numpy --seq_file name to display"
+        )
         exit(1)
     if not is_inference:
         name = args.seq_file
-        data_root = './data/GRAB'
-        motion_dir = pjoin(data_root, 'joints')
+        data_root = "./data/GRAB"
+        motion_dir = pjoin(data_root, "joints")
     else:
-        log.info(f"converting prompt into file name")
+        log.info("converting prompt into file name")
         name = get_numpy_file_path(prompt, args.which_epoch, args.max_t - args.min_t)
         model_type = args.model_path
-        motion_dir = pjoin(model_type, 'outputs')
-    motion_path = pjoin(motion_dir, name + '.npy')
+        motion_dir = pjoin(model_type, "outputs")
+    motion_path = pjoin(motion_dir, name + ".npy")
     log.info(f"loading motion from {motion_path}")
     motion_arr = np.load(motion_path)
     t = 999
-    mean_path = '/work3/s222376/MotionDiffuse2/text2motion/checkpoints/grab/md_fulem_2g_excl_196_seed42/meta/mean.npy'
-    std_path = '/work3/s222376/MotionDiffuse2/text2motion/checkpoints/grab/md_fulem_2g_excl_196_seed42/meta/std.npy'
+    mean_path = "/work3/s222376/MotionDiffuse2/text2motion/checkpoints/grab/md_fulem_2g_excl_196_seed42/meta/mean.npy"
+    std_path = "/work3/s222376/MotionDiffuse2/text2motion/checkpoints/grab/md_fulem_2g_excl_196_seed42/meta/std.npy"
     mean = np.load(mean_path)
     std = np.load(std_path)
     # do range skipping by 100
@@ -481,7 +549,9 @@ if __name__ == "__main__":
     for t in list_:
         name = f"sample_tensor([{t}])"
         # breakpoint()
-        motion_arr = np.load(f"/work3/s222376/MotionDiffuse2/text2motion/generation_samples/{name}.npy")
+        motion_arr = np.load(
+            f"/work3/s222376/MotionDiffuse2/text2motion/generation_samples/{name}.npy"
+        )
         motion_arr = np.squeeze(motion_arr)
 
         motion_arr = motion_arr * std + mean
@@ -491,7 +561,9 @@ if __name__ == "__main__":
             # directly get smplx dimensionality by dropping body and face shape data
             print("warning, dropping body and face shape data")
             motion_arr = drop_shapes_from_motion_arr(motion_arr)
-            assert motion_arr.shape[1] == 212, f"expected 212 dims, got {motion_arr.shape[1]}"
+            assert (
+                motion_arr.shape[1] == 212
+            ), f"expected 212 dims, got {motion_arr.shape[1]}"
 
         # our MotionDiffuse predicts motion data that doesn't include face and body shape
         motion_dict = motion_arr_to_dict(motion_arr, shapes_dropped=True)
@@ -522,23 +594,23 @@ if __name__ == "__main__":
         log.info(f"TOTAL SMPLX dims: {tot_smplx_dims}\n")
 
         if not is_inference:
-            action_label_path = pjoin(data_root, 'texts', name + '.txt')
+            action_label_path = pjoin(data_root, "texts", name + ".txt")
             action_label = load_label_from_file(action_label_path)
-            emotion_label_path = pjoin(data_root, 'face_texts', name + '.txt')
+            emotion_label_path = pjoin(data_root, "face_texts", name + ".txt")
             emotion_label = load_label_from_file(emotion_label_path)
             log.info(f"action: {action_label}")
             log.info(f"emotion: {emotion_label}")
 
         if is_inference:
-            emotion_label = args.prompt.split(' ')[0]
-        
+            emotion_label = args.prompt.split(" ")[0]
+
         if args.display_mesh:
             model_folder = os.path.join(MY_REPO, MODELS_DIR, "smplx")
             batch_size = max_t - min_t
             log.info(f"calculating mesh with batch size {batch_size}")
             model = smplx.SMPLX(
                 model_folder,
-                use_pca=False, # our joints are not in pca space
+                use_pca=False,  # our joints are not in pca space
                 num_expression_coeffs=NUM_FACIAL_EXPRESSION_DIMS,
                 batch_size=batch_size,
             )
@@ -546,7 +618,9 @@ if __name__ == "__main__":
             log.info(f"output size {output.vertices.shape}")
             log.info(f"output size {output.joints.shape}")
             log.info("rendering mesh")
-            model_name = args.model_path.split('/')[-1] if args.model_path else "ground_truth"
+            model_name = (
+                args.model_path.split("/")[-1] if args.model_path else "ground_truth"
+            )
             gif_path = f"gifs/{model_name}/{name}_{emotion_label}.gif"
             render_meshes(output, gif_path=gif_path, should_save_gif=args.save_gif)
             log.warning(

text2motion/tools/inference.py

@@ -14,6 +14,24 @@ from utils.plot_script import *
 from utils.utils import *
 from utils.word_vectorizer import POS_enumerator
 
+# def plot_t2m(opt, data, result_path, caption):
+#     joint = recover_from_ric(torch.from_numpy(data).float(), opt.joints_num).numpy()
+#     # joint = motion_temporal_filter(joint, sigma=1)
+#     plot_3d_motion(result_path, paramUtil.t2m_kinematic_chain, joint, title=caption, fps=20)
+
+
+# def process(trainer, opt, device, mean, std, text, motion_length, result_path):
+
+#     result_dict = {}
+#     with torch.no_grad():
+#         if motion_length != -1:
+#             caption = [text]
+#             m_lens = torch.LongTensor([motion_length]).to(device)
+#             pred_motions = trainer.generate(caption, m_lens, opt.dim_pose)
+#             motion = pred_motions[0].cpu().numpy()
+#             motion = motion * std + mean
+#             title = text + " #%d" % motion.shape[0]
+#             plot_t2m(opt, motion, result_path, title
 
 def plot_t2m(data, result_path, npy_path, caption, joints_n):
     joint = recover_from_ric(torch.from_numpy(data).float(), joints_n).numpy()