File size: 5,097 Bytes
bc45d7d 7279c87 bc45d7d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 | import os
import glob
import argparse
import json
import numpy as np
import pandas as pd
from tqdm import tqdm
import torch
import torch.nn as nn
from torch.utils import data
from generate_keypoints import process_video
from models import Transformer
from configs import TransformerConfig
from utils import load_json, load_label_map
import shutil
parser = argparse.ArgumentParser(description="Evaluate function")
parser.add_argument("--data_dir", required=True, help="data directory")
args = parser.parse_args()
class KeypointsDataset(data.Dataset):
def __init__(
self,
keypoints_dir,
max_frame_len=200,
frame_length=1080,
frame_width=1920,
):
self.files = sorted(glob.glob(os.path.join(keypoints_dir, "*.json")))
self.max_frame_len = max_frame_len
self.frame_length = frame_length
self.frame_width = frame_width
def interpolate(self, arr):
arr_x = arr[:, :, 0]
arr_x = pd.DataFrame(arr_x)
arr_x = arr_x.interpolate(method="linear", limit_direction="both").to_numpy()
arr_y = arr[:, :, 1]
arr_y = pd.DataFrame(arr_y)
arr_y = arr_y.interpolate(method="linear", limit_direction="both").to_numpy()
if np.count_nonzero(~np.isnan(arr_x)) == 0:
arr_x = np.zeros(arr_x.shape)
if np.count_nonzero(~np.isnan(arr_y)) == 0:
arr_y = np.zeros(arr_y.shape)
arr_x = arr_x * self.frame_width
arr_y = arr_y * self.frame_length
return np.stack([arr_x, arr_y], axis=-1)
def combine_xy(self, x, y):
x, y = np.array(x), np.array(y)
_, length = x.shape
x = x.reshape((-1, length, 1))
y = y.reshape((-1, length, 1))
return np.concatenate((x, y), -1).astype(np.float32)
def __getitem__(self, idx):
file_path = self.files[idx]
row = pd.read_json(file_path, typ="series")
label = row.label
label = "".join([i for i in label if i.isalpha()]).lower()
pose = self.combine_xy(row.pose_x, row.pose_y)
h1 = self.combine_xy(row.hand1_x, row.hand1_y)
h2 = self.combine_xy(row.hand2_x, row.hand2_y)
pose = self.interpolate(pose)
h1 = self.interpolate(h1)
h2 = self.interpolate(h2)
df = pd.DataFrame.from_dict(
{
"uid": row.uid,
"pose": pose.tolist(),
"hand1": h1.tolist(),
"hand2": h2.tolist(),
"label": label,
}
)
pose = (
np.array(list(map(np.array, df.pose.values)))
.reshape(-1, 50)
.astype(np.float32)
)
h1 = (
np.array(list(map(np.array, df.hand1.values)))
.reshape(-1, 42)
.astype(np.float32)
)
h2 = (
np.array(list(map(np.array, df.hand2.values)))
.reshape(-1, 42)
.astype(np.float32)
)
final_data = np.concatenate((pose, h1, h2), -1)
final_data = np.pad(
final_data,
((0, self.max_frame_len - final_data.shape[0]), (0, 0)),
"constant",
)
return {
"uid": row.uid,
"data": torch.FloatTensor(final_data),
}
def __len__(self):
return len(self.files)
@torch.no_grad()
def inference(dataloader, model, device, label_map):
model.eval()
predictions = []
for batch in tqdm(dataloader, desc="Eval"):
input_data = batch["data"].to(device)
output = model(input_data).detach().cpu()
output = torch.argmax(torch.softmax(output, dim=-1), dim=-1).numpy()
predictions.append({"uid": batch["uid"][0], "predicted_label": label_map[output[0]]})
return predictions
video_paths = glob.glob(os.path.join(args.data_dir, "*"))
save_dir = "keypoints_dir"
if os.path.isdir(save_dir):
shutil.rmtree(save_dir)
os.mkdir(save_dir)
for path in tqdm(video_paths, desc="Processing Videos"):
process_video(path, save_dir)
label_map = load_label_map("include")
dataset = KeypointsDataset(
keypoints_dir=save_dir,
max_frame_len=169,
)
dataloader = data.DataLoader(
dataset,
batch_size=1,
shuffle=False,
num_workers=4,
pin_memory=True,
)
label_map = dict(zip(label_map.values(), label_map.keys()))
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
config = TransformerConfig(size="large", max_position_embeddings=256)
model = Transformer(config=config, n_classes=263)
model = model.to(device)
pretrained_model_name = "include_no_cnn_transformer_large.pth"
pretrained_model_links = load_json("pretrained_links.json")
if not os.path.isfile(pretrained_model_name):
link = pretrained_model_links[pretrained_model_name]
torch.hub.download_url_to_file(link, pretrained_model_name, progress=True)
ckpt = torch.load(pretrained_model_name, weights_only=False)
model.load_state_dict(ckpt["model"])
print("### Model loaded ###")
preds = inference(dataloader, model, device, label_map)
print(json.dumps(preds, indent=2))
|