File size: 7,018 Bytes
f17ae24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
171
172
173
174
175
176
177
178
179
180
import os
import torch
import torch.nn.functional as F
import yaml
import numpy as np
import matplotlib.pyplot as plt
from tqdm import tqdm
import argparse
import sys

# Add project root to path
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), "../..")))

from wm.model.interface import get_dynamics_class
from wm.dataset.dataset import RoboticsDatasetWrapper
from torch.utils.data import DataLoader

def load_checkpoint(model, path, device):
    print(f"Loading checkpoint: {path}")
    checkpoint = torch.load(path, map_location=device, weights_only=False)
    state_dict = checkpoint['model_state_dict']
    
    # Filter out scheduler buffers that might cause size mismatches
    scheduler_buffers = [
        'scheduler.sigmas', 
        'scheduler.timesteps', 
        'scheduler.linear_timesteps_weights'
    ]
    
    new_state_dict = {}
    for k, v in state_dict.items():
        # Remove "module." prefix from DDP
        name = k[7:] if k.startswith('module.') else k
        if name not in scheduler_buffers:
            new_state_dict[name] = v
            
    model.load_state_dict(new_state_dict, strict=False)
    return checkpoint.get('step', 0)

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=str, required=True, help="Path to the yaml config file")
    parser.add_argument("--ckpt_dir", type=str, default="checkpoints/lang_table_fulltraj_dit_v1", help="Directory containing checkpoints")
    parser.add_argument("--results_dir", type=str, default="results", help="Directory to save plots")
    parser.add_argument("--num_samples", type=int, default=-1, help="Number of validation samples to evaluate (-1 for all)")
    parser.add_argument("--batch_size", type=int, default=64, help="Batch size for evaluation")
    parser.add_argument("--inference_steps", type=int, default=50, help="Number of denoising steps")
    parser.add_argument("--noise_level", type=float, default=0.0, help="Noise level (t0) for the first frame (0.0 to 1.0)")
    parser.add_argument("--label", type=str, default="", help="Label for the evaluation run")
    args = parser.parse_args()

    os.makedirs(args.results_dir, exist_ok=True)
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    # If label is not provided, generate one
    if not args.label:
        args.label = f"{args.inference_steps}steps"
        if args.noise_level > 0:
            args.label += f"_noise{args.noise_level}"

    # 1. Load Config
    with open(args.config, 'r') as f:
        config = yaml.safe_load(f)

    # 2. Setup Dataset
    print("Initializing dataset...")
    full_dataset = RoboticsDatasetWrapper.get_dataset(config['dataset']['name'])
    
    # Pick validation samples from the end as in training split
    num_train = int(len(full_dataset) * (config['dataset']['train_test_split'] / (config['dataset']['train_test_split'] + 1)))
    val_indices = list(range(num_train, len(full_dataset)))
    
    # Limit number of samples if requested
    if args.num_samples > 0:
        import random
        random.seed(42)
        random.shuffle(val_indices)
        val_indices = val_indices[:args.num_samples]
    
    print(f"Evaluating on {len(val_indices)} samples.")
    val_subset = torch.utils.data.Subset(full_dataset, val_indices)
    val_loader = DataLoader(val_subset, batch_size=args.batch_size, shuffle=False, num_workers=8)

    # 3. Initialize Model
    print("Initializing model...")
    dynamics_class = get_dynamics_class(config['dynamics_class'])
    model = dynamics_class(config['model_name'], config['model_config'])
    model.to(device)
    model.eval()

    # 4. Find Checkpoints
    ckpt_files = [f for f in os.listdir(args.ckpt_dir) if f.endswith(".pt") and "checkpoint_" in f]
    
    def get_step(f):
        try:
            return int(f.split('_')[1].split('.')[0])
        except:
            return -1
            
    ckpt_files.sort(key=get_step)
    
    if not ckpt_files:
        print(f"No checkpoints found in {args.ckpt_dir}")
        return

    results = []

    # 5. Evaluate each checkpoint
    for ckpt_file in ckpt_files:
        ckpt_path = os.path.join(args.ckpt_dir, ckpt_file)
        step = load_checkpoint(model, ckpt_path, device)
        
        all_mse = []
        
        with torch.no_grad():
            for batch in tqdm(val_loader, desc=f"Eval Step {step}"):
                obs = batch['obs'].to(device) # [B, T, C, H, W]
                action = batch['action'].to(device) # [B, T, A]
                
                # First frame: [B, H, W, 3] -> permute from [B, C, H, W]
                o_0 = obs[:, 0].permute(0, 2, 3, 1).contiguous()
                
                # Generate rollout
                pred_video = model.generate(
                    o_0, action, 
                    num_inference_steps=args.inference_steps,
                    noise_level=args.noise_level
                )
                
                # Ground truth: [B, T, C, H, W] -> [B, T, H, W, 3]
                gt_video = obs.permute(0, 1, 3, 4, 2).contiguous()
                
                # Calculate MSE per sample (average over time and space)
                mse = (pred_video - gt_video)**2
                mse_per_sample = mse.mean(dim=(1, 2, 3, 4)) # [B]
                all_mse.append(mse_per_sample.cpu().numpy())
                
        # Calculate statistics across all samples
        all_mse_np = np.concatenate(all_mse, axis=0)
        mean_mse = np.mean(all_mse_np)
        p25 = np.percentile(all_mse_np, 25)
        p75 = np.percentile(all_mse_np, 75)
        
        print(f"Step {step} | Mean MSE: {mean_mse:.6f} | P25: {p25:.6f} | P75: {p75:.6f}")
        results.append({
            'step': step,
            'mean': mean_mse,
            'p25': p25,
            'p75': p75
        })

    # 6. Plot and Save
    results.sort(key=lambda x: x['step'])
    steps = [r['step'] for r in results]
    means = [r['mean'] for r in results]
    p25s = [r['p25'] for r in results]
    p75s = [r['p75'] for r in results]

    plt.figure(figsize=(10, 6))
    plt.plot(steps, means, marker='o', linestyle='-', color='b', label='Mean MSE')
    plt.fill_between(steps, p25s, p75s, color='b', alpha=0.2, label='25th-75th Percentile')
    plt.title(f"Evaluation MSE Curve - {config['dataset']['name']} ({args.label})")
    plt.xlabel("Training Steps")
    plt.ylabel("Mean RGB MSE (Full Trajectory)")
    plt.legend()
    plt.grid(True)
    
    plot_path = os.path.join(args.results_dir, f"mse_curve_{config['dataset']['name']}_{args.label}.png")
    plt.savefig(plot_path)
    print(f"Saved MSE curve to {plot_path}")
    
    # Save raw results to a text file
    with open(os.path.join(args.results_dir, f"mse_results_{config['dataset']['name']}_{args.label}.txt"), "w") as f:
        f.write("Step,Mean_MSE,P25,P75\n")
        for r in results:
            f.write(f"{r['step']},{r['mean']:.8f},{r['p25']:.8f},{r['p75']:.8f}\n")

if __name__ == "__main__":
    main()