File size: 6,255 Bytes
4fa8bcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import pandas as pd
from model_training.model_torch import EncoderDecoder
from tqdm import tqdm
from torch.utils.data import Dataset, DataLoader
from torchmetrics.functional.image import structural_similarity_index_measure
import numpy as np
from config import normalize_input, get_input_fields
from utils.path_utils import resolve_path
import os

class PhysicsTrajectoryDataset(Dataset):
    def __init__(self, filepath, physics_type):
        self.df = pd.read_pickle(filepath)
        self.physics_type = physics_type
        
        # Detect the number of time steps from the first sample
        sample_trajectory = self.df.iloc[0]['trajectory']
        self.time_steps = len(sample_trajectory)
        
        # Sanity check
        if not all(len(traj) == self.time_steps for traj in self.df['trajectory']):
            raise ValueError("Not all trajectories have the same number of time steps!")

    def __len__(self):
        return len(self.df)

    def __getitem__(self, idx):
        row = self.df.iloc[idx]
        fields = get_input_fields(self.physics_type)
        inputs = torch.tensor(normalize_input(self.physics_type, *[row[f] for f in fields]), dtype=torch.float32)
        
        trajectory_np = np.array(row['trajectory'])
        trajectory = torch.from_numpy(trajectory_np).float()
        

        return inputs, trajectory

def compute_ssim(pred, target, device):
    """
    Computes SSIM between predicted and target tensors.
    If the input is 1D (H=1), attempts to squarify.
    Falls back to MSE-only if squarify fails.
    """
    B, C, H, W = pred.shape
    try:
        if H == 1:
            # 1D sequence → squarify to [B, C, S, S]
            pred_sq = squarify_1d_sequence(pred)
            target_sq = squarify_1d_sequence(target)
            return structural_similarity_index_measure(pred_sq, target_sq, data_range=1.0)
        else:
            # Already 2D image-like
            return structural_similarity_index_measure(pred, target, data_range=1.0)
    except ValueError:
        # Fallback: SSIM not computable
        return torch.tensor(0.0, device=device)
      
def squarify_1d_sequence(x):
    """Reshape 1D sequence into square 2D frames for SSIM."""
    B, C, _, T = x.shape
    S = int(T ** 0.5)
    if S * S != T:
        raise ValueError(f"T={T} must be a perfect square for squarify, got {T}.")
    return x.view(B, C, S, S)

def bce_dot_loss(pred, target, device):
    """
    Compute binary cross entropy loss between predicted and target dot-maps.
    Assumes pred and target have shape [B, T, H, W] with values in [0, 1].
    """
    if pred.shape != target.shape:
        raise ValueError(f"Shape mismatch: pred {pred.shape} vs target {target.shape}")

    # Compute per-pixel weights: 6.0 where target=1, 1.0 where target=0
    weight = (target * 10.0 + 1.0)  # weight = 610 where dot is, 1 elsewhere
    loss = nn.functional.binary_cross_entropy(pred, target, weight=weight, reduction='mean')
    return loss
  
def train_model(
    physics_type, 
    hidden_size=128, 
    lr=0.002, 
    epochs=20, 
    early_stopping=False, 
    patience=5, 
    batch_size=256, 
    clip_grad=1.0
):
    dataset_path = resolve_path(f"{physics_type}_data.pkl")
    model_path = resolve_path(f"{physics_type}_model.pth", write_mode=True)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

    # Dataset and DataLoader
    dataset = PhysicsTrajectoryDataset(dataset_path, physics_type)
    dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True, num_workers=0)
    
    # Inspect shape of sample
    sample_input, sample_target = dataset[0]
    timesteps, coord_dims = sample_target.shape
    assert coord_dims == 2, "Expected target shape [T, 2] for (x, y) coordinates"

    model = EncoderDecoder(
        input_dim=sample_input.shape[0],
        hidden_size=hidden_size,
        output_seq_len=timesteps,
        output_shape=None  # Not used in coordinate mode
    ).to(device)

    optimizer = torch.optim.Adam(model.parameters(), lr=lr)
    scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=3)

    loss_fn = nn.MSELoss()
    best_loss = float('inf')
    wait = 0
    losses = []

    for epoch in range(epochs):
        model.train()
        total_loss = 0
        progress_bar = tqdm(dataloader, desc=f"Epoch {epoch+1}/{epochs}", leave=False)

        for batch_inputs, batch_targets in progress_bar:
            batch_inputs = batch_inputs.to(device)            # [B, F]
            batch_targets = batch_targets.to(device)          # [B, T, 2]

            optimizer.zero_grad()
            outputs = model(batch_inputs)                     # [B, T, 2]
            loss = loss_fn(outputs, batch_targets)

            loss.backward()
            if clip_grad:
                torch.nn.utils.clip_grad_norm_(model.parameters(), clip_grad)
            optimizer.step()

            total_loss += loss.item()

        avg_loss = total_loss / len(dataloader)
        losses.append(avg_loss)
        print(f"Epoch {epoch+1}/{epochs}, Loss: {avg_loss:.6f}")
        scheduler.step(avg_loss)

        # Debug: check one sample trajectory
        # if epoch % 5 == 0:
        #     with torch.no_grad():
        #         coords = outputs[0].detach().cpu().numpy()
        #         print(f"🔍 Predicted trajectory sample (epoch {epoch+1}):")
        #         for t, (x, y) in enumerate(coords[:5]):
        #             print(f"t={t}: (x={x:.3f}, y={y:.3f})")

        # Early Stopping Logic
        if early_stopping:
            if avg_loss < best_loss - 1e-6:
                best_loss = avg_loss
                wait = 0
            else:
                wait += 1
                if wait >= patience:
                    print(f"⏹️ Early stopping at epoch {epoch+1}")
                    break

    # Save model
    torch.save({
        'model_state': model.state_dict(),
        'input_dim': sample_input.shape[0],
        'output_seq_len': timesteps,
    }, model_path)
    print(f"✅ Model saved to {model_path}")
    return f"✅ Trained and saved {physics_type} model to {model_path}", losses

if __name__ == "__main__":
    train_model("ball_motion")