Spaces:
Running on Zero
Running on Zero
File size: 9,242 Bytes
becf13a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 | """
Multi-DataLoader for combining multiple dataloaders with weighted batch sizes.
This module provides a way to train on multiple datasets simultaneously by
creating separate dataloaders for each dataset and combining their batches.
Batch sizes directly control the sampling weight for each dataset.
"""
import torch
from typing import List, Optional
from torch.utils.data import DataLoader
def collate_objaverse_batch(batch):
"""Collate function for Objaverse datasets (SimplifiedViewpointDataset).
Converts tuple format to dict format compatible with CombinedDataset.
Args:
batch: List of tuples (image, rotation, translation, relative_rotation, spherical_angular)
Returns:
Dict with batched tensors and dataset_type
"""
images = torch.stack([item[0] for item in batch])
rotations = torch.stack([item[1] for item in batch])
translations = torch.stack([item[2] for item in batch])
relative_rotations = torch.stack([item[3] for item in batch])
spherical_angular = torch.stack([item[4] for item in batch])
return {
'image': images,
'rotation': rotations,
'translation': translations,
'relative_rotation': relative_rotations,
'spherical_angular': spherical_angular,
'dataset_type': torch.tensor([0] * len(batch), dtype=torch.long), # 0 = objaverse
}
def collate_compass_batch(batch):
"""Collate function for Compass datasets (CompassDataset).
Converts tuple format to dict format compatible with CombinedDataset.
Args:
batch: List of tuples (image, azimuth, category)
Note: category is ignored (3rd element not used)
Returns:
Dict with batched tensors and dataset_type
"""
images = torch.stack([item[0] for item in batch])
azimuths = torch.stack([item[1] for item in batch])
# Note: item[2] is category but we don't use it
# Build spherical_angular with only azimuth (rest are zeros)
# Format: [sin(az), cos(az), sin(el), cos(el), norm_radius, norm_yaw, norm_pitch]
sin_az = torch.sin(azimuths)
cos_az = torch.cos(azimuths)
zeros = torch.zeros_like(azimuths)
spherical_angular = torch.stack([
sin_az, cos_az, zeros, zeros, zeros, zeros, zeros
], dim=1) # (B, 7)
return {
'image': images,
'spherical_angular': spherical_angular,
'dataset_type': torch.tensor([1] * len(batch), dtype=torch.long), # 1 = compass
}
def combine_batches(batches: List[dict]) -> dict:
"""Combine multiple batched dicts into a single batch.
Args:
batches: List of batch dicts from different dataloaders
Returns:
Combined batch dict with concatenated tensors
"""
if len(batches) == 1:
return batches[0]
combined = {}
# Get all keys
all_keys = set()
for batch in batches:
all_keys.update(batch.keys())
for key in all_keys:
values = [batch[key] for batch in batches if key in batch]
# Concatenate tensors, extend lists
if torch.is_tensor(values[0]):
combined[key] = torch.cat(values, dim=0)
elif isinstance(values[0], list):
combined[key] = sum(values, []) # Flatten lists
else:
combined[key] = values
return combined
class MultiDataLoader:
"""Combines multiple dataloaders with specified batch sizes.
This dataloader iterates multiple dataloaders in parallel, combining their
batches into a single batch. The epoch length is determined by the longest
dataloader, and shorter dataloaders are restarted when exhausted (cyclic).
Args:
dataloaders: List of DataLoader instances to combine
batch_sizes: List of batch sizes for each dataloader (must sum to total batch size)
collate_fn: Optional collate function to combine batches from different dataloaders.
If None, batches are returned as a list.
Example:
>>> dl1 = DataLoader(dataset1, batch_size=28, num_workers=4, shuffle=True)
>>> dl2 = DataLoader(dataset2, batch_size=4, num_workers=4, shuffle=True)
>>> multi_dl = MultiDataLoader([dl1, dl2], batch_sizes=[28, 4])
>>> for batch in multi_dl:
... # batch contains 32 samples (28 from dl1 + 4 from dl2)
... outputs = model(batch)
"""
def __init__(
self,
dataloaders: List[DataLoader],
batch_sizes: List[int],
collate_fn: Optional[callable] = None,
):
assert len(dataloaders) > 0, "Must provide at least one dataloader"
assert len(dataloaders) == len(batch_sizes), \
f"Number of dataloaders ({len(dataloaders)}) must match number of batch_sizes ({len(batch_sizes)})"
# Verify batch sizes match dataloader configurations
for i, (dl, expected_bs) in enumerate(zip(dataloaders, batch_sizes)):
actual_bs = dl.batch_size
if actual_bs != expected_bs:
print(f"Warning: DataLoader {i} has batch_size={actual_bs} but expected {expected_bs}")
self.dataloaders = dataloaders
self.batch_sizes = batch_sizes
self.collate_fn = collate_fn
# Calculate total batch size
self.total_batch_size = sum(batch_sizes)
# Calculate epoch length (longest dataloader)
self._length = max(len(dl) for dl in dataloaders)
print(f"MultiDataLoader initialized with {len(dataloaders)} dataloaders:")
for i, (dl, bs) in enumerate(zip(dataloaders, batch_sizes)):
print(f" [{i}] batch_size={bs}, length={len(dl)} iterations")
print(f"Total batch size: {self.total_batch_size}")
print(f"Epoch length: {self._length} iterations (determined by longest dataloader)")
def __len__(self) -> int:
"""Return the number of iterations per epoch (length of longest dataloader)."""
return self._length
def __iter__(self):
"""Iterate through all dataloaders in parallel, combining their batches."""
# Create iterators for all dataloaders
iterators = [iter(dl) for dl in self.dataloaders]
# Iterate for the length of the longest dataloader
for iteration_idx in range(len(self)):
batches = []
# Get one batch from each dataloader
for i, iterator in enumerate(iterators):
try:
batch = next(iterator)
except StopIteration:
# This dataloader is exhausted, restart it (cyclic behavior)
iterators[i] = iter(self.dataloaders[i])
batch = next(iterators[i])
batches.append(batch)
# Combine batches
if self.collate_fn is not None:
# Use custom collate function
combined_batch = self.collate_fn(batches)
elif len(batches) > 1:
# Default: return batches as-is (no combination)
# This is for the default PyTorch collate which returns lists/tuples
combined_batch = batches
else:
# Single batch, return as-is
combined_batch = batches[0]
yield combined_batch
def test_multi_dataloader():
"""Test MultiDataLoader with dummy datasets."""
import torch
from torch.utils.data import TensorDataset, DataLoader
print("Testing MultiDataLoader...")
# Create dummy datasets with different sizes
data1 = torch.randn(100, 3, 224, 224)
labels1 = torch.randint(0, 10, (100,))
dataset1 = TensorDataset(data1, labels1)
data2 = torch.randn(30, 3, 224, 224)
labels2 = torch.randint(0, 10, (30,))
dataset2 = TensorDataset(data2, labels2)
# Create dataloaders with different batch sizes
dl1 = DataLoader(dataset1, batch_size=8, shuffle=True)
dl2 = DataLoader(dataset2, batch_size=2, shuffle=True)
print(f"\nDataLoader 1: {len(dataset1)} samples, batch_size=8 → {len(dl1)} iterations")
print(f"DataLoader 2: {len(dataset2)} samples, batch_size=2 → {len(dl2)} iterations")
# Create multi-dataloader
multi_dl = MultiDataLoader([dl1, dl2], batch_sizes=[8, 2])
print(f"\nMultiDataLoader length: {len(multi_dl)} iterations")
print(f"Expected samples per iteration: 10 (8+2)")
# Test iteration
print("\nTesting first 5 iterations:")
for i, batches in enumerate(multi_dl):
if i >= 5:
break
batch1, batch2 = batches
img1, lbl1 = batch1
img2, lbl2 = batch2
print(f" Iteration {i}: batch1={img1.shape}, batch2={img2.shape}")
# Verify batch sizes
assert img1.shape[0] == 8, f"Expected batch size 8, got {img1.shape[0]}"
assert img2.shape[0] == 2, f"Expected batch size 2, got {img2.shape[0]}"
# Test full epoch
print("\nTesting full epoch...")
total_iterations = 0
for _ in multi_dl:
total_iterations += 1
print(f"Total iterations in one epoch: {total_iterations}")
assert total_iterations == len(multi_dl), \
f"Expected {len(multi_dl)} iterations, got {total_iterations}"
print("\nTest passed!")
if __name__ == "__main__":
test_multi_dataloader()
|