Update README.md
Browse files
README.md
CHANGED
@@ -185,31 +185,54 @@ MatSynth is accessible through the datasets python library.
|
|
185 |
Following a usage example:
|
186 |
|
187 |
```python
|
|
|
188 |
from datasets import load_dataset
|
189 |
from torch.utils.data import DataLoader
|
190 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
# load the dataset in streaming mode
|
192 |
ds = load_dataset(
|
193 |
"gvecchio/MatSynth",
|
194 |
streaming = True,
|
195 |
)
|
196 |
|
197 |
-
# remove
|
198 |
ds = ds.remove_columns(["diffuse", "specular", "displacement", "opacity", "blend_mask"])
|
199 |
-
# keep only specified columns
|
200 |
-
ds = ds.select_columns(["metadata", "basecolor"
|
|
|
|
|
|
|
201 |
|
202 |
# filter data matching a specific criteria, e.g.: only CC0 materials
|
203 |
ds = ds.filter(lambda x: x["metadata"]["license"] == "CC0")
|
|
|
|
|
204 |
|
205 |
-
#
|
206 |
-
ds = ds.
|
207 |
|
208 |
# set format for usage in torch
|
209 |
ds = ds.with_format("torch")
|
210 |
-
|
|
|
|
|
|
|
|
|
211 |
```
|
212 |
|
|
|
|
|
213 |
## 📜 Citation
|
214 |
|
215 |
```
|
|
|
185 |
Following a usage example:
|
186 |
|
187 |
```python
|
188 |
+
import torchvision.transforms.functional as TF
|
189 |
from datasets import load_dataset
|
190 |
from torch.utils.data import DataLoader
|
191 |
|
192 |
+
# image processing function
|
193 |
+
def process_img(x):
|
194 |
+
x = TF.resize(x, (1024, 1024))
|
195 |
+
x = TF.to_tensor(x)
|
196 |
+
return x
|
197 |
+
|
198 |
+
# item processing function
|
199 |
+
def process_batch(examples):
|
200 |
+
examples["basecolor"] = [process_img(x) for x in examples["basecolor"]]
|
201 |
+
return examples
|
202 |
+
|
203 |
# load the dataset in streaming mode
|
204 |
ds = load_dataset(
|
205 |
"gvecchio/MatSynth",
|
206 |
streaming = True,
|
207 |
)
|
208 |
|
209 |
+
# remove unwanted columns
|
210 |
ds = ds.remove_columns(["diffuse", "specular", "displacement", "opacity", "blend_mask"])
|
211 |
+
# or keep only specified columns
|
212 |
+
ds = ds.select_columns(["metadata", "basecolor"])
|
213 |
+
|
214 |
+
# shuffle data
|
215 |
+
ds = ds.shuffle(buffer_size=100)
|
216 |
|
217 |
# filter data matching a specific criteria, e.g.: only CC0 materials
|
218 |
ds = ds.filter(lambda x: x["metadata"]["license"] == "CC0")
|
219 |
+
# filter out data from Deschaintre et al. 2018
|
220 |
+
ds = ds.filter(lambda x: x["metadata"]["source"] != "deschaintre_2020")
|
221 |
|
222 |
+
# Set up processing
|
223 |
+
ds = ds.map(process_batch, batched=True, batch_size=8)
|
224 |
|
225 |
# set format for usage in torch
|
226 |
ds = ds.with_format("torch")
|
227 |
+
|
228 |
+
# iterate over the dataset
|
229 |
+
for x in ds:
|
230 |
+
print(x)
|
231 |
+
|
232 |
```
|
233 |
|
234 |
+
⚠️ **Note**: Streaming can be slow. We strongly suggest to cache data locally.
|
235 |
+
|
236 |
## 📜 Citation
|
237 |
|
238 |
```
|