Update processing_prismatic.py
Browse files- processing_prismatic.py +74 -17
processing_prismatic.py
CHANGED
@@ -15,12 +15,21 @@ from torchvision.transforms import CenterCrop, Compose, Normalize, Resize, ToTen
|
|
15 |
from transformers import PreTrainedTokenizerBase
|
16 |
from transformers.image_processing_utils import BatchFeature, ImageProcessingMixin
|
17 |
from transformers.processing_utils import ProcessorMixin
|
18 |
-
from transformers.tokenization_utils import
|
|
|
|
|
|
|
|
|
|
|
19 |
from transformers.utils import TensorType
|
20 |
|
|
|
|
|
21 |
|
22 |
# === Image Processing ===
|
23 |
-
def letterbox_pad_transform(
|
|
|
|
|
24 |
"""Given a PIL.Image, pad to square by adding a symmetric border around the height/width."""
|
25 |
(w, h), max_wh = image.size, max(image.size)
|
26 |
horizontal_pad, vertical_pad = int((max_wh - w) / 2), int((max_wh - h) / 2)
|
@@ -62,10 +71,19 @@ class PrismaticImageProcessor(ImageProcessingMixin):
|
|
62 |
stds = [(0.5, 0.5, 0.5)] if stds is None else stds
|
63 |
|
64 |
# TIMM `data_cfg` Parameters
|
65 |
-
self.input_sizes, self.interpolations, self.means, self.stds =
|
|
|
|
|
|
|
|
|
|
|
66 |
|
67 |
# Grab torchvision transforms via TIMM =>> need to parse for specific "functional" transform values!
|
68 |
-
self.tvf_resize_params, self.tvf_crop_params, self.tvf_normalize_params =
|
|
|
|
|
|
|
|
|
69 |
self.tvf_do_letterbox, self.tvf_letterbox_fill = False, None
|
70 |
|
71 |
for idx in range(len(input_sizes)):
|
@@ -90,11 +108,17 @@ class PrismaticImageProcessor(ImageProcessingMixin):
|
|
90 |
and (transform.transforms[0].size == self.input_sizes[idx][-1])
|
91 |
and (transform.transforms[1].size == self.input_sizes[idx][-2:])
|
92 |
):
|
93 |
-
raise ValueError(
|
|
|
|
|
94 |
|
95 |
# HF Image Processors *must* be JSON-serializable; as such, cannot have torchvision. as an attribute.
|
96 |
# => Instead, we're going to parse the transform and call "torchvision.transforms.functional" (`tvf`)
|
97 |
-
resize_t, crop_t, norm_t =
|
|
|
|
|
|
|
|
|
98 |
self.tvf_resize_params.append(
|
99 |
{
|
100 |
"size": resize_t.size,
|
@@ -117,11 +141,15 @@ class PrismaticImageProcessor(ImageProcessingMixin):
|
|
117 |
if self.image_resize_strategy == "resize-naive":
|
118 |
self.tvf_resize_params[idx]["size"] = (resize_t.size, resize_t.size)
|
119 |
elif self.image_resize_strategy == "letterbox":
|
120 |
-
self.tvf_do_letterbox, self.tvf_letterbox_fill = True, tuple(
|
|
|
|
|
121 |
elif self.image_resize_strategy == "resize-crop":
|
122 |
pass
|
123 |
else:
|
124 |
-
raise ValueError(
|
|
|
|
|
125 |
|
126 |
# Dispatch **kwargs to super()
|
127 |
super().__init__(**kwargs)
|
@@ -164,12 +192,19 @@ class PrismaticImageProcessor(ImageProcessingMixin):
|
|
164 |
images = [images]
|
165 |
|
166 |
# Apply `self.img_transform` to each image (will return list of torch.Tensors); stack into "batched" Tensor
|
167 |
-
pixel_values = torch.stack(
|
|
|
|
|
168 |
|
169 |
# Return BatchFeature =>> note that for compatibility, constructor expects Dict[str, np.ndarray], so we convert
|
170 |
-
return BatchFeature(
|
|
|
|
|
|
|
171 |
|
172 |
-
def __call__(
|
|
|
|
|
173 |
return self.preprocess(images, **kwargs)
|
174 |
|
175 |
|
@@ -189,7 +224,9 @@ class PrismaticProcessor(ProcessorMixin):
|
|
189 |
|
190 |
def __call__(
|
191 |
self,
|
192 |
-
text: Union[
|
|
|
|
|
193 |
images: Union[Image.Image, List[Image.Image]],
|
194 |
padding: Union[bool, str, PaddingStrategy] = False,
|
195 |
truncation: Optional[Union[bool, str, TruncationStrategy]] = None,
|
@@ -209,21 +246,31 @@ class PrismaticProcessor(ProcessorMixin):
|
|
209 |
|
210 |
@return: BatchFeature with keys for `input_ids`, `attention_mask` and `pixel_values`.
|
211 |
"""
|
212 |
-
pixel_values = self.image_processor(images, return_tensors=return_tensors)[
|
|
|
|
|
213 |
text_inputs = self.tokenizer(
|
214 |
-
text,
|
|
|
|
|
|
|
|
|
215 |
)
|
216 |
|
217 |
# [Validate] Need same number of images and text inputs!
|
218 |
if pixel_values.shape[0] != text_inputs.input_ids.shape[0]:
|
219 |
-
raise ValueError(
|
|
|
|
|
220 |
|
221 |
return BatchFeature(data={**text_inputs, "pixel_values": pixel_values})
|
222 |
|
223 |
# === Tokenizer Dispatch Utilities =>> check `PreTrainedTokenizerBase` for documentation ===
|
224 |
def batch_decode(
|
225 |
self,
|
226 |
-
sequences: Union[
|
|
|
|
|
227 |
skip_special_tokens: bool = False,
|
228 |
clean_up_tokenization_spaces: Optional[bool] = None,
|
229 |
**kwargs: str,
|
@@ -237,7 +284,9 @@ class PrismaticProcessor(ProcessorMixin):
|
|
237 |
|
238 |
def decode(
|
239 |
self,
|
240 |
-
token_ids: Union[
|
|
|
|
|
241 |
skip_special_tokens: bool = False,
|
242 |
clean_up_tokenization_spaces: Optional[bool] = None,
|
243 |
**kwargs: str,
|
@@ -255,3 +304,11 @@ class PrismaticProcessor(ProcessorMixin):
|
|
255 |
image_processor_input_names = self.image_processor.model_input_names
|
256 |
|
257 |
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
from transformers import PreTrainedTokenizerBase
|
16 |
from transformers.image_processing_utils import BatchFeature, ImageProcessingMixin
|
17 |
from transformers.processing_utils import ProcessorMixin
|
18 |
+
from transformers.tokenization_utils import (
|
19 |
+
PaddingStrategy,
|
20 |
+
PreTokenizedInput,
|
21 |
+
TextInput,
|
22 |
+
TruncationStrategy,
|
23 |
+
)
|
24 |
from transformers.utils import TensorType
|
25 |
|
26 |
+
from .gripper_position import get_gripper_pos_raw
|
27 |
+
|
28 |
|
29 |
# === Image Processing ===
|
30 |
+
def letterbox_pad_transform(
|
31 |
+
image: Image.Image, padding_fill_value: Tuple[int, int, int]
|
32 |
+
) -> Image.Image:
|
33 |
"""Given a PIL.Image, pad to square by adding a symmetric border around the height/width."""
|
34 |
(w, h), max_wh = image.size, max(image.size)
|
35 |
horizontal_pad, vertical_pad = int((max_wh - w) / 2), int((max_wh - h) / 2)
|
|
|
71 |
stds = [(0.5, 0.5, 0.5)] if stds is None else stds
|
72 |
|
73 |
# TIMM `data_cfg` Parameters
|
74 |
+
self.input_sizes, self.interpolations, self.means, self.stds = (
|
75 |
+
input_sizes,
|
76 |
+
interpolations,
|
77 |
+
means,
|
78 |
+
stds,
|
79 |
+
)
|
80 |
|
81 |
# Grab torchvision transforms via TIMM =>> need to parse for specific "functional" transform values!
|
82 |
+
self.tvf_resize_params, self.tvf_crop_params, self.tvf_normalize_params = (
|
83 |
+
[],
|
84 |
+
[],
|
85 |
+
[],
|
86 |
+
)
|
87 |
self.tvf_do_letterbox, self.tvf_letterbox_fill = False, None
|
88 |
|
89 |
for idx in range(len(input_sizes)):
|
|
|
108 |
and (transform.transforms[0].size == self.input_sizes[idx][-1])
|
109 |
and (transform.transforms[1].size == self.input_sizes[idx][-2:])
|
110 |
):
|
111 |
+
raise ValueError(
|
112 |
+
f"Unexpected TIMM image transformation structure/sizes: `{transform}`"
|
113 |
+
)
|
114 |
|
115 |
# HF Image Processors *must* be JSON-serializable; as such, cannot have torchvision. as an attribute.
|
116 |
# => Instead, we're going to parse the transform and call "torchvision.transforms.functional" (`tvf`)
|
117 |
+
resize_t, crop_t, norm_t = (
|
118 |
+
transform.transforms[0],
|
119 |
+
transform.transforms[1],
|
120 |
+
transform.transforms[3],
|
121 |
+
)
|
122 |
self.tvf_resize_params.append(
|
123 |
{
|
124 |
"size": resize_t.size,
|
|
|
141 |
if self.image_resize_strategy == "resize-naive":
|
142 |
self.tvf_resize_params[idx]["size"] = (resize_t.size, resize_t.size)
|
143 |
elif self.image_resize_strategy == "letterbox":
|
144 |
+
self.tvf_do_letterbox, self.tvf_letterbox_fill = True, tuple(
|
145 |
+
[int(x * 255) for x in self.means[idx]]
|
146 |
+
)
|
147 |
elif self.image_resize_strategy == "resize-crop":
|
148 |
pass
|
149 |
else:
|
150 |
+
raise ValueError(
|
151 |
+
f"Image resize strategy `{self.image_resize_strategy}` is not supported!"
|
152 |
+
)
|
153 |
|
154 |
# Dispatch **kwargs to super()
|
155 |
super().__init__(**kwargs)
|
|
|
192 |
images = [images]
|
193 |
|
194 |
# Apply `self.img_transform` to each image (will return list of torch.Tensors); stack into "batched" Tensor
|
195 |
+
pixel_values = torch.stack(
|
196 |
+
[self.apply_transform(img.convert("RGB")) for img in images]
|
197 |
+
)
|
198 |
|
199 |
# Return BatchFeature =>> note that for compatibility, constructor expects Dict[str, np.ndarray], so we convert
|
200 |
+
return BatchFeature(
|
201 |
+
data={"pixel_values": pixel_values.float().numpy()},
|
202 |
+
tensor_type=return_tensors,
|
203 |
+
)
|
204 |
|
205 |
+
def __call__(
|
206 |
+
self, images: Union[Image.Image, List[Image.Image]], **kwargs
|
207 |
+
) -> BatchFeature:
|
208 |
return self.preprocess(images, **kwargs)
|
209 |
|
210 |
|
|
|
224 |
|
225 |
def __call__(
|
226 |
self,
|
227 |
+
text: Union[
|
228 |
+
TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]
|
229 |
+
],
|
230 |
images: Union[Image.Image, List[Image.Image]],
|
231 |
padding: Union[bool, str, PaddingStrategy] = False,
|
232 |
truncation: Optional[Union[bool, str, TruncationStrategy]] = None,
|
|
|
246 |
|
247 |
@return: BatchFeature with keys for `input_ids`, `attention_mask` and `pixel_values`.
|
248 |
"""
|
249 |
+
pixel_values = self.image_processor(images, return_tensors=return_tensors)[
|
250 |
+
"pixel_values"
|
251 |
+
]
|
252 |
text_inputs = self.tokenizer(
|
253 |
+
text,
|
254 |
+
return_tensors=return_tensors,
|
255 |
+
padding=padding,
|
256 |
+
truncation=truncation,
|
257 |
+
max_length=max_length,
|
258 |
)
|
259 |
|
260 |
# [Validate] Need same number of images and text inputs!
|
261 |
if pixel_values.shape[0] != text_inputs.input_ids.shape[0]:
|
262 |
+
raise ValueError(
|
263 |
+
"Batch is malformed; expected same number of images and text inputs!"
|
264 |
+
)
|
265 |
|
266 |
return BatchFeature(data={**text_inputs, "pixel_values": pixel_values})
|
267 |
|
268 |
# === Tokenizer Dispatch Utilities =>> check `PreTrainedTokenizerBase` for documentation ===
|
269 |
def batch_decode(
|
270 |
self,
|
271 |
+
sequences: Union[
|
272 |
+
List[int], List[List[int]], torch.Tensor, Any
|
273 |
+
], # `Any` = np.ndarray | tf.Tensor
|
274 |
skip_special_tokens: bool = False,
|
275 |
clean_up_tokenization_spaces: Optional[bool] = None,
|
276 |
**kwargs: str,
|
|
|
284 |
|
285 |
def decode(
|
286 |
self,
|
287 |
+
token_ids: Union[
|
288 |
+
int, List[int], torch.Tensor, Any
|
289 |
+
], # `Any` = np.ndarray | tf.Tensor
|
290 |
skip_special_tokens: bool = False,
|
291 |
clean_up_tokenization_spaces: Optional[bool] = None,
|
292 |
**kwargs: str,
|
|
|
304 |
image_processor_input_names = self.image_processor.model_input_names
|
305 |
|
306 |
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
|
307 |
+
|
308 |
+
def get_prompt(self, task_label, image: Image.Image):
|
309 |
+
image = image.convert("RGB")
|
310 |
+
image = image.resize((224, 224))
|
311 |
+
gripper_pos, mask, prediction = get_gripper_pos_raw(image)
|
312 |
+
|
313 |
+
prompt = f"In: What action should the robot take to achieve the instruction\nINSTRUCTION: \n{task_label}\nCURRENT GRIPPER: {list(gripper_pos)}\n\nOut:"
|
314 |
+
return prompt, image
|