Add DETR (#11653)

* Squash all commits of modeling_detr_v7 branch into one

* Improve docs

* Fix tests

* Style

* Improve docs some more and fix most tests

* Fix slow tests of ViT, DeiT and DETR

* Improve replacement of batch norm

* Restructure timm backbone forward

* Make DetrForSegmentation support any timm backbone

* Fix name of output

* Address most comments by @LysandreJik

* Give better names for variables

* Conditional imports + timm in setup.py

* Address additional comments by @sgugger

* Make style, add require_timm and require_vision to testsé

* Remove train_backbone attribute of DetrConfig, add methods to freeze/unfreeze backbone

* Add png files to fixtures

* Fix type hint

* Add timm to workflows

* Add `BatchNorm2d` to the weight initialization

* Fix retain_grad test

* Replace model checkpoints by Facebook namespace

* Fix name of checkpoint in test

* Add user-friendly message when scipy is not available

* Address most comments by @patrickvonplaten

* Remove return_intermediate_layers attribute of DetrConfig and simplify Joiner

* Better initialization

* Scipy is necessary to get sklearn metrics

* Rename TimmBackbone to DetrTimmConvEncoder and rename DetrJoiner to DetrConvModel

* Make style

* Improve docs and add 2 community notebooks

Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

tests/test_feature_extraction_common.py

@@ -18,6 +18,57 @@ import json
 import os
 import tempfile
 
+from transformers.file_utils import is_torch_available, is_vision_available
+
+
+if is_torch_available():
+    import numpy as np
+    import torch
+
+if is_vision_available():
+    from PIL import Image
+
+
+def prepare_image_inputs(feature_extract_tester, equal_resolution=False, numpify=False, torchify=False):
+    """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+    or a list of PyTorch tensors if one specifies torchify=True.
+    """
+
+    assert not (numpify and torchify), "You cannot specify both numpy and PyTorch tensors at the same time"
+
+    if equal_resolution:
+        image_inputs = []
+        for i in range(feature_extract_tester.batch_size):
+            image_inputs.append(
+                np.random.randint(
+                    255,
+                    size=(
+                        feature_extract_tester.num_channels,
+                        feature_extract_tester.max_resolution,
+                        feature_extract_tester.max_resolution,
+                    ),
+                    dtype=np.uint8,
+                )
+            )
+    else:
+        image_inputs = []
+        for i in range(feature_extract_tester.batch_size):
+            width, height = np.random.choice(
+                np.arange(feature_extract_tester.min_resolution, feature_extract_tester.max_resolution), 2
+            )
+            image_inputs.append(
+                np.random.randint(255, size=(feature_extract_tester.num_channels, width, height), dtype=np.uint8)
+            )
+
+    if not numpify and not torchify:
+        # PIL expects the channel dimension as last dimension
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+    if torchify:
+        image_inputs = [torch.from_numpy(x) for x in image_inputs]
+
+    return image_inputs
+
 
 class FeatureExtractionSavingTestMixin:
     def test_feat_extract_to_json_string(self):