Anthony Miyaguchi
commited on
Commit
•
c10f559
1
Parent(s):
43c4ba2
test for random sizes in images
Browse files- generate_dummy_testset.py +4 -1
- script.py +15 -21
generate_dummy_testset.py
CHANGED
@@ -14,8 +14,11 @@ if __name__ == "__main__":
|
|
14 |
with tempfile.TemporaryDirectory() as tmpdir:
|
15 |
tmp_path = Path(tmpdir)
|
16 |
for row in metadata.itertuples():
|
|
|
|
|
|
|
17 |
img = PIL.Image.fromarray(
|
18 |
-
np.random.randint(0, 255, (
|
19 |
)
|
20 |
img.save(tmp_path / row.filename)
|
21 |
|
|
|
14 |
with tempfile.TemporaryDirectory() as tmpdir:
|
15 |
tmp_path = Path(tmpdir)
|
16 |
for row in metadata.itertuples():
|
17 |
+
# random dimensions
|
18 |
+
x = np.random.randint(100, 300)
|
19 |
+
y = np.random.randint(100, 300)
|
20 |
img = PIL.Image.fromarray(
|
21 |
+
np.random.randint(0, 255, (x, y, 3), dtype=np.uint8)
|
22 |
)
|
23 |
img.save(tmp_path / row.filename)
|
24 |
|
script.py
CHANGED
@@ -13,10 +13,12 @@ from transformers import AutoImageProcessor, AutoModel
|
|
13 |
|
14 |
|
15 |
class ImageDataset(Dataset):
|
16 |
-
def __init__(self, metadata_path, images_root_path):
|
17 |
self.metadata_path = metadata_path
|
18 |
self.metadata = pd.read_csv(metadata_path)
|
19 |
self.images_root_path = images_root_path
|
|
|
|
|
20 |
|
21 |
def __len__(self):
|
22 |
return len(self.metadata)
|
@@ -24,9 +26,18 @@ class ImageDataset(Dataset):
|
|
24 |
def __getitem__(self, idx):
|
25 |
row = self.metadata.iloc[idx]
|
26 |
image_path = Path(self.images_root_path) / row.filename
|
27 |
-
|
28 |
-
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
|
32 |
class LinearClassifier(nn.Module):
|
@@ -40,21 +51,6 @@ class LinearClassifier(nn.Module):
|
|
40 |
return torch.log_softmax(self.model(x), dim=1)
|
41 |
|
42 |
|
43 |
-
class TransformDino:
|
44 |
-
def __init__(self, model_name="./dinov2"):
|
45 |
-
self.processor = AutoImageProcessor.from_pretrained(model_name)
|
46 |
-
self.model = AutoModel.from_pretrained(model_name)
|
47 |
-
|
48 |
-
def forward(self, batch):
|
49 |
-
model_inputs = self.processor(images=batch["features"], return_tensors="pt")
|
50 |
-
with torch.no_grad():
|
51 |
-
outputs = self.model(**model_inputs)
|
52 |
-
last_hidden_states = outputs.last_hidden_state
|
53 |
-
# extract the cls token
|
54 |
-
batch["features"] = last_hidden_states[:, 0]
|
55 |
-
return batch
|
56 |
-
|
57 |
-
|
58 |
def make_submission(
|
59 |
test_metadata,
|
60 |
model_path,
|
@@ -66,13 +62,11 @@ def make_submission(
|
|
66 |
model = LinearClassifier(hparams["num_features"], hparams["num_classes"])
|
67 |
model.load_state_dict(checkpoint["state_dict"])
|
68 |
|
69 |
-
transform = TransformDino()
|
70 |
dataloader = DataLoader(
|
71 |
ImageDataset(test_metadata, images_root_path), batch_size=32, num_workers=4
|
72 |
)
|
73 |
rows = []
|
74 |
for batch in dataloader:
|
75 |
-
batch = transform.forward(batch)
|
76 |
observation_ids = batch["observation_id"]
|
77 |
logits = model(batch["features"])
|
78 |
class_ids = torch.argmax(logits, dim=1)
|
|
|
13 |
|
14 |
|
15 |
class ImageDataset(Dataset):
|
16 |
+
def __init__(self, metadata_path, images_root_path, model_name="./dinov2"):
|
17 |
self.metadata_path = metadata_path
|
18 |
self.metadata = pd.read_csv(metadata_path)
|
19 |
self.images_root_path = images_root_path
|
20 |
+
self.processor = AutoImageProcessor.from_pretrained(model_name)
|
21 |
+
self.model = AutoModel.from_pretrained(model_name)
|
22 |
|
23 |
def __len__(self):
|
24 |
return len(self.metadata)
|
|
|
26 |
def __getitem__(self, idx):
|
27 |
row = self.metadata.iloc[idx]
|
28 |
image_path = Path(self.images_root_path) / row.filename
|
29 |
+
|
30 |
+
model_inputs = self.processor(
|
31 |
+
images=Image.open(image_path), return_tensors="pt"
|
32 |
+
)
|
33 |
+
with torch.no_grad():
|
34 |
+
outputs = self.model(**model_inputs)
|
35 |
+
last_hidden_states = outputs.last_hidden_state
|
36 |
+
# extract the cls token
|
37 |
+
return {
|
38 |
+
"features": last_hidden_states[0, 0],
|
39 |
+
"observation_id": row.observation_id,
|
40 |
+
}
|
41 |
|
42 |
|
43 |
class LinearClassifier(nn.Module):
|
|
|
51 |
return torch.log_softmax(self.model(x), dim=1)
|
52 |
|
53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
def make_submission(
|
55 |
test_metadata,
|
56 |
model_path,
|
|
|
62 |
model = LinearClassifier(hparams["num_features"], hparams["num_classes"])
|
63 |
model.load_state_dict(checkpoint["state_dict"])
|
64 |
|
|
|
65 |
dataloader = DataLoader(
|
66 |
ImageDataset(test_metadata, images_root_path), batch_size=32, num_workers=4
|
67 |
)
|
68 |
rows = []
|
69 |
for batch in dataloader:
|
|
|
70 |
observation_ids = batch["observation_id"]
|
71 |
logits = model(batch["features"])
|
72 |
class_ids = torch.argmax(logits, dim=1)
|