data_transforms_test = albumentations.Compose([
albumentations.Resize(RESIZE_H, RESIZE_W),
albumentations.Normalize(),
AT.ToTensor()
])
# load the best model and initialize EmbeddingNet
siamese_net = torch.load("best_net.pth")
state_dict = siamese_net.state_dict()
embed_net = EmbeddingNet()
temp_dict = {}
for key in state_dict.keys():
if key.startswith("embedding_net"):
temp_dict[key[14:len(key)]] = state_dict[key]
embed_net.load_state_dict(temp_dict)
# function to generate embedding
def getEmbedding(file_path, x):
file_name = os.path.join(file_path, x)
bbox = bbox_df.loc[bbox_df.Image == x, :].values[0, 1:]
img_pil = Image.open(file_name).crop(bbox).convert('RGB')
img = np.array(img_pil)
image = data_transforms_test(image=img)['image'].unsqueeze(0)
vector = embed_net(image)
return vector
# test
embed_net.eval()
labels = np.zeros(len(dataloader.dataset))
k = 0
for images, target in dataloader:
if cuda:
images = images.cuda()
embeddings[k:k + len(images)] = model.get_embedding(
images).data.cpu().numpy()
labels[k:k + len(images)] = target.numpy()
k += len(images)
return embeddings, labels
# In[4]:
model = EmbeddingNet()
model.load_state_dict(torch.load('./saved_model/titi'))
model.eval()
# In[5]:
final_test_epoch('/Users/ayush/projects/my_pytorch/probe',
'/Users/ayush/projects/my_pytorch/gallery',
'/Users/ayush/projects/my_pytorch/fp_output_txt',
model,
metrics=[AverageNonzeroTripletsMetric()],
transform=transforms.Compose([transforms.ToTensor()]))
# In[7]:
get_ipython().system('rm /Users/ayush/projects/my_pytorch/probe/.DS_Store')