'
DATA_PATH = './data'
ft_visual_features_file = DATA_PATH + '/dataset_v1_visual_features_database.npy'
#load database features:
ft_visual_features_database = np.load(ft_visual_features_file)
database_csv_path = os.path.join(DATA_PATH, 'dataset_v1.csv')
database_df = pd.read_csv(database_csv_path)
class NearestNeighbors:
"""
Class for NearestNeighbors.
"""
def __init__(self, n_neighbors=10, metric='cosine', rerank_from=-1):
"""
metric = 'cosine' / 'binary'
if metric ~= 'cosine' and rerank_from > n_neighbors then a cosine rerank will be performed
"""
self.n_neighbors = n_neighbors
self.metric = metric
self.rerank_from = rerank_from
def normalize(self, a):
return a / np.sum(a**2, axis=1, keepdims=True)
def fit(self, data, o_data=None):
if self.metric == 'cosine':
data = self.normalize(data)
self.index = faiss.IndexFlatIP(data.shape[1])
elif self.metric == 'binary':
self.o_data = data if o_data is None else o_data
#assuming data already packed
self.index = faiss.IndexBinaryFlat(data.shape[1]*8)
self.index.add(np.ascontiguousarray(data))
def kneighbors(self, q_data):
if self.metric == 'cosine':
q_data = self.normalize(q_data)
sim, idx = self.index.search(q_data, self.n_neighbors)
else:
if self.metric == 'binary':
print('binary search: ')
bq_data = np.packbits((q_data > 0.0).astype(bool), axis=1)
print(bq_data.shape, self.index.d)
sim, idx = self.index.search(bq_data, max(self.rerank_from, self.n_neighbors))
if self.rerank_from > self.n_neighbors:
rerank_data = self.o_data[idx[0]]
rerank_search = NearestNeighbors(n_neighbors=self.n_neighbors, metric='cosine')
rerank_search.fit(rerank_data)
sim, re_idxs = rerank_search.kneighbors(q_data)
idx = [idx[0][re_idxs[0]]]
return sim, idx
model = CLIPTextModelWithProjection.from_pretrained("Diangle/clip4clip-webvid")
tokenizer = AutoTokenizer.from_pretrained("Diangle/clip4clip-webvid")
def search(search_sentence):
inputs = tokenizer(text=search_sentence , return_tensors="pt", padding=True)
outputs = model(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"], return_dict=False)
# Customized projection layer
text_projection = model.state_dict()['text_projection.weight']
text_embeds = outputs[1] @ text_projection
final_output = text_embeds[torch.arange(text_embeds.shape[0]), inputs["input_ids"].argmax(dim=-1)]
# Normalization
final_output = final_output / final_output.norm(dim=-1, keepdim=True)
final_output = final_output.cpu().detach().numpy()
sequence_output = final_output / np.sum(final_output**2, axis=1, keepdims=True)
nn_search = NearestNeighbors(n_neighbors=5, metric='binary', rerank_from=100)
nn_search.fit(np.packbits((ft_visual_features_database > 0.0).astype(bool), axis=1), o_data=ft_visual_features_database)
sims, idxs = nn_search.kneighbors(sequence_output)
return database_df.iloc[idxs[0]]['contentUrl'].to_list()
with gr.Blocks() as demo:
gr.HTML(TITLE)
gr.Markdown(DESCRIPTION)
gr.HTML(IMAGE)
gr.Markdown("Retrieval of top 5 videos relevant to the input sentence: ")
with gr.Row():
with gr.Column():
inp = gr.Textbox(placeholder="Write a sentence.")
btn = gr.Button(value="Retrieve")
ex = [["a woman waving to the camera"],["a basketball player performing a slam dunk"], ["how to bake a chocolate cake"], ["birds fly in the sky"]]
gr.Examples(examples=ex,
inputs=[inp],
)
with gr.Column():
out = [gr.Video(format='mp4') for _ in range(5)]
btn.click(search, inputs=inp, outputs=out)
demo.launch()