def update_graph(amt, hrs):
t0 = time.time()
# First, filter based on the slider values
time_mask = (gdf.Time >= hrs[0]) & (gdf.Time <= hrs[1])
amount_mask = (gdf.Amount >= amt[0]) & (gdf.Amount <= amt[1])
filt_df = gdf.loc[time_mask & amount_mask]
# Then, select the features and train a UMAP model with cuML
features = filt_df.loc[:, "V1":"V28"].values
reducer = cuml.UMAP()
embedding = reducer.fit_transform(features)
# Convert the embedding back to numpy
embedding = cp.asnumpy(embedding)
amount = cp.asnumpy(filt_df.Amount.values.round(2))
# Create a plotly.express scatter plot
fig = px.scatter(
x=embedding[:, 0],
y=embedding[:, 1],
color=amount,
labels={"color": "Amount ($)"},
title="UMAP projection of credit card transactions",
)
t1 = time.time()
out_msg = f"Projected {embedding.shape[0]} transactions in {t1-t0:.2f}s."
alert = dbc.Alert(out_msg, color="success", dismissable=True)
return fig, alert
def get_UMAP_prjs(input_data, cpu=True, **kwargs):
"Compute the projections of `input_data` using UMAP, with a configuration contained in `**kwargs`."
warnings.filterwarnings(
"ignore",
category=NumbaPerformanceWarning) # silence NumbaPerformanceWarning
reducer = umap.UMAP(**kwargs) if cpu else cuml.UMAP(**kwargs)
projections = reducer.fit_transform(input_data)
return projections
def getMapper(self, X, y=None, **kwargs):
if self._mapper is None:
t0 = time.time()
print(f"Computing embedding, input shape = {X.shape}")
input_data: cudf.DataFrame = self.getDataFrame(X)
self._mapper = cuml.UMAP(init=self.init,
n_neighbors=self.n_neighbors,
n_components=self.n_components,
n_epochs=self.n_epochs,
min_dist=self.min_dist,
output_type="numpy")
self._mapper.fit(input_data)
print(
f"Completed umap fit in time {time.time() - t0} sec, embedding shape = {self._embedding_.shape}"
)
return self._mapper
def cuml_umap(config, feature):
# Import RAPIDS
import cudf, cuml
print("INSIDE CUML_UMAP")
print(feature.shape)
num_fr = feature.shape[0]
embed = np.zeros((num_fr, config['n_components']))
# TRY THIS LATER!!!!!!!!!!!!!!!!! IF YOU EVER RUN OUT OF SPACE; COMPARE EMBEDDINGS AND SEE IF SMALLER DATA MAKES A DIFFERENCE
# df = cudf.DataFrame(feature, dtype='float32')
df = cudf.DataFrame(feature)
cu_embed = cuml.UMAP(n_components=config['n_components'], n_neighbors=config['n_neighbors'], n_epochs=config['n_epochs'],
min_dist=config['min_dist'], spread=config['spread'], negative_sample_rate=config['negative_sample_rate'],
init=config['init'], repulsion_strength=config['repulsion_strength'], output_type='numpy').fit_transform(df)
embed[:,0:config['n_components']] = cu_embed
return embed
def prep_trumap_inputs(
interpret_session: InterpretationSession, sid_idx: torch.Tensor,
topk_sim_full_embeds: List, target_text: str,
target_pred_ind: int) -> Tuple[List[Tuple], Tuple, Tuple]:
full_embed_df = cudf.DataFrame([tuple(k) for k in topk_sim_full_embeds])
umap_xformed_embeds = cuml.UMAP(
n_neighbors=5, n_components=3, n_epochs=500,
min_dist=0.1).fit_transform(full_embed_df).as_matrix()
umap_normed_embeds = (umap_xformed_embeds /
np.linalg.norm(umap_xformed_embeds)).tolist()
max_truth_umap, max_falsehood_umap, target_sent_umap = [
umap_normed_embeds.pop() for _ in range(3)
]
trumap_spectrum = []
for idx, umapval in zip(sid_idx.tolist(), umap_normed_embeds):
trumap_spectrum.append(
(interpret_session.stmt_embed_dict['stext'][idx],
interpret_session.stmt_embed_dict['labels'][idx], tuple(umapval)))
target_token_tup = (target_text, target_pred_ind, tuple(target_sent_umap))
umap_bounds_tup = (tuple(max_falsehood_umap), tuple(max_truth_umap))
return trumap_spectrum, target_token_tup, umap_bounds_tup
def reduce_to_3D(data, labels, dimReductionMethod, trainedEmbeddingModel=None):
startTime = time.time()
preTrainedStr = ''
'''
if dimReductionMethod == 'TSNE':
embeddingModel = None
embeddedData = cuml.TSNE( n_components = 2 ).fit_transform ( X = data )
embeddedData.add_column('3', cudf.Series(np.zeros((data.shape[0]))) )
else:
'''
if trainedEmbeddingModel is not None:
preTrainedStr = 'pre-trained '
embeddingModel = trainedEmbeddingModel
else:
if dimReductionMethod == 'PCA':
embeddingModel = cuml.PCA(copy=True,
n_components=3,
random_state=0,
svd_solver='full',
verbose=True,
whiten=False).fit(X=data)
elif dimReductionMethod == 'UMAP':
embeddingModel = cuml.UMAP(n_components=3).fit(X=data, y=labels)
else:
assert ('unable to find embedding model match to user query')
embeddedData = embeddingModel.transform(X=data)
elapsedTime = time.time() - startTime
print(
f'{embeddedData.shape} via {preTrainedStr}{dimReductionMethod} -- completed in: {elapsedTime:.3f} seconds'
)
return embeddedData, embeddingModel
cluster_models = dict(KMeans=cuml.KMeans())
decomposition_models = dict(
PCA=cuml.PCA(),
TruncatedSVD=cuml.TruncatedSVD(),
)
decomposition_models_xfail = dict(
GaussianRandomProjection=cuml.GaussianRandomProjection(),
SparseRandomProjection=cuml.SparseRandomProjection())
neighbor_models = dict(NearestNeighbors=cuml.NearestNeighbors())
dbscan_model = dict(DBSCAN=cuml.DBSCAN())
umap_model = dict(UMAP=cuml.UMAP())
def unit_param(*args, **kwargs):
return pytest.param(*args, **kwargs, marks=pytest.mark.unit)
def quality_param(*args, **kwargs):
return pytest.param(*args, **kwargs, marks=pytest.mark.quality)
def stress_param(*args, **kwargs):
return pytest.param(*args, **kwargs, marks=pytest.mark.stress)
def pickle_save_load(tmpdir, model):
decomposition_models = {
"PCA": lambda: cuml.PCA(),
"TruncatedSVD": lambda: cuml.TruncatedSVD(),
}
decomposition_models_xfail = {
"GaussianRandomProjection": lambda: cuml.GaussianRandomProjection(),
"SparseRandomProjection": lambda: cuml.SparseRandomProjection()
}
neighbor_models = {"NearestNeighbors": lambda: cuml.NearestNeighbors()}
dbscan_model = {"DBSCAN": lambda: cuml.DBSCAN()}
umap_model = {"UMAP": lambda: cuml.UMAP()}
rf_models = {
"rfc": lambda: cuml.RandomForestClassifier(),
"rfr": lambda: cuml.RandomForestRegressor()
}
all_models = {
**regression_models,
**solver_models,
**cluster_models,
**decomposition_models,
**decomposition_models_xfail,
**neighbor_models,
**dbscan_model,
**umap_model,
def plot_comparison(checkpoint_id,
num_progressions=5,
num_classes=20,
num_samples=10000,
algorithm='umap'):
checkpoints = os.listdir('pcl_cifar10_{}'.format(checkpoint_id[0]))
checkpoints.sort()
assert num_progressions <= len(
checkpoints), 'Not enough checkpoints saved.'
checkpoints = [
checkpoints[i]
for i in range(0, len(checkpoints),
len(checkpoints) // (num_progressions - 1))
][:num_progressions - 1] + [checkpoints[-1]]
# fig, axes = plt.subplots(len(checkpoints), 2, figsize=(30,50))
fig, axes = plt.subplots(2, len(checkpoints), figsize=(60, 30))
print(checkpoints)
for i, checkpoint_file in enumerate(checkpoints):
# print(checkpoints)
# print(checkpoint_file)
epoch = checkpoint_file[11:15] # just the 4-digit checkpoint epoch
print('epoch: {}'.format(epoch))
checkpoint = torch.load('pcl_cifar10_{}/{}'.format(
checkpoint_id[0], checkpoint_file))
model.load_state_dict(checkpoint['state_dict'])
features, classes = compute_features(eval_loader,
model,
low_dim=low_dim,
gpu=gpu)
features[
torch.norm(features, dim=1) >
1.5] /= 2 #account for the few samples that are computed twice
features = features.numpy()
restricted_classes = np.array([i for i in classes if i < num_classes])
features = features[np.array(classes) < num_classes]
features = features[:num_samples]
restricted_classes = restricted_classes[:num_samples]
if algorithm == 'umap':
# reducer = umap.UMAP(n_neighbors = 60, min_dist=0.1, n_components=2, metric='cosine')
reducer = cuml.UMAP(n_neighbors=60,
min_dist=0.1,
n_components=2,
n_epochs=1000)
y = reducer.fit_transform(features)
elif algorithm == 'tsne':
tsne = cudaTSNE(n_components=2,
perplexity=50,
learning_rate=600,
verbose=1,
n_iter=2500,
metric='euclidean')
y = tsne.fit_transform(features)
scatter = axes.flat[i].scatter(y[:, 0],
y[:, 1],
c=restricted_classes,
cmap='Spectral',
s=3)
with torch.no_grad():
results = run_dbscan(features,
minPts=200,
minSamples=0,
temperature=0.2,
eps=0.3)
# results = run_kmeans(features, num_cluster=['250'])
im2cluster = results['im2cluster'][0].tolist(
) # remember to turn this back to a list
scatter = axes.flat[i + num_progressions].scatter(
y[:, 0], y[:, 1], c=im2cluster, cmap='Spectral',
s=3) # restricting num_classes does not work here
# legend = axes.flat[i].legend(*scatter.legend_elements(), loc='lower left', title="Classes")
# axes.flat[i].add_artist(legend)
axes.flat[i].set_title('Cifar Classes, epoch: {}'.format(epoch))
axes.flat[i + num_progressions].set_title(
'Clustering Classes, epoch: {}'.format(epoch))
# axes.flat[-1].legend(*scatter.legend_elements(), loc='lower left', title="Classes", bbox_to_anchor=(1.00, 0), prop={'size': 25})
fig.suptitle('{}_{} Comparison'.format(algorithm, checkpoint_id[0]),
fontsize=50)
if not os.path.exists('imgs/{}_{}'.format(algorithm, checkpoint_id[0])):
os.makedirs('imgs/{}_{}'.format(algorithm, checkpoint_id[0]))
save_path = 'imgs/{}_{}/comparison'.format(algorithm, checkpoint_id[0])
fig.savefig(save_path)
print('Figure saved to : {}'.format(save_path))
print('PCA has been skipped')
task_start_time = datetime.now()
n_clusters = 7
if enable_gpu:
kmeans_float = cuml.KMeans(n_clusters=n_clusters)
else:
kmeans_float = sklearn.cluster.KMeans(n_clusters=n_clusters)
kmeans_float.fit(df_fingerprints)
print('Runtime Kmeans time (hh:mm:ss.ms) {}'.format(datetime.now() -
task_start_time))
# UMAP
task_start_time = datetime.now()
if enable_gpu:
umap = cuml.UMAP(n_neighbors=100, a=1.0, b=1.0, learning_rate=1.0)
else:
umap = umap.UMAP()
Xt = umap.fit_transform(df_fingerprints)
print('Runtime UMAP time (hh:mm:ss.ms) {}'.format(datetime.now() -
task_start_time))
if enable_gpu:
df_fingerprints.add_column('x', Xt[0].to_array())
df_fingerprints.add_column('y', Xt[1].to_array())
df_fingerprints.add_column('cluster', kmeans_float.labels_)
else:
df_fingerprints['x'] = Xt[:, 0]
df_fingerprints['y'] = Xt[:, 1]
df_fingerprints['cluster'] = kmeans_float.labels_
distances = cupy.ravel(cupy.fromDlpack(dist_mlarr.to_dlpack()))
indices = cupy.ravel(cupy.fromDlpack(ind_mlarr.to_dlpack()))
print(
f"Computed KNN graph, distances shape = {distances.shape}, indices shape = {indices.shape}, distances[0:5]= {distances[0:5]}, indices[0:5]= {indices[0:5]}"
)
n_samples = indices.shape[0]
n_nonzero = n_samples * n_neighbors
rowptr = cupy.arange(0, n_nonzero + 1, n_neighbors)
knn_graph = cupyx.scipy.sparse.csr_matrix((distances, indices, rowptr),
shape=(n_samples, n_samples))
print(f"Completed KNN, graph shape = {knn_graph.shape}")
reducer = cuml.UMAP(n_neighbors=15,
n_components=3,
n_epochs=500,
min_dist=0.1,
output_type="numpy")
embedding = reducer.fit_transform(data, knn_graph=knn_graph)
print(f"Completed embedding, shape = {embedding.shape}")
# df = embedding.to_pandas()
# df.columns = ["x", "y"]
# df['class'] = pd.Series([str(x) for x in target.to_array()], dtype="category")
#
# cvs = ds.Canvas(plot_width=400, plot_height=400)
# agg = cvs.points(df, 'x', 'y', ds.count_cat('class'))
# img = tf.shade(agg, color_key=color_key, how='eq_hist')
#
# utils.export_image(img, filename='fashion-mnist', background='black')
#
decomposition_models_xfail = {
"GaussianRandomProjection": lambda: cuml.GaussianRandomProjection(),
"SparseRandomProjection": lambda: cuml.SparseRandomProjection()
}
neighbor_models = {
"NearestNeighbors": lambda: cuml.NearestNeighbors()
}
dbscan_model = {
"DBSCAN": lambda: cuml.DBSCAN()
}
umap_model = {
"UMAP": lambda: cuml.UMAP()
}
rf_classification_model = {
"rfc": lambda: cuml.RandomForestClassifier()
}
rf_regression_model = {
"rfr": lambda: cuml.RandomForestRegressor()
}
def pickle_save_load(tmpdir, func_create_model, func_assert):
model, X_test = func_create_model()
pickle_file = tmpdir.join('cu_model.pickle')
def __init__(self, df, n_clusters, chembl_ids, enable_gpu=True, pca_model=False):
self.app = dash.Dash(
__name__, external_stylesheets=external_stylesheets)
self.df = df
self.n_clusters = n_clusters
self.chembl_ids = chembl_ids
self.enable_gpu = enable_gpu
self.pca = pca_model
# Fetch relavant properties from database.
self.prop_df = self.create_dataframe_molecule_properties(chembl_ids)
self.df['chembl_id'] = chembl_ids
self.df['id'] = self.df.index
self.orig_df = df.copy()
# initialize UMAP
if enable_gpu:
self.umap = cuml.UMAP(n_neighbors=100,
a=1.0,
b=1.0,
learning_rate=1.0)
else:
self.umap = umap.UMAP()
# Construct the UI
self.app.layout = self.constuct_layout()
# Register callbacks for selection inside main figure
self.app.callback(
[Output('selected_clusters', 'value'),
Output('selected_point_cnt', 'children')],
[Input('main-figure', 'clickData'),
Input('main-figure', 'selectedData'),
Input('bt_recluster_clusters', 'n_clicks'),
Input('bt_recluster_points', 'n_clicks'),
Input('northstar_cluster', 'children')],
[State("selected_clusters", "value")]) (self.handle_data_selection)
# Register callbacks for buttons for reclustering selected data
self.app.callback(
[Output('main-figure', 'figure'),
Output('northstar_cluster', 'children'),
Output('north_star_clusterid_map', 'children')],
[Input('bt_recluster_clusters', 'n_clicks'),
Input('bt_recluster_points', 'n_clicks'),
Input('bt_north_star', 'n_clicks'),
Input('hidden_northstar', 'value'),
Input('sl_prop_gradient', 'value'),
Input('sl_nclusters', 'value'),],
[State("selected_clusters", "value"),
State("main-figure", "selectedData"),
State('north_star', 'value'),]) (self.handle_re_cluster)
# Register callbacks for selection inside main figure to update module details
self.app.callback(
[Output('tb_selected_molecules', 'children'),
Output('sl_mol_props', 'options'),
Output('current_page', 'children'),
Output('total_page', 'children'),
Output('section_molecule_details', 'style')],
[Input('main-figure', 'selectedData'),
Input('sl_mol_props', 'value'),
Input('sl_prop_gradient', 'value'),
Input('bt_page_prev', 'n_clicks'),
Input('bt_page_next', 'n_clicks'),
Input('north_star_clusterid_map', 'children')],
State('current_page', 'children')) (self.handle_molecule_selection)
self.app.callback(
Output("hidden1", "children"),
[Input("bt_reset", "n_clicks")]) (self.handle_reset)
self.app.callback(
[Output('north_star', 'value'),
Output('hidden_northstar', 'value')],
[Input({'role': 'bt_star_candidate', 'index': ALL}, 'n_clicks')],
State('north_star', 'value')) \
(self.handle_mark_north_star)
import pandas as pd
import cudf, cuml
df = pd.read_csv("data/data.csv")
columns = [
'name', 'artists', 'acousticness', 'danceability', 'energy',
'instrumentalness', 'key', 'liveness', 'loudness', 'speechiness', 'tempo',
'valence'
]
df_mod = df[columns]
keys = df_mod.iloc[:, :2].values.tolist()
features = df_mod.iloc[:, 2:].to_numpy()
features = (features - features.min()) / (features.max() - features.min())
df = cudf.DataFrame(features)
embed = cuml.UMAP(n_neighbors=20, n_epochs=100, min_dist=0.1,
init='spectral').fit_transform(df)
np_embed = embed.to_pandas().to_numpy()
np.save("result/embeddings.npy", np_embed)
