def get_2d_coordinates_tsne(multinet, output_format="json", verbose=True):
embedding = multinet.embedding
X = embedding[0]
indices = embedding[1]
if verbose:
multinet.monitor("Doing the TSNE reduction to 2 dimensions!")
if parallel_tsne:
X_embedded = TSNE(n_components=2,
n_jobs=mp.cpu_count()).fit_transform(X)
else:
X_embedded = TSNE(n_components=2).fit_transform(X)
dfr = pd.DataFrame(X_embedded, columns=['dim1', 'dim2'])
dfr['node_names'] = [n for n in multinet.get_nodes()]
dfr['node_codes'] = indices
if output_format == "json":
return dfr.to_json(orient='records')
elif output_format == "dataframe":
# pure pandas dataframe
return dfr
elif output_format == "pos_dict":
output_dict = {}
for index, row in dfr.iterrows():
output_dict[row['node_names']] = (row['dim1'], row['dim2'])
return output_dict
else:
return None
def compare_embeddings_in_group(config, data, n_samples, point_size=10, log=True, images_path=None):
sample_idx = random.sample(range(len(data)), n_samples)
raw_obs, raw_action, raw_reward = data[0]
raw_action = torch.repeat_interleave(raw_action[1:], 2, dim=0)
action_dim = raw_action.size(1)
initial_episode_size = raw_action.size(0)
actual_episode_size = initial_episode_size - (initial_episode_size % config.traj_len)
rewards_ak = []
embeddings_ak = []
for k in sample_idx:
raw_obs, raw_action, raw_reward = data[k]
raw_embeddings = torch.repeat_interleave(raw_action[1:], 2, dim=0)[:actual_episode_size] \
.reshape([actual_episode_size // config.traj_len,
config.traj_len * action_dim])
rewards = torch.repeat_interleave(raw_reward[1:]/2, 2, dim=0)[:actual_episode_size] \
.reshape([actual_episode_size // config.traj_len,
config.traj_len]).sum(axis=1)
for idx, vector in enumerate(raw_embeddings):
embeddings_ak.append(vector.numpy())
rewards_ak.append(rewards[idx])
tsne_ak_2d = TSNE(perplexity=30, n_components=2, init='random', n_iter=3500, random_state=32, n_jobs=8)
embeddings_ak_2d = tsne_ak_2d.fit_transform(np.array(embeddings_ak))
dyne_emb_ak = []
for k in sample_idx:
mu, logvar, _ = data.transform_episode(k)
for idx, vector in enumerate(mu):
dyne_emb_ak.append(vector.numpy())
tsne_dyne_ak_2d = TSNE(perplexity=30, n_components=2, init='random', n_iter=3500, random_state=32, n_jobs=8)
embeddings_dyne_ak_2d = tsne_dyne_ak_2d.fit_transform(np.array(dyne_emb_ak))
fig, axes = plt.subplots(2, 2, figsize=(20, 20), constrained_layout=True)
tsne_plot_2d('raw actions by time'.format(config.env), embeddings_ak_2d,
color_style='time', rewards=rewards_ak, size=point_size,
log=False, ax=axes[0][0], episodes_num=n_samples)
tsne_plot_2d('dyne actions by time', embeddings_dyne_ak_2d,
color_style='time', rewards=rewards_ak, size=point_size,
log=False, ax=axes[0][1], episodes_num=n_samples)
tsne_plot_2d('raw actions by rewards', embeddings_ak_2d,
color_style='rewards', rewards=rewards_ak, size=point_size,
log=False, ax=axes[1][0], episodes_num=n_samples)
tsne_plot_2d('dyne actions by rewards', embeddings_dyne_ak_2d,
color_style='rewards', rewards=rewards_ak, size=point_size,
log=False, ax=axes[1][1], episodes_num=n_samples)
fig.suptitle("{}_DynE-{}".format(config.env, config.traj_len), fontsize=16)
fig.savefig(images_path / "{}_emb_comparison_{}_samples.png".format(config.env, config.n_samples),
format='png', dpi=150, bbox_inches='tight')
if log:
wandb.log({'Embeddings Comparison': wandb.Image(fig)})
def create_compute_tsne_components_function(input_dim, target_dim,
save_folder):
# Get t-SNE function
tsne = TSNE(n_jobs=4)
if False:
tsne = TSNE(n_jobs=number_of_jobs,
n_components=target_dim,
perplexity=perplexity,
early_exaggeration=early_exaggeration,
learning_rate=learning_rate,
n_iter=n_iter,
n_iter_without_progress=n_iter_without_progress,
min_grad_norm=min_grad_norm,
metric=metric,
init=init,
verbose=verbose,
random_state=random_state,
method=method,
angle=angle)
# Create the function which we are interested in
def compute_tsne_components(in_feature_file_path, out_feature_file_path):
xprint("Processing '{}'...".format(in_feature_file_path))
with open(in_feature_file_path, "rb") as in_feature_file:
# Read features from file
features = np.reshape(
np.fromfile(in_feature_file, dtype=feature_dtype),
(-1, input_dim))
# Compute t-SNE components
components = tsne.fit_transform(features)
xprint("features.shape:", features.shape)
xprint("components.shape:", components.shape)
if procuce_output_files:
xprint("Creating '{}'...".format(out_feature_file_path))
with open(out_feature_file_path, "wb") as out_feature_file:
# Write standardized features to file
components.tofile(out_feature_file)
if produce_plots:
create_plots(components,
title=os.path.basename(in_feature_file_path),
total_variance=np.sum(np.var(features, axis=0)),
save_folder=save_folder)
return compute_tsne_components
def main():
parser = argparse.ArgumentParser(description='main function parser')
parser.add_argument('--path', type=str, help='load file path', required=True)
parser.add_argument('--dump_dir', type=str, help='dump directory', default=None)
parser.add_argument('--size', type=int, default=1000, help='embedding vector size')
args = parser.parse_args()
embeddings, labels = load(args.path)
embeddings = np.array(embeddings)
output = args.path.split('/')[-1]
# # UMAP
n_neighbors = [15] #, 35, 55, 75]
min_dists = [0.1] #0.001, 0.01, 0.1]
for min_dist in min_dists:
for n_neighbor in n_neighbors:
start = time.time()
weights = umap.UMAP(n_neighbors=n_neighbor, min_dist=min_dist).fit_transform(embeddings)
finish = time.time()
print(f'time: {finish-start} s', flush=True)
os.makedirs(f'graph/umap/{output}', exist_ok=True)
show(weights, labels, f'graph/umap/{output}/min_dist:{min_dist}_neighbor:{n_neighbor}.svg')
# t-SNE
perplexities = [30] #10, 20, 30, 40, 50]
for perplexity in perplexities:
start = time.time()
tsne_model = TSNE(n_components=2, perplexity=perplexity, n_jobs=10)
weights = tsne_model.fit_transform(embeddings)
finish = time.time()
print(f'time: {finish-start} s', flush=True)
os.makedirs(f'graph/tsne/{output}', exist_ok=True)
show(weights, labels, f'graph/tsne/{output}/perplexity:{perplexity}.svg')
def main(path):
embs = pool_sentence_embs(path)
print("Dimension",
embs.shape) #number sentences X BERT hidden dimension (768)
df = pd.read_csv('master_df_ALL.csv', encoding='utf - 8', index_col=False)
filter_name = 'Coreference'
target = df[filter_name]
#df['Sentences'].to_csv("master_ALL", header=None, index=False)
#target = [1., 0., 0., 1., 1., 1., 0., 1., 0., 1., 1., 0., 0., 1., 0., 0., 0., 0., 0., 1.] + [0. for i in range(22)]
print(len(target))
embeddings = TSNE(n_jobs=4, random_state=1).fit_transform(
embs) #t-SNE reduces 768 dimension to 2D or 3D
vis_x = embeddings[:, 0]
vis_y = embeddings[:, 1]
plt.scatter(vis_x,
vis_y,
c=target,
cmap=ListedColormap(["blue", "red"]),
marker='.',
s=50)
plt.title(filter_name +
" filter (red=passed filter, blue=did not pass filter)")
# plt.colorbar(ticks=range(10))
# plt.clim(-0.5, 9.5)
plt.ioff()
#plt.show()
plt.savefig(filter_name)
def run_tSNE(natural_embed, n_jobs, perplexity):
'''
The GPU version requires CUDA 9.0 and install the tsnecuda package by running
conda install tsnecuda -c cannylab
The Multicore CPU version can be installed by running
pip install MulticoreTSNE
Apply t-SNE to the input data
INPUT:
natural_embed: 2d numpy array with size [number of points, embedding length]
n_jobs:
perplexity:
OUTPUT:
natural_2d: 2d numpy array with size [number of points, 2]
adversary_2d: 2d numpy array with size [number of points, 2]
'''
X = natural_embed
# CPU Sklearn
# from sklearn.manifold import TSNE
# tsne = TSNE(perplexity=perplexity, n_iter=5000, n_iter_without_progress=800, learning_rate=20, metric='cosine')
# X_embedded = tsne.fit_transform(X)
# CPU
from MulticoreTSNE import MulticoreTSNE as TSNE
tsne = TSNE(n_jobs=n_jobs, perplexity=perplexity, n_iter=5000, n_iter_without_progress=800, learning_rate=20, metric='cosine')
X_embedded = tsne.fit_transform(X)
# GPU
# from tsnecuda import TSNE
# X_embedded = TSNE(n_components=2, perplexity=30, learning_rate=10).fit_transform(X)
return X_embedded
def classifier_choice(method='tsne', neighbors=30, dimensions=2):
if method in "tsne":
return TSNE(n_components=dimensions, perplexity=30, verbose=1)
elif method in "pca":
return decomposition.TruncatedSVD(n_components=dimensions)
elif method in "isomap":
return manifold.Isomap(n_neighbors=neighbors, n_components=dimensions)
elif method in "lle":
return manifold.LocallyLinearEmbedding(n_neighbors=neighbors,
n_components=dimensions,
method='standard')
elif method in "mlle":
return manifold.LocallyLinearEmbedding(n_neighbors=neighbors,
n_components=dimensions,
method='modified')
elif method in "hlle":
return manifold.LocallyLinearEmbedding(n_neighbors=neighbors,
n_components=dimensions,
method='hessian')
elif method in "ltsa":
return manifold.LocallyLinearEmbedding(n_neighbors=neighbors,
n_components=dimensions,
method='ltsa')
elif method in "mds":
return manifold.MDS(n_components=dimensions, n_init=1, max_iter=100)
elif method in "trees":
trees = ensemble.RandomTreesEmbedding(n_estimators=200, max_depth=5)
pca = decomposition.TruncatedSVD(n_components=dimensions)
return Pipeline([('Random Tree Embedder', trees), ('PCA', pca)])
elif method in "spectral":
return manifold.SpectralEmbedding(n_components=dimensions,
eigen_solver="arpack")
else:
print('Please use valid method')
def get_data(n_cmd, n_spk, only_missed=False):
if only_missed:
# most popular MIS-CLASSIFIED command based on utterances count
top_cmd = itemfreq(y_command[y_missed.astype('int32')])
top_spk = itemfreq(y_speaker[y_missed.astype('int32')])
else:
top_spk = itemfreq(y_speaker)
top_cmd = itemfreq(y_command)
top_cmd = top_cmd[np.argsort(top_cmd[:, 1])][::-1]
top_cmd = top_cmd[:, 0]
# most speaker command based on utterances count
top_spk = top_spk[np.argsort(top_spk[:, 1].astype('int32'))][::-1]
top_spk = top_spk[:, 0]
spk = top_spk[:n_spk]
cmd = top_cmd[:n_cmd]
ids = get_indices(speaker_set=spk, command_set=cmd)
if only_missed:
ids = np.array([i for i in ids if i in y_missed], dtype='int32')
y_cmd = y_command[ids]
y_spk = y_speaker[ids]
z_org = Z_original[ids]
z_max = Z_maximize[ids]
tsne = TSNE(random_state=SEED)
t = tsne.fit_transform(np.concatenate((z_org, z_max), axis=0))
t_org = t[:z_org.shape[0]]
t_max = t[z_org.shape[0]:]
return t_org, t_max, y_cmd, y_spk
def tsne_reduction(samples,
perplexity,
data=None,
n_components=2,
l_r=200,
dim=2,
ex=12,
iterations=5000,
verbosity=0):
if (samples is None) and (data is not None):
samples = data[:, :-1]
targets = data[:, -1]
# tsne = manifold.TSNE(n_components = dim, init='pca', learning_rate = l_r,
# perplexity=perplexity, early_exaggeration = ex,
# n_iter = iterations, random_state=data_handling.RANDOM_SEED,
# verbose = verbosity)
tsne = TSNE(n_components=dim,
n_jobs=-1,
learning_rate=l_r,
perplexity=perplexity,
early_exaggeration=ex,
n_iter=iterations,
random_state=data_handling.RANDOM_SEED,
verbose=verbosity)
reduced_samples = tsne.fit_transform(samples)
return reduced_samples, tsne
def calcTSNEMulti(data, iterations, perplexity, learning_rate):
tsne = TSNE(n_jobs=4,
perplexity=perplexity,
n_iter=iterations,
learning_rate=learning_rate)
Y = tsne.fit_transform(data)
return data.assign(x=Y[:, 0], y=Y[:, 1])
def train(self, parameters):
tsne = TSNE(**parameters)
tsne_outputs = tsne.fit_transform(self.x_train)
utils.save_data_to_pkl(tsne_outputs,
tsne_outputs_path + 'tsne_outputs.p')
def main(feats_path):
with open(feats_path, 'rb') as handle:
unpickler = pickle.Unpickler(handle)
labels = unpickler.load()
labels = {
name: vector
for name, vector in labels.items() if vector is not None
}
features = np.asarray(list(labels.values()))
print('[INFO] Conducting t-SNE on ' + feats_path)
tsne = TSNE(metric='braycurtis',
verbose=1,
n_iter=5000,
random_state=42,
n_jobs=-1)
projection = tsne.fit_transform(features)
# save reduced vectors
base = path.basename(feats_path)
name = path.splitext(base)[0]
output = name + '_tsne.pickle'
print('[INFO] Saving reduced vectors to ' + output)
with open(output, 'wb') as handle:
pickle.dump(projection, handle)
def dim_red_plot(plt_type,
emb,
vocab,
output_dir,
n_components=2,
random_state=42):
print(f"-- Start {plt_type} --")
if plt_type == 'tsne':
new_values = TSNE(n_components=n_components,
random_state=random_state,
n_jobs=10,
verbose=2).fit_transform(emb)
x = []
y = []
for value in new_values:
x.append(value[0])
y.append(value[1])
elif plt_type == 'umap':
new_values = umap.UMAP(n_components=n_components,
random_state=random_state).fit_transform(emb)
x, y = new_values[:, 0], new_values[:, 1]
print("-- Start ploting --")
plt.figure(figsize=(16, 16))
plt.scatter(x, y)
# for i in range(len(x)):
# plt.annotate(vocab[i], xy=(x[i], y[i]), xytext=(
# 5, 2), textcoords="offset points", ha="right", va="bottom")
plt.savefig(os.path.join(output_dir, f'viz/emb_{plt_type}.png'))
def draw(x, y):
from matplotlib.colors import ListedColormap
from MulticoreTSNE import MulticoreTSNE as TSNE
print("TSNE: fitting start...")
tsne = TSNE(2, n_jobs=4, perplexity=30)
Y = tsne.fit_transform(x)
# matplotlib_axes_logger.setLevel('ERROR')
labels = [
'c1', 'c2', 'c3', 'c4', 'c5', 'c6', 'c7', 'c8', 'c9', 'c10', 'open'
]
id_to_label = {i: label for i, label in enumerate(labels)}
y_true = pd.Series(y)
plt.style.use('ggplot')
n_class = y_true.unique().shape[0]
colors = ('gray', 'lightgreen', 'plum', 'DarkMagenta', 'SkyBlue',
'PaleTurquoise', 'DeepPink', 'Gold', 'Orange', 'Brown',
'DarkKhaki')
fig, ax = plt.subplots(figsize=(9, 6), )
la = [i for i in range(n_class)]
la = sorted(la, reverse=True)
cmap = ListedColormap(colors)
for idx, label in enumerate(la):
ix = y_true[y_true == label].index
x = Y[:, 0][ix]
y = Y[:, 1][ix]
ax.scatter(x, y, c=cmap(idx), label=id_to_label[label], alpha=0.5)
# Shrink current axis by 20%
ax.set_title('proto_loss')
box = ax.get_position()
ax.set_position([box.x0, box.y0, box.width * 0.8, box.height])
ax.legend(loc='center left', bbox_to_anchor=(1, 0.5))
def generate2dftEmb():
global w2id, w, i, word, tsne, post_2d
####
# Loading glove embeddings from pickle file glove_new.pickle and writing into embedding map and a text file which
# can be used to gensim model
####
file = open(finetuned_path, 'rb')
embedding_map = pickle.load(file)
# In[470]:
##########
##Converting glove embeddings to numpy matrix where each row contains embedding of a word.
##Adding words to "word to id" and "id to word" maps
##########
w2id = {}
id2w = {}
w = np.zeros((len(embedding_map.keys()), 300))
for i, word in enumerate(embedding_map.keys()):
w2id[word] = i
id2w[i] = word
w[i] = embedding_map[word]
# In[6]:
######
##Applying t-SNE to reduce the dimension of the embedding from 300D to 2D.
######
tsne = TSNE(n_jobs=12)
post_2d = tsne.fit_transform(w)
# In[486]:
return post_2d, w2id, w
def generate2dpre():
global word, i, pre_w2id, tsne, pre_2d
pre_vocab = []
pre = open(pretrained_path, 'r')
for line in pre:
embeds = line.rstrip().split(" ")
word = embeds[0]
pre_vocab.append(word)
# In[37]:
pre_w = np.zeros((len(pre_vocab), 300))
for i, line in enumerate(pre):
embeds = line.rstrip().split(" ")
word = embeds[0]
pre_w[i, :] = embeds[1:]
# In[ ]:
##########
##Converting pre glove embeddings to numpy matrix where each row contains embedding of a word.
##Adding words to "word to id" and "id to word" maps
##########
pre_w2id = {}
for i in range(len(pre_vocab)):
pre_w2id[pre_vocab[i]] = i
# In[39]:
######
##Applying t-SNE to reduce the dimension of the embedding from 300D to 2D.
######
tsne = TSNE(n_jobs=12)
pre_2d = tsne.fit_transform(pre_w)
return pre_2d, pre_w2id, pre_w
def plot_distribution(
epoch,
train,
# acc,
path,
data_x,
# true_y,
pred_y,
learning_rate=100,
n_jobs=-1):
print("plotting image on " + path + "...")
if (os.path.exists(path) == False):
os.makedirs(path)
tsne_model = TSNE(n_components=2,
learning_rate=learning_rate,
n_jobs=n_jobs)
# pca_model = PCA(n_components=2)
data_x = np.array(data_x)
if (len(data_x.shape) > 2):
data_temp = []
for data in data_x:
data_temp.append(data.rehsape(-1))
data_x = np.array(data_temp)
transformed = tsne_model.fit_transform(data_x)
# transformed = pca_model.fit_transform(data_x)
xs = transformed[:, 0]
ys = transformed[:, 1]
# draw_plot(xs, ys, train, epoch, true_y, os.path.join(path, "true_label"))
draw_plot(xs, ys, train, epoch, pred_y, path)
def tsne_executor(X, y, logger, path_logs):
check_input_type(
['epi'],
"t-SNE experiment work just with epigenetic data, {} found".format(
config['general']['input_type']))
cell_lines = config['general']['cell_lines']
tasks_dict = config['general']['tasks']
results = {}
for t in tasks_dict:
task_name, X_filtered, y_filtered = filter_labels(X, y, t)
logger.debug("TASK: {}".format(task_name))
cpus = multiprocessing.cpu_count(
) // 2 # we use just half of avaible cpus to not overload the machine
logger.debug("Using {} cpus".format(cpus))
for cl, data, labels in zip(cell_lines, X_filtered, y_filtered):
logger.debug("Computing t-SNE for {}".format(cl))
tsne = TSNE(perplexity=config['tsne']['perplexity'],
n_jobs=cpus) # TODO: add parameters
tsne_results = tsne.fit_transform(data)
assert len(tsne_results) == len(labels)
tsne_results = np.c_[
tsne_results,
labels] # to save the labels with the tsne results
results["{}_{}".format(task_name, cl)] = tsne_results
save_tsne(path_logs, "tsne_results", results)
if config['tsne']['save_plots']:
plot_tsne(results, path_logs, "tsne_plot")
def plot_tsne(experience=None, latent_states=None, rewards=None):
if latent_states is None or rewards is None:
latent_states = np.array([
list(rssm_state.prev_state.stoch)
for rssm_state in experience['agent_infos']
])
rewards = np.array(experience['reward'])
np.random.seed(0)
perm = np.random.permutation(10000)
latent_states = latent_states[perm]
rewards = rewards[perm]
feature_cols = ['axis_' + str(i) for i in range(latent_states.shape[1])]
df = DataFrame(latent_states, columns=feature_cols)
df['y'] = rewards
time_start = time()
tsne = TSNE(n_components=2,
verbose=1,
perplexity=1000,
n_iter=1000,
n_jobs=16)
tsne_results = tsne.fit_transform(df[feature_cols].values)
print('t-SNE done! Time elapsed: {} seconds'.format(time() - time_start))
pickle.dump(tsne_results, open('tsne_results.pkl', 'wb'))
df['tsne-2d-one'] = tsne_results[:, 0]
df['tsne-2d-two'] = tsne_results[:, 1]
sns.scatterplot(x="tsne-2d-one",
y="tsne-2d-two",
hue="y",
palette=sns.color_palette("flare", as_cmap=True),
data=df,
alpha=0.6,
s=5)
plt.show()
def dimensionality_reduction(X, algorithm="PCA"):
"""Reduce the dimensionality of the AISNPs
:param X: One-hot encoded 1kG AISNPs.
:type X: pandas DataFrame
:param algorithm: The type of dimensionality reduction to perform.
One of {PCA, UMAP, t-SNE}
:type algorithm: str
:returns: The transformed X DataFrame, reduced to 3 components by <algorithm>,
and the dimensionality reduction Transformer object.
"""
n_components = 3
if algorithm == "PCA":
reducer = PCA(n_components=n_components)
elif algorithm == "t-SNE":
reducer = TSNE(n_components=n_components, n_jobs=4)
elif algorithm == "UMAP":
reducer = umap.UMAP(n_components=n_components,
min_dist=0.2,
metric="dice",
random_state=42)
else:
return None, None
X_reduced = reducer.fit_transform(X.values)
return pd.DataFrame(X_reduced, columns=["x", "y", "z"],
index=X.index), reducer
def decompose(dimred, dim, nneigh):
if dimred == 'MDS': # slowest!
embedding = MDS(n_components=dim,
n_init=__inits,
max_iter=__iters,
n_jobs=-1,
dissimilarity=__dis)
elif dimred == 'ISOMAP': # slow
embedding = Isomap(n_neighbors=nneigh, n_components=dim, n_jobs=-1)
elif dimred == 'LLE': # slow-acceptable
embedding = LocallyLinearEmbedding(n_neighbors=nneigh,
n_components=dim,
n_jobs=-1)
elif dimred == 'TSNE': # acceptable
embedding = TSNE(n_components=dim,
n_iter=__iters,
metric='precomputed',
learning_rate=__lrate,
perplexity=__perplexity)
elif dimred == 'UMAP': # fast
# embedding = umap.UMAP(n_neighbors=nneigh, n_components=dim, metric=__dis, min_dist=0.1)
embedding = umap.UMAP(n_neighbors=nneigh,
n_components=dim,
min_dist=0.1)
elif dimred == 'PCA': # fastest!
embedding = PCA(n_components=dim)
else:
raise ValueError('dimension reduction method not recognized')
positions = embedding.fit_transform(An)
return positions
def reduce_dim(df, algorithm='pca'):
"""Reduce the dimensionality of the 55 AISNPs
:param X: One-hot encoded 1kG 55 AISNPs.
:type X: pandas DataFrame
:param algorithm: The type of dimensionality reduction to perform.
One of {pca, umap, tsne}
:type algorithm: str
:returns: The transformed X DataFrame, reduced to 3 components by <algorithm>.
"""
ncols = len(df.columns)
ohe = OneHotEncoder(categories=[range(4)] * ncols, sparse=False)
n_components = 3
X = ohe.fit_transform(df.values)
if algorithm == 'pca':
X_red = PCA(n_components=n_components).fit_transform(X)
elif algorithm == 'tsne':
# TSNE, Barnes-Hut have dim <= 3
if n_components > 3:
print(
'The Barnes-Hut method requires the dimensionaility to be <= 3'
)
return None
else:
X_red = TSNE(n_components=n_components,
n_jobs=4).fit_transform(X)
elif algorithm == 'umap':
X_red = umap.UMAP(n_components=n_components).fit_transform(X)
else:
return None
return pd.DataFrame(X_red, columns=['x', 'y', 'z'], index=df.index)
def display_closestwords_tsnescatterplot(arg_path_to_model, word):
model = word2vec.Word2Vec.load(arg_path_to_model)
for i in range(len(word)):
arr = np.empty((0, 300), dtype='f')
word_labels = [word[i]]
# get close words
close_words = model.similar_by_word(word[i])
# add the vector for each of the closest words to the array
arr = np.append(arr, np.array([model[word[i]]]), axis=0)
for wrd_score in close_words:
wrd_vector = model[wrd_score[0]]
word_labels.append(wrd_score[0])
arr = np.append(arr, np.array([wrd_vector]), axis=0)
# find tsne coords for 2 dimensions
tsne = TSNE(n_components=2, random_state=0)
np.set_printoptions(suppress=True)
Y = tsne.fit_transform(arr)
x_coords = Y[:, 0]
y_coords = Y[:, 1]
# display scatter plot
plt.scatter(x_coords, y_coords)
for label, x, y in zip(word_labels, x_coords, y_coords):
plt.annotate(label, xy=(x, y), xytext=(0, 0), textcoords='offset points')
# Zmiana mnoznika powoduje zmiane 'przyblizenia' wykresu (mniejszy mnoznik = wieksze przyblizenie)
plt.xlim(x_coords.min()*1, x_coords.max()*1)
plt.ylim(y_coords.min()*1, y_coords.max()*1)
plt.show()
def tsne_image(
features, images, img_res=64, res=4000, background_color=255, max_feature_size=-1, labels=None, point_radius=20, n_threads=0
):
"""
Embeds images via tsne into a scatter plot.
Parameters
---------
features: numpy array
Features to visualize
images: list or numpy array
Corresponding images to features.
img_res: int
Resolution to embed images at
res: int
Size of embedding image in pixels
background_color: float or numpy array
Background color value
max_feature_size: int
If input_feature_size > max_feature_size> 0, features are first
reduced using PCA to the desired size.
point_radius: int
Size of the circle for the label image.
n_threads: int
Number of threads to use for t-SNE
labels: List or numpy array if provided
Label for each image for drawing circle image.
"""
features = np.asarray(features, dtype=np.float32)
assert len(features.shape) == 2
print("Starting TSNE")
s_time = time.time()
if 0 < max_feature_size < features.shape[-1]:
pca = PCA(n_components=max_feature_size)
features = pca.fit_transform(features)
if n_threads <= 0:
n_threads = multiprocessing.cpu_count()
model = TSNE(n_components=2, verbose=1, random_state=0, n_jobs=n_threads)
f2d = model.fit_transform(features)
print("TSNE done.", (time.time() - s_time))
print("Starting drawing.")
x_coords = f2d[:, 0]
y_coords = f2d[:, 1]
return image_util.draw_images_at_locations(images, x_coords, y_coords, img_res, res, background_color, labels, point_radius)
def compute_tsne(X, y, n_class=2,
savepath=None,
xlim=(-50,50), ylim=(-50,50),
cls_lbl=['Benign','Tumor'],
title=' ',PCADIM=50):
tsne = TSNE(n_jobs=4, random_state=1337)
#X = PCA(n_components=PCADIM).fit_transform(X)
embs = tsne.fit_transform(X)
plt.figure(figsize=(10,10))
for i in range(n_class):
inds = np.where(y == i)[0]
plt.scatter(embs[inds, 0], embs[inds, 1], color=colors[i], marker='*', s=30)
if xlim:
plt.xlim(xlim[0], xlim[1])
if ylim:
plt.ylim(ylim[0], ylim[1])
plt.legend(cls_lbl)
plt.grid(b=None)
plt.title(title)
if savepath:
plt.savefig(savepath, dpi=300, bbox_inches='tight')
plt.savefig(savepath.replace('.png','.pdf'), dpi=300, bbox_inches='tight')
else:
plt.show()
plt.clf()
def tsne_main(args):
verbose_print(args, f'Loaded niche labels from {args.labels}')
labels = np.load(args.labels)
verbose_print(args, f'Running t-SNE based on {args.proximity}')
proximities = np.load(args.proximity)
x_tsne = TSNE(n_components=2, n_jobs=-1, perplexity=800,
learning_rate=100).fit_transform(proximities)
if args.plot:
# Show tSNE
for i in range(4):
idx = np.where(labels == i)[0]
if len(idx) == 0:
continue
plt.plot(x_tsne[idx, 0], x_tsne[idx, 1], '.', label=f'Cluster {i}')
plt.legend()
plt.show()
# Save the t-SNE coordinates
np.save(args.tsne, x_tsne)
verbose_print(args, f't-SNE coordinates saved to {args.tsne}')
verbose_print(args, f'Niche clustering done!')
def calc_tsne(
X,
n_jobs,
n_components,
perplexity,
early_exaggeration,
learning_rate,
random_state,
init="random",
n_iter=1000,
n_iter_early_exag=250,
):
"""
TODO: Typing
"""
tsne = TSNE(
n_jobs=n_jobs,
n_components=n_components,
perplexity=perplexity,
early_exaggeration=early_exaggeration,
learning_rate=learning_rate,
random_state=random_state,
verbose=1,
init=init,
n_iter=n_iter,
n_iter_early_exag=n_iter_early_exag,
)
X_tsne = tsne.fit_transform(X)
logger.info("Final error = {}".format(tsne.kl_divergence_))
return X_tsne
def plot_conti_code_tsne():
data = pickle.load(
open(
"/home/patrick/repositories/hyperspectral_phenotyping_gan/experiments_{}/generated_code_noise{}_disc{}_conti{}_epoch{}.p"
.format(opt.dataset, opt.n_noise, opt.n_dis, opt.n_conti,
opt.epoch), "rb"))
labels = np.array(data["y"]).squeeze()
labels_unique = np.unique(labels)
code = np.array(data["z"]).copy()
z = np.array(data["z"]).copy()
# print(code[0])
# code = code[:, -5:-2]
code = code[:, -2:]
# print(code[0])
# 1 / 0
signatures = np.array(data["x"])
tsne = TSNE(n_jobs=26, n_components=2, learning_rate=100)
Y = tsne.fit_transform(code)
colors = ["red", "green", "blue"]
for idx, label in enumerate(labels_unique):
data_tsne = Y[labels == label]
plt.scatter(data_tsne[:, 0],
data_tsne[:, 1],
c=colors[idx],
alpha=0.3,
label=str(label))
plt.legend()
plt.show()
def main():
parser = argparse.ArgumentParser(description='main function parser')
parser.add_argument('--path',
type=str,
help='load file path',
required=True)
parser.add_argument('--dump_dir',
type=str,
help='dump directory',
default=None)
parser.add_argument('--size',
type=int,
default=1000,
help='embedding vector size')
args = parser.parse_args()
embeddings, labels = load(args.path, args.size)
output = args.path.split('/')[-1]
# # UMAP
# weights = umap.UMAP().fit_transform(embeddings)
# show(weights, labels, 'umap.svg')
# t-SNE
tsne_model = TSNE(n_components=2)
weights = tsne_model.fit_transform(embeddings)
show(weights, labels, f'graph/{output}.svg')
def tsne(codewords, label, num_of_class):
"""plot the T-SNE based on codewords and label
Params:
------------------
codewords: (num_of_samples, dims_of_feature) numpy array
codewords to be dimension reduction
label: (num_of_samples,) numpy array
data label
num_of_class: int
number of class
Returns:
------------------
None
"""
starter_time = time.time()
embeddings = TSNE(n_components=2, perplexity=50,
n_jobs=4).fit_transform(codewords)
vis_x = embeddings[:, 0]
vis_y = embeddings[:, 1]
plt.scatter(vis_x,
vis_y,
c=label,
cmap=plt.cm.get_cmap("jet", num_of_class),
marker='.',
s=100)
plt.colorbar(ticks=range(num_of_class))
plt.clim(-0.5, 9.5)
print('TSNE TIME: {} seconds'.format(time.time() - starter_time))
