def get_statistics(model, data_loader=None, sequences=False):
if data_loader is None:
data_loader = test_loader_MNIST()
model.eval()
num_errs = 0.0
num_examples = 0
results = np.zeros((10, model.output_size), dtype='int32')
with torch.no_grad():
for x, y in data_loader:
# x = x.to(DEVICE).view(-1, 1, 28, 28).float()
y = y.to(DEVICE)
predictions = 0
if sequences:
x = x.to(DEVICE).view(-1, model.n, 28 * 28).float()
outputs = model.forward_sequences(x)
_, predictions = outputs.max(dim=2)
else:
x = x.to(DEVICE).view(-1, 28 * 28).float()
outputs = model.forward(x)
_, predictions = outputs.data.max(dim=1)
for i in range(10):
x_ = predictions[y.data == i].cpu().numpy()
x_unique, x_unique_count = np.unique(x_, return_counts=True)
# x_unique_count = torch.stack([(x_ == x_u).sum() for x_u in x_unique])
for idx, occ in zip(x_unique, x_unique_count):
results[i, idx] += occ
num_errs += (predictions != y.data).sum().item()
num_examples += x.size(0)
model.train()
return results
def get_statistics(model,
data_loader=None,
sequences=False,
return_ngram=False):
if data_loader is None:
data_loader = test_loader_MNIST()
model.eval()
num_errs = 0.0
num_examples = 0
results = np.zeros((10, model.output_size), dtype='int32')
stats = defaultdict(lambda: defaultdict(int))
with torch.no_grad():
for x, y in data_loader:
# x = x.to(DEVICE).view(-1, 1, 28, 28).float()
y = y.to(DEVICE)
predictions = 0
if sequences:
x = x.to(DEVICE).view(-1, model.n, 28 * 28).float()
outputs = model.forward_sequences(x)
_, predictions = outputs.max(dim=2)
if return_ngram:
for i in model.ngram:
preds = predictions.cpu().numpy()[(y.cpu().numpy(
) == np.array(i)).prod(1).astype('bool')]
if len(preds) > 0:
idxs, cnts = np.unique(preds,
axis=0,
return_counts=True)
for idx, cnt in zip(idxs, cnts):
stats[tuple(i)][tuple(idx)] += cnt
else:
x = x.to(DEVICE).view(-1, 28 * 28).float()
outputs = model.forward(x)
_, predictions = outputs.data.max(dim=1)
for i in range(10):
x_ = predictions[y.data == i].cpu().numpy()
x_unique, x_unique_count = np.unique(x_, return_counts=True)
# x_unique_count = torch.stack([(x_ == x_u).sum() for x_u in x_unique])
for idx, occ in zip(x_unique, x_unique_count):
results[i, idx] += occ
num_errs += (predictions != y.data).sum().item()
num_examples += x.size(0)
model.train()
stats.default_factory = None
if return_ngram:
return results, stats
return results
from src.remote import mpl from src.training import SGD, SPDG from src.statistics import get_statistics import matplotlib.pyplot as plt torch.manual_seed(671470554) np.random.seed(70914308) # %% GENERATING DATASET ngram = Ngram(3) ngram[(0, 1, 2)] = 6. ngram[(1, 2, 3)] = 4. ngram.norm() data_loader = train_loader_MNIST() test_loader = test_loader_MNIST() sequence_loader = sequence_loader_MNIST(ngram, num_samples=40000) # %% REGULAR TRAINING (SGD) # model = Model(ngram) # model.to(DEVICE) # model.init_weights() # optimizer = torch.optim.Adam(model.primal.parameters()) # history = SGD(model, optimizer, data_loader, test_loader, num_epochs=1, log_every=50, test_every=1) # %% DUAL TRAINING model = Model(ngram, output_size=4) model.to(DEVICE) model.init_weights()
