def get_input_from_batch(batch, use_cuda):
batch_size = len(batch.enc_lens)
enc_batch = Variable(torch.from_numpy(batch.enc_batch).long())
enc_padding_mask = Variable(torch.from_numpy(batch.enc_padding_mask)).float()
enc_lens = batch.enc_lens
extra_zeros = None
enc_batch_extend_vocab = None
if config.pointer_gen:
enc_batch_extend_vocab = Variable(torch.from_numpy(batch.enc_batch_extend_vocab).long())
# max_art_oovs is the max over all the article oov list in the batch
if batch.max_art_oovs > 0:
extra_zeros = Variable(torch.zeros((batch_size, batch.max_art_oovs)))
c_t_1 = Variable(torch.zeros((batch_size, 2 * config.hidden_dim)))
coverage = None
if config.is_coverage:
coverage = Variable(torch.zeros(enc_batch.size()))
if use_cuda:
enc_batch = enc_batch.cuda()
enc_padding_mask = enc_padding_mask.cuda()
if enc_batch_extend_vocab is not None:
enc_batch_extend_vocab = enc_batch_extend_vocab.cuda()
if extra_zeros is not None:
extra_zeros = extra_zeros.cuda()
c_t_1 = c_t_1.cuda()
if coverage is not None:
coverage = coverage.cuda()
return enc_batch, enc_padding_mask, enc_lens, enc_batch_extend_vocab, extra_zeros, c_t_1, coverage
def _run_mst_decoding(batch_energy: torch.Tensor, lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
heads = []
head_tags = []
for energy, length in zip(batch_energy.detach().cpu(), lengths):
scores, tag_ids = energy.max(dim=0)
# Although we need to include the root node so that the MST includes it,
# we do not want any word to be the parent of the root node.
# Here, we enforce this by setting the scores for all word -> ROOT edges
# edges to be 0.
scores[0, :] = 0
# Decode the heads. Because we modify the scores to prevent
# adding in word -> ROOT edges, we need to find the labels ourselves.
instance_heads, _ = decode_mst(scores.numpy(), length, has_labels=False)
# Find the labels which correspond to the edges in the max spanning tree.
instance_head_tags = []
for child, parent in enumerate(instance_heads):
instance_head_tags.append(tag_ids[parent, child].item())
# We don't care what the head or tag is for the root token, but by default it's
# not necesarily the same in the batched vs unbatched case, which is annoying.
# Here we'll just set them to zero.
instance_heads[0] = 0
instance_head_tags[0] = 0
heads.append(instance_heads)
head_tags.append(instance_head_tags)
return torch.from_numpy(numpy.stack(heads)), torch.from_numpy(numpy.stack(head_tags))
def __init__(self, input_dim, hidden_dim, word_drop, word_alphabet_size, word_emb_dim, pretrain_word_emb=None, left2right=True, fix_word_emb=True, gpu=True, use_bias = True):
super(LatticeLSTM, self).__init__()
skip_direction = "forward" if left2right else "backward"
print ("build LatticeLSTM... ", skip_direction, ", Fix emb:", fix_word_emb, " gaz drop:", word_drop)
self.gpu = gpu
self.hidden_dim = hidden_dim
self.word_emb = nn.Embedding(word_alphabet_size, word_emb_dim)
if pretrain_word_emb is not None:
print ("load pretrain word emb...", pretrain_word_emb.shape)
self.word_emb.weight.data.copy_(torch.from_numpy(pretrain_word_emb))
else:
self.word_emb.weight.data.copy_(torch.from_numpy(self.random_embedding(word_alphabet_size, word_emb_dim)))
if fix_word_emb:
self.word_emb.weight.requires_grad = False
self.word_dropout = nn.Dropout(word_drop)
self.rnn = MultiInputLSTMCell(input_dim, hidden_dim)
self.word_rnn = WordLSTMCell(word_emb_dim, hidden_dim)
self.left2right = left2right
if self.gpu:
self.rnn = self.rnn.cuda()
self.word_emb = self.word_emb.cuda()
self.word_dropout = self.word_dropout.cuda()
self.word_rnn = self.word_rnn.cuda()
def get_hierarchical_features_till_spawn(self, actions, backtrack_steps=0):
action_length = len(actions)-1
pa, ca, pq_idx, cq_idx, ph_idx = flat_to_hierarchical_actions(actions)
target_pos_idx = action_length - backtrack_steps
controller_step = True
if target_pos_idx in pq_idx:
controller_step = False
pq_idx_pruned = [v for v in pq_idx if v <= target_pos_idx]
pa_pruned = pa[:len(pq_idx_pruned)+1]
images = self.get_frames(
self.episode_house,
self.episode_pos_queue,
preprocess=True)
raw_img_feats = self.cnn(
Variable(torch.FloatTensor(images)
.cuda())).data.cpu().numpy().copy()
controller_img_feat, controller_action_in = False, False
if controller_step == True:
controller_img_feat = torch.from_numpy(raw_img_feats[target_pos_idx].copy())
controller_action_in = pa_pruned[-1] - 2
planner_img_feats = torch.from_numpy(raw_img_feats[pq_idx_pruned].copy())
planner_actions_in = torch.from_numpy(np.array(pa_pruned[:-1]) - 1)
return planner_actions_in, planner_img_feats, controller_step, controller_action_in, controller_img_feat, self.episode_pos_queue[target_pos_idx]
def evaluate(attention_model,x_test,y_test):
"""
cv results
Args:
attention_model : {object} model
x_test : {nplist} x_test
y_test : {nplist} y_test
Returns:
cv-accuracy
"""
attention_model.batch_size = x_test.shape[0]
attention_model.hidden_state = attention_model.init_hidden()
x_test_var = Variable(torch.from_numpy(x_test).type(torch.LongTensor))
y_test_pred,_ = attention_model(x_test_var)
if bool(attention_model.type):
y_preds = torch.max(y_test_pred,1)[1]
y_test_var = Variable(torch.from_numpy(y_test).type(torch.LongTensor))
else:
y_preds = torch.round(y_test_pred.type(torch.DoubleTensor).squeeze(1))
y_test_var = Variable(torch.from_numpy(y_test).type(torch.DoubleTensor))
return torch.eq(y_preds,y_test_var).data.sum()/x_test_var.size(0)
def predict(self, dataset):
"""Predict target for dataset.
Parameters:
----------
dataset (dict): dictionary with the testing dataset -
X_wide_test, X_deep_test, target
Returns:
--------
array-like with the target for dataset
"""
X_w = Variable(torch.from_numpy(dataset.wide)).float()
X_d = Variable(torch.from_numpy(dataset.deep))
if use_cuda:
X_w, X_d = X_w.cuda(), X_d.cuda()
# set the model in evaluation mode so dropout is not applied
net = self.eval()
pred = net(X_w,X_d).cpu()
if self.method == "regression":
return pred.squeeze(1).data.numpy()
if self.method == "logistic":
return (pred > 0.5).squeeze(1).data.numpy()
if self.method == "multiclass":
_, pred_cat = torch.max(pred, 1)
return pred_cat.data.numpy()
def predict_proba(self, dataset):
"""Predict predict probability for dataset.
This method will only work with method logistic/multiclass
Parameters:
----------
dataset (dict): dictionary with the testing dataset -
X_wide_test, X_deep_test, target
Returns:
--------
array-like with the probability for dataset.
"""
X_w = Variable(torch.from_numpy(dataset.wide)).float()
X_d = Variable(torch.from_numpy(dataset.deep))
if use_cuda:
X_w, X_d = X_w.cuda(), X_d.cuda()
# set the model in evaluation mode so dropout is not applied
net = self.eval()
pred = net(X_w,X_d).cpu()
if self.method == "logistic":
pred = pred.squeeze(1).data.numpy()
probs = np.zeros([pred.shape[0],2])
probs[:,0] = 1-pred
probs[:,1] = pred
return probs
if self.method == "multiclass":
return pred.data.numpy()
def fit(self, X, Y, n_epoch=10, batch_size=128, en_shuffle=True):
global_step = 0
n_batch = len(X) / batch_size
total_steps = int(n_epoch * n_batch)
for epoch in range(n_epoch):
if en_shuffle:
shuffled = np.random.permutation(len(X))
X = X[shuffled]
Y = Y[shuffled]
state = None
for local_step, (X_batch, Y_batch) in enumerate(zip(self.gen_batch(X, batch_size),
self.gen_batch(Y, batch_size))):
y_batch = Y_batch.ravel()
inputs = torch.autograd.Variable(torch.from_numpy(X_batch.astype(np.int64)))
labels = torch.autograd.Variable(torch.from_numpy(y_batch.astype(np.int64)))
if (self.stateful) and (len(X_batch) == batch_size):
preds, state = self.forward(inputs, state)
state = (torch.autograd.Variable(state[0].data), torch.autograd.Variable(state[1].data))
else:
preds, _ = self.forward(inputs)
loss = self.criterion(preds, labels) # cross entropy loss
self.optimizer, lr = self.adjust_lr(self.optimizer, global_step, total_steps)
self.optimizer.zero_grad() # clear gradients for this training step
loss.backward() # backpropagation, compute gradients
self.optimizer.step() # apply gradients
global_step += 1
preds = torch.max(preds, 1)[1].data.numpy().squeeze()
acc = (preds == y_batch).mean()
if local_step % 100 == 0:
print ('Epoch [%d/%d] | Step [%d/%d] | Loss: %.4f | Acc: %.4f | LR: %.4f'
%(epoch+1, n_epoch, local_step, n_batch, loss.data[0], acc, lr))
def get_test_data(num_train=1000, num_test=500,
input_shape=(10,), output_shape=(2,),
classification=True, num_classes=2):
"""Generates test data to train a model on.
classification=True overrides output_shape
(i.e. output_shape is set to (1,)) and the output
consists in integers in [0, num_class-1].
Otherwise: float output with shape output_shape.
"""
samples = num_train + num_test
if classification:
y = np.random.randint(0, num_classes, size=(samples,))
X = np.zeros((samples,) + input_shape)
for i in range(samples):
X[i] = np.random.normal(loc=y[i], scale=0.7, size=input_shape)
else:
y_loc = np.random.random((samples,))
X = np.zeros((samples,) + input_shape)
y = np.zeros((samples,) + output_shape)
for i in range(samples):
X[i] = np.random.normal(loc=y_loc[i], scale=0.7, size=input_shape)
y[i] = np.random.normal(loc=y_loc[i], scale=0.7, size=output_shape)
return (th.from_numpy(X[:num_train]), th.from_numpy(y[:num_train])), \
(th.from_numpy(X[num_train:]), th.from_numpy(y[num_train:]))
def __iter__(self):
for batch in self.data:
batch_size = len(batch)
batch = list(zip(*batch))
if self.eval:
assert len(batch) == 7
else:
assert len(batch) == 9
context_len = max(len(x) for x in batch[0])
context_id = torch.LongTensor(batch_size, context_len).fill_(0)
context_order = torch.LongTensor(batch_size,context_len).fill_(0)
for i, doc in enumerate(batch[0]):
context_id[i, :len(doc)] = torch.LongTensor(doc)
context_order[i,:len(doc)] = torch.from_numpy(np.arange(1,len(doc)+1))
feature_len = len(batch[1][0][0])
context_feature = torch.Tensor(batch_size, context_len, feature_len).fill_(0)
for i, doc in enumerate(batch[1]):
for j, feature in enumerate(doc):
context_feature[i, j, :] = torch.Tensor(feature)
context_tag = torch.LongTensor(batch_size, context_len).fill_(0)
for i, doc in enumerate(batch[2]):
context_tag[i, :len(doc)] = torch.LongTensor(doc)
context_ent = torch.LongTensor(batch_size, context_len).fill_(0)
for i, doc in enumerate(batch[3]):
context_ent[i, :len(doc)] = torch.LongTensor(doc)
question_len = max(len(x) for x in batch[4])
question_id = torch.LongTensor(batch_size, question_len).fill_(0)
question_order = torch.LongTensor(batch_size,question_len).fill_(0)
for i, doc in enumerate(batch[4]):
question_id[i, :len(doc)] = torch.LongTensor(doc)
question_order[i,:len(doc)] = torch.from_numpy(np.arange(1,len(doc)+1))
context_mask = torch.eq(context_id, 0)
question_mask = torch.eq(question_id, 0)
if not self.eval:
y_s = torch.LongTensor(batch[5])
y_e = torch.LongTensor(batch[6])
text = list(batch[-2])
span = list(batch[-1])
if self.gpu:
context_id = context_id.pin_memory()
context_feature = context_feature.pin_memory()
context_tag = context_tag.pin_memory()
context_ent = context_ent.pin_memory()
context_mask = context_mask.pin_memory()
question_id = question_id.pin_memory()
question_mask = question_mask.pin_memory()
context_order = context_order.pin_memory()
question_order = question_order.pin_memory()
if self.eval:
yield (context_id, context_feature, context_tag, context_ent, context_mask,
question_id, question_mask, context_order, question_order, text, span)
else:
yield (context_id, context_feature, context_tag, context_ent, context_mask,
question_id, question_mask, context_order, question_order, y_s, y_e, text, span)
def im_detect(net, im):
blobs, im_scales = _get_blobs(im)
assert len(im_scales) == 1, "Only single-image batch implemented"
im_blob = blobs['data']
blobs['im_info'] = np.array(
[im_blob.shape[1], im_blob.shape[2], im_scales[0]], dtype=np.float32)
_, scores, bbox_pred, rois = net.test_image(blobs['data'],
blobs['im_info'])
boxes = rois[:, 1:5] / im_scales[0]
scores = np.reshape(scores, [scores.shape[0], -1])
bbox_pred = np.reshape(bbox_pred, [bbox_pred.shape[0], -1])
if cfg.TEST.BBOX_REG:
# Apply bounding-box regression deltas
box_deltas = bbox_pred
pred_boxes = bbox_transform_inv(
torch.from_numpy(boxes), torch.from_numpy(box_deltas)).numpy()
pred_boxes = _clip_boxes(pred_boxes, im.shape)
else:
# Simply repeat the boxes, once for each class
pred_boxes = np.tile(boxes, (1, scores.shape[1]))
return scores, pred_boxes
def load_pretrained_npy(faster_rcnn_model, fname):
params = np.load(fname).item()
# vgg16
vgg16_dict = faster_rcnn_model.rpn.features.state_dict()
for name, val in list(vgg16_dict.items()):
# # print name
# # print val.size()
# # print param.size()
if name.find('bn.') >= 0:
continue
i, j = int(name[4]), int(name[6]) + 1
ptype = 'weights' if name[-1] == 't' else 'biases'
key = 'conv{}_{}'.format(i, j)
param = torch.from_numpy(params[key][ptype])
if ptype == 'weights':
param = param.permute(3, 2, 0, 1)
val.copy_(param)
# fc6 fc7
frcnn_dict = faster_rcnn_model.state_dict()
pairs = {'fc6.fc': 'fc6', 'fc7.fc': 'fc7'}
for k, v in list(pairs.items()):
key = '{}.weight'.format(k)
param = torch.from_numpy(params[v]['weights']).permute(1, 0)
frcnn_dict[key].copy_(param)
key = '{}.bias'.format(k)
param = torch.from_numpy(params[v]['biases'])
frcnn_dict[key].copy_(param)
def interpolate(ae, gg, z1, z2, vocab,
steps=5, sample=None, maxlen=None):
"""
Interpolating in z space
Assumes that type(z1) == type(z2)
"""
if type(z1) == Variable:
noise1 = z1
noise2 = z2
elif type(z1) == torch.FloatTensor or type(z1) == torch.cuda.FloatTensor:
noise1 = Variable(z1, volatile=True)
noise2 = Variable(z2, volatile=True)
elif type(z1) == np.ndarray:
noise1 = Variable(torch.from_numpy(z1).float(), volatile=True)
noise2 = Variable(torch.from_numpy(z2).float(), volatile=True)
else:
raise ValueError("Unsupported input type (noise): {}".format(type(z1)))
# interpolation weights
lambdas = [x*1.0/(steps-1) for x in range(steps)]
gens = []
for L in lambdas:
gens.append(generate(ae, gg, (1-L)*noise1 + L*noise2,
vocab, sample, maxlen))
interpolations = []
for i in range(len(gens[0])):
interpolations.append([s[i] for s in gens])
return interpolations
def l2l_train(model, cluster_center, n_epoch=10000, trunc_step=10):
optimizer = optim.Adam(model.parameters(), lr=0.01)
M_all = Variable(torch.zeros(model.n_class, model.n_dim))
B_all = Variable(torch.zeros(model.n_class))
for epoch in range(n_epoch):
loss = 0
M_step, B_step = [], []
for step in range(trunc_step):
data = generate_data(cluster_center)
optimizer.zero_grad()
x, y = Variable(torch.from_numpy(data[0])).float(), Variable(torch.from_numpy(data[1]))
w, b = model(x)
M = Variable(torch.zeros(model.n_class_n, model.n_dim))
B = Variable(torch.zeros(model.n_class_n))
for k in range(model.n_class_n):
M[k] = torch.cat((w[:, 0][y == model.n_class_l + k].view(-1, 1),
w[:, 1][y == model.n_class_l + k].view(-1, 1)), 1).mean(0)
B[k] = b[y == model.n_class_l + k].mean()
if step == 0:
M_ = M
B_ = B
else:
M_ = step / (step + 1) * M_step[-1] + 1 / (step + 1) * M
B_ = step / (step + 1) * B_step[-1] + 1 / (step + 1) * B
M_step.append(M_)
B_step.append(B_)
pred = torch.mm(x, M_.t()) + B_.view(1, -1).expand_as(torch.mm(x, M_.t()))
loss += F.cross_entropy(pred, y)
loss.backward()
optimizer.step()
print('Train Epoch: {}\tLoss: {:.6f}'.format(epoch, loss.data[0]))
return M_all, B_all, cluster_center
def resize(original, um_sizes, desired_res):
""" Resize array originally of um_sizes size to have desired_res resolution.
We preserve the center of original and resized arrays exactly in the middle. We also
make sure resolution is exactly the desired resolution. Given these two constraints,
we cannot hold FOV of original and resized arrays to be exactly the same.
:param np.array original: Array to resize.
:param tuple um_sizes: Size in microns of the array (one per axis).
:param int or tuple desired_res: Desired resolution (um/px) for the output array.
:return: Output array (np.float32) resampled to the desired resolution. Size in pixels
is round(um_sizes / desired_res).
"""
import torch.nn.functional as F
# Create grid to sample in microns
grid = create_grid(um_sizes, desired_res) # d x h x w x 3
# Re-express as a torch grid [-1, 1]
um_per_px = np.array([um / px for um, px in zip(um_sizes, original.shape)])
torch_ones = np.array(um_sizes) / 2 - um_per_px / 2 # sample position of last pixel in original
grid = grid / torch_ones[::-1].astype(np.float32)
# Resample
input_tensor = torch.from_numpy(original.reshape(1, 1, *original.shape).astype(
np.float32))
grid_tensor = torch.from_numpy(grid.reshape(1, *grid.shape))
resized_tensor = F.grid_sample(input_tensor, grid_tensor, padding_mode='border')
resized = resized_tensor.numpy().squeeze()
return resized
def imgEncodeTorch(self, abimg):
abimg = abimg.cuda()
w, h = abimg.shape[1], abimg.shape[2]
label = torch.zeros((w*h, 313))
label = label.cuda()
(dists, indexes) = self.nbrs.kneighbors(
abimg.view(abimg.shape[0], -1).t(), self.NN)
dists = torch.from_numpy(dists).float().cuda()
indexes = torch.from_numpy(indexes).cuda()
weights = torch.exp(-dists**2/(2*self.sigma**2)).cuda()
weights = weights/torch.sum(weights, dim=1).view(-1, 1)
pixel_indexes = torch.Tensor.long(torch.arange(
start=0, end=abimg.shape[1]*abimg.shape[2])[:, np.newaxis])
pixel_indexes = pixel_indexes.cuda()
label[pixel_indexes, indexes] = weights
label = label.t().contiguous().view(313, w, h)
rebal_indexes = indexes[:, 0]
rebal_weights = self.weights[rebal_indexes]
rebal_weights = rebal_weights.view(w, h)
rebal_label = rebal_weights * label
return rebal_label
def l2l_validate(model, cluster_center, n_epoch=100):
val_accuracy = []
for epoch in range(n_epoch):
data_l = generate_data_l(cluster_center)
data_n = generate_data_n(cluster_center, model.n_class_n)
x_l, y_l = Variable(torch.from_numpy(data_l[0])).float(), Variable(
torch.from_numpy(data_l[1]))
x_n, y_n = Variable(torch.from_numpy(data_n[0])).float(), Variable(
torch.from_numpy(data_n[1]))
pred_ll, pred_nl, w, b = model(x_l, x_n)
M = Variable(torch.zeros(model.n_class_n, model.n_dim))
B = Variable(torch.zeros(model.n_class_n))
for k in range(model.n_class_n):
M[k] = torch.cat((w[:, 0][y_n == model.n_class_l + k].view(-1, 1),
w[:, 1][y_n == model.n_class_l + k].view(-1, 1)), 1).mean(0)
B[k] = b[y_n == model.n_class_l + k].mean()
pred_ln = torch.mm(x_l, M.t()) + B.view(1, -1).expand_as(torch.mm(x_l, M.t()))
pred_nn = torch.mm(x_n, M.t()) + B.view(1, -1).expand_as(torch.mm(x_n, M.t()))
pred = torch.cat((torch.cat((pred_ll, pred_nl)), torch.cat((pred_ln, pred_nn))), 1)
pred = pred.data.max(1)[1]
y = torch.cat((y_l, y_n))
accuracy = pred.eq(y.data).cpu().sum() * 1.0 / y.size()[0]
# print('accuracy: %.2f' % accuracy)
val_accuracy.append(accuracy)
acc_l = pred.eq(y.data).cpu()[0:100].sum() * 1.0 / 100
acc_n = pred.eq(y.data).cpu()[100:150].sum() * 1.0 / 50
print('accuracy: %.2f, lifelong accuracy: %.2f, new accuracy: %.2f' % (accuracy, acc_l, acc_n))
return numpy.mean(numpy.asarray(val_accuracy))
def __call__(self, pic):
if isinstance(pic, np.ndarray):
# handle numpy array
img = torch.from_numpy(pic.transpose((2, 0, 1)))
# backard compability
return img.float().div(255)
# handle PIL Image
if pic.mode == 'I':
img = torch.from_numpy(np.array(pic, np.int32, copy=False))
elif pic.mode == 'I;16':
img = torch.from_numpy(np.array(pic, np.int16, copy=False))
else:
img = torch.ByteTensor(torch.ByteStorage.from_buffer(pic.tobytes()))
# PIL image mode: 1, L, P, I, F, RGB, YCbCr, RGBA, CMYK
if pic.mode == 'YCbCr':
nchannel = 3
elif pic.mode == 'I;16':
nchannel = 1
else:
nchannel = len(pic.mode)
img = img.view(pic.size[1], pic.size[0], nchannel)
# put it from HWC to CHW format
# yikes, this transpose takes 80% of the loading time/CPU
img = img.transpose(0, 1).transpose(0, 2).contiguous()
if isinstance(img, torch.ByteTensor):
return img.float().div(255)
else:
return img
def collate_data(self, data):
video_tensor = np.zeros((len(data), 2048))
flow_tensor = np.zeros((len(data), 1024))
face_tensor = np.zeros((len(data), 128))
audio_tensor = np.zeros((len(data), self.max_words,128))
text_tensor = np.zeros((len(data), self.max_words, 300))
coco_ind = np.zeros((len(data)))
face_ind = np.zeros((len(data)))
for i in range(len(data)):
coco_ind[i] = data[i]['coco_ind']
face_ind[i] = data[i]['face_ind']
video_tensor[i] = data[i]['video']
flow_tensor[i] = data[i]['flow']
if len(data[i]['face']) > 0:
face_tensor[i] = data[i]['face']
la = len(data[i]['audio'])
audio_tensor[i,:min(la,self.max_words), :] = data[i]['audio'][:min(self.max_words,la)]
lt = len(data[i]['text'])
text_tensor[i,:min(lt,self.max_words), :] = data[i]['text'][:min(self.max_words,lt)]
return {'video': th.from_numpy(video_tensor).float(),
'flow': th.from_numpy(flow_tensor).float(),
'face': th.from_numpy(face_tensor).float(),
'coco_ind': coco_ind,
'face_ind': face_ind,
'text': th.from_numpy(text_tensor).float(),
'audio': th.from_numpy(audio_tensor).float()}
def corrupt(self, src, rel, dst):
prob = self.bern_prob[rel]
selection = torch.bernoulli(prob).numpy().astype('int64')
ent_random = choice(self.n_ent, len(src))
src_out = (1 - selection) * src.numpy() + selection * ent_random
dst_out = selection * dst.numpy() + (1 - selection) * ent_random
return torch.from_numpy(src_out), torch.from_numpy(dst_out)
def evalAccuracy(model, device, args, outPutType, test_X, test_Y, validEval = False) :
model_training_orig = model.training
setModelMode(model, False)
N_test = len(test_Y)*len(test_Y[0])
totals = 0
for b_data, b_labels in zip(test_X, test_Y):
b_data = torch.from_numpy(b_data).to(device)
b_labels = torch.from_numpy(b_labels).to(device)
b_labels = b_labels.view(b_labels.shape[0],-1 ) # make it the same shape as output
yhat = model(b_data) # need to compute the Yhat again, as this is the yhat AFTER updating the weights, not before as in 'learn()' function
b_data = None
# depending on if we are in a classification or regression problem, we evaluate performance differently
if outPutType == OUT_REGRESSION :
currentRate = torch_pearsonr( yhat.view(-1) , b_labels.view(-1))**2
N_test = len(test_Y) # as we are testing correlation, the N should refer to the number of batches, and NOT the total number of observations
elif outPutType == OUT_MULTICLASS :
currentRate = calc_Accuracy(yhat,b_labels ) # calculate accuracy, this is ROUNDED
else : # mean absolute error
currentRate = -torch.mean(torch.abs(yhat - b_labels)) # negative as the rest of the metrics are accuracy, IE the greater the error, the lower the accuracy
N_test = len(test_Y) # as we are testing correlation, the N should refer to the number of batches, and NOT the total number of observations
currentRate = float(currentRate.detach().cpu().numpy() ) # need to move it back to CPU
totals = totals +currentRate # sum in all minibatches
accuracy = round( float(totals)/N_test,5)
setModelMode(model, model_training_orig)
return(accuracy)
def __call__(self, inputs):
if isinstance(inputs, Image.Image):
inputs = torch.from_numpy(np.array(inputs).transpose(2, 0, 1))
else:
inputs = torch.from_numpy(inputs.transpose(2, 0, 1))
return inputs.float()
def average_without_padding(x, ids, padding_id, cuda=False, eps=1e-8):
if cuda:
mask = Variable(torch.from_numpy(np.not_equal(ids, padding_id).astype(int)[:,:,np.newaxis])).float().cuda().permute(1, 2, 0).expand_as(x)
else:
mask = Variable(torch.from_numpy(np.not_equal(ids, padding_id).astype(int)[:,:,np.newaxis])).float().permute(1, 2, 0).expand_as(x)
s = torch.sum(x*mask, dim=2) / (torch.sum(mask, dim=2)+eps)
return s
def run_test(self, args):
print("testing...")
self.eval()
game = DoomInstance(
args.vizdoom_config, args.wad_path, args.skiprate, visible=True, actions=args.action_set, bot_cmd=args.bot_cmd)
step_state = game.get_state_normalized()
state = NormalizedState(screen=None, depth=None, labels=None, variables=None)
state.screen = torch.Tensor(1, *args.screen_size)
state.variables = torch.Tensor(1, args.variable_num)
while True:
# convert state to torch tensors
state.screen[0, :] = torch.from_numpy(step_state.screen)
state.variables[0, :] = torch.from_numpy(step_state.variables)
# compute an action
action = self.get_action(state)
# render
step_state, _, finished = game.step_normalized(action[0][0])
#img = step_state.screen[0:3, :]
#img = img.transpose(1, 2, 0)
#plt.imsave('depth-plan.png', img)
if finished:
print("episode return: {}".format(game.get_episode_return()))
self.set_terminal(torch.zeros(1))
def flow_resnet50_aux(pretrained=False, **kwargs):
"""Constructs a ResNet-50 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
if pretrained:
# model.load_state_dict(model_zoo.load_url(model_urls['resnet50']))
pretrained_dict = model_zoo.load_url(model_urls['resnet50'])
model_dict = model.state_dict()
fc_origin_weight = pretrained_dict["fc.weight"].data.numpy()
fc_origin_bias = pretrained_dict["fc.bias"].data.numpy()
# 1. filter out unnecessary keys
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
# 2. overwrite entries in the existing state dict
model_dict.update(pretrained_dict)
# print(model_dict)
fc_new_weight = model_dict["fc_aux.weight"].numpy()
fc_new_bias = model_dict["fc_aux.bias"].numpy()
fc_new_weight[:1000, :] = fc_origin_weight
fc_new_bias[:1000] = fc_origin_bias
model_dict["fc_aux.weight"] = torch.from_numpy(fc_new_weight)
model_dict["fc_aux.bias"] = torch.from_numpy(fc_new_bias)
# 3. load the new state dict
model.load_state_dict(model_dict)
return model
def train():
num_classses = 21
net = ssd.SSD300(num_classes=num_classses)
ssd_box_coder = ssd.SSDBoxCoder(net)
C, H, W = (3, 300, 300)
x = Variable(torch.randn(1, C, H, W))
boxes = torch.from_numpy(np.array([(0, 0, 100, 100), (25, 25, 125, 125), (200, 200, 250, 250), (0, 0, 300, 300)], dtype=np.float32))
labels = torch.from_numpy(np.array([1, 2, 3, 4], dtype=np.long))
loc_targets, cls_targets = ssd_box_coder.encode(boxes, labels)
loc_targets = loc_targets[None, :]
cls_targets = cls_targets[None, :]
# print('loc_targets.size():{}'.format(loc_targets.size()))
# print('cls_targets.size():{}'.format(cls_targets.size()))
optimizer = optim.SGD(net.parameters(), lr=1e-5, momentum=0.9, weight_decay=5e-4)
criterion = ssd.SSDLoss(num_classes=num_classses)
for epoch in range(100):
loc_preds, cls_preds = net(x)
# print('loc_preds.size():{}'.format(loc_preds.size()))
# print('cls_preds.size():{}'.format(cls_preds.size()))
optimizer.zero_grad()
loss = criterion(loc_preds, loc_targets, cls_preds, cls_targets)
loss.backward()
optimizer.step()
def __init__(self, height, width, lr = 1, aux_loss = False, ray_tracing = False):
super(Depth3DGridGen_with_mask, self).__init__()
self.height, self.width = height, width
self.aux_loss = aux_loss
self.lr = lr
self.ray_tracing = ray_tracing
self.grid = np.zeros( [self.height, self.width, 3], dtype=np.float32)
self.grid[:,:,0] = np.expand_dims(np.repeat(np.expand_dims(np.arange(-1, 1, 2.0/self.height), 0), repeats = self.width, axis = 0).T, 0)
self.grid[:,:,1] = np.expand_dims(np.repeat(np.expand_dims(np.arange(-1, 1, 2.0/self.width), 0), repeats = self.height, axis = 0), 0)
self.grid[:,:,2] = np.ones([self.height, width])
self.grid = torch.from_numpy(self.grid.astype(np.float32))
self.theta = self.grid[:,:,0] * np.pi/2 + np.pi/2
self.phi = self.grid[:,:,1] * np.pi
self.x = torch.sin(self.theta) * torch.cos(self.phi)
self.y = torch.sin(self.theta) * torch.sin(self.phi)
self.z = torch.cos(self.theta)
self.grid3d = torch.from_numpy(np.zeros( [self.height, self.width, 4], dtype=np.float32))
self.grid3d[:,:,0] = self.x
self.grid3d[:,:,1] = self.y
self.grid3d[:,:,2] = self.z
self.grid3d[:,:,3] = self.grid[:,:,2]
def artist_works_with_labels(): # painting from the famous artist (real target)
a = np.random.uniform(1, 2, size=BATCH_SIZE)[:, np.newaxis]
paintings = a * np.power(PAINT_POINTS, 2) + (a-1)
labels = (a-1) > 0.5 # upper paintings (1), lower paintings (0), two classes
paintings = torch.from_numpy(paintings).float()
labels = torch.from_numpy(labels.astype(np.float32))
return paintings, labels
def transform(self, img, lbl):
"""transform
:param img:
:param lbl:
"""
img = img[:, :, ::-1]
img = img.astype(np.float64)
img -= self.mean
img = m.imresize(img, (self.img_size[0], self.img_size[1]))
# Resize scales images from 0 to 255, thus we need
# to divide by 255.0
img = img.astype(float) / 255.0
# NHWC -> NCWH
img = img.transpose(2, 0, 1)
classes = np.unique(lbl)
lbl = lbl.astype(float)
lbl = m.imresize(lbl, (self.img_size[0], self.img_size[1]), 'nearest', mode='F')
lbl = lbl.astype(int)
if not np.all(classes == np.unique(lbl)):
print("WARN: resizing labels yielded fewer classes")
if not np.all(np.unique(lbl[lbl!=self.ignore_index]) < self.n_classes):
print('after det', classes, np.unique(lbl))
raise ValueError("Segmentation map contained invalid class values")
img = torch.from_numpy(img).float()
lbl = torch.from_numpy(lbl).long()
return img, lbl
def data_generator(args, screens, variables, labels, episodes):
# remove short episodes
episode_min_size = args.episode_size*args.skiprate
episodes = episodes[episodes[:, 1]-episodes[:, 0] > episode_min_size]
episodes_num = len(episodes)
#
batch_size = args.batch_size
step_idx = episodes[:, 0].copy() + np.random.randint(args.skiprate, size=episodes_num)
step_screens = np.ndarray(shape=(batch_size, *screens.shape[1:]), dtype=np.float32)
step_variables = np.ndarray(shape=(batch_size, *variables.shape[1:]), dtype=np.float32)
step_labels = np.ndarray(shape=(batch_size,), dtype=np.int)
step_terminals = np.ones(shape=(batch_size, 1), dtype=np.float32)
# select episodes for the initial batch
batch_episodes = np.random.randint(episodes_num, size=batch_size)
while True:
for i in range(batch_size):
idx = batch_episodes[i]
step_screens[i, :] = screens[step_idx[idx]] / 127.5 - 1.0
step_variables[i, :] = variables[step_idx[idx]] / 100
step_labels[i] = labels[step_idx[idx]]
step_idx[idx] += args.skiprate
if step_idx[idx] > episodes[idx][1]:
step_idx[idx] = episodes[idx][0] + np.random.randint(args.skiprate)
step_terminals[i] = 0
# reached terminal state, select a new episode
batch_episodes[i] = np.random.randint(episodes_num)
else:
step_terminals[i] = 1
yield torch.from_numpy(step_screens), \
torch.from_numpy(step_variables), \
torch.from_numpy(step_labels), \
torch.from_numpy(step_terminals)
pred, uncert = sgp.predict(X_train, 0 * X_train)
error = np.sqrt(np.mean((pred - y_train)**2))
trainll = np.mean(sps.norm.logpdf(pred - y_train, scale=np.sqrt(uncert)))
print('Train RMSE: ', error)
print('Train ll: ', trainll)
# We pick the next 60 inputs
next_inputs = sgp.batched_greedy_ei(60, np.min(X_train, 0),
np.max(X_train, 0))
valid_smiles = []
new_features = []
for i in range(60):
all_vec = next_inputs[i].reshape((1, -1))
tree_vec, mol_vec = np.hsplit(all_vec, 2)
tree_vec = create_var(torch.from_numpy(tree_vec).float())
mol_vec = create_var(torch.from_numpy(mol_vec).float())
s = model.decode(tree_vec, mol_vec, prob_decode=False)
if s is not None:
valid_smiles.append(s)
new_features.append(all_vec)
print(len(valid_smiles), "molecules are found")
valid_smiles = valid_smiles[:50]
new_features = next_inputs[:50]
new_features = np.vstack(new_features)
save_object(valid_smiles,
opts.save_dir + "/valid_smiles{}.dat".format(iteration))
import sascorer
import networkx as nx
def run(self, img):
""" Process a numpy image to extract points and descriptors.
Input
img - HxW numpy float32 input image in range [0,1].
Output
corners - 3xN numpy array with corners [x_i, y_i, confidence_i]^T.
desc - 256xN numpy array of corresponding unit normalized descriptors.
heatmap - HxW numpy heatmap in range [0,1] of point confidences.
"""
assert img.ndim == 2, 'Image must be grayscale.'
assert img.dtype == np.float32, 'Image must be float32.'
H, W = img.shape[0], img.shape[1]
inp = img.copy()
inp = (inp.reshape(1, H, W))
inp = torch.from_numpy(inp)
inp = torch.autograd.Variable(inp).view(1, 1, H, W)
if self.cuda:
inp = inp.cuda()
# Forward pass of network.
outs = self.net.forward(inp)
semi = outs
# Convert pytorch -> numpy.
semi = semi.data.cpu().numpy().squeeze()
# --- Process points.
dense = np.exp(semi) # Softmax.
dense = dense / (np.sum(dense, axis=0) + .00001) # Should sum to 1.
# Remove dustbin.
nodust = dense[:-1, :, :]
# Reshape to get full resolution heatmap.
Hc = int(H / self.cell)
Wc = int(W / self.cell)
nodust = nodust.transpose(1, 2, 0)
heatmap = np.reshape(nodust, [Hc, Wc, self.cell, self.cell])
heatmap = np.transpose(heatmap, [0, 2, 1, 3])
heatmap = np.reshape(heatmap, [Hc * self.cell, Wc * self.cell])
xs, ys = np.where(heatmap >= self.conf_thresh) # Confidence threshold.
if len(xs) == 0:
return np.zeros((3, 0)), None
pts = np.zeros((3, len(xs))) # Populate point data sized 3xN.
pts[0, :] = ys
pts[1, :] = xs
pts[2, :] = heatmap[xs, ys]
pts, _ = self.nms_fast(pts, H, W,
dist_thresh=self.nms_dist) # Apply NMS.
inds = np.argsort(pts[2, :])
pts = pts[:, inds[::-1]] # Sort by confidence.
# Remove points along border.
if self.border_remove != 0:
bord = self.border_remove
toremoveW = np.logical_or(pts[0, :] < bord, pts[0, :] >=
(W - bord))
toremoveH = np.logical_or(pts[1, :] < bord, pts[1, :] >=
(H - bord))
toremove = np.logical_or(toremoveW, toremoveH)
pts = pts[:, ~toremove]
# --- Process descriptor.
# D = coarse_desc.shape[1]
# if pts.shape[1] == 0:
# desc = np.zeros((D, 0))
# else:
# # Interpolate into descriptor map using 2D point locations.
# samp_pts = torch.from_numpy(pts[:2, :].copy())
# samp_pts[0, :] = (samp_pts[0, :] / (float(W) / 2.)) - 1.
# samp_pts[1, :] = (samp_pts[1, :] / (float(H) / 2.)) - 1.
# samp_pts = samp_pts.transpose(0, 1).contiguous()
# samp_pts = samp_pts.view(1, 1, -1, 2)
# samp_pts = samp_pts.float()
# if self.cuda:
# samp_pts = samp_pts.cuda()
# desc = torch.nn.functional.grid_sample(coarse_desc, samp_pts)
# desc = desc.data.cpu().numpy().reshape(D, -1)
# desc /= np.linalg.norm(desc, axis=0)[np.newaxis, :]
return pts, heatmap
if image.shape[0] > image.shape[1]:
image = imutils.resize(image, width=450)
else:
image = imutils.resize(image, width=900)
original_height = image.shape[0]
original_width = image.shape[1]
resized_image = utils.resize_image(image, model.WIDTH)
resized_image = np.expand_dims(resized_image, 0)
start = time.time()
tf.logging.info('Predicting {}...'.format(
os.path.basename(image_list[index])))
detections = sess.run(predictions, feed_dict={inputs: resized_image})
tf.logging.info('Performing non-maximum suppression...')
result = utils.torch_non_max_suppression(
detections=torch.from_numpy(detections),
confidence_threshold=0.25,
num_classes=model.NUM_CLASSES,
nms_conf=0.4)
end = time.time()
tf.logging.info('Total time: {}'.format(end - start))
if result is not None:
for box in result:
p1 = (int(box[1]), int(box[2]))
p1 = utils.convert_to_original_size(p1, original_height,
original_width,
model.HEIGHT, model.WIDTH)
p2 = (int(box[3]), int(box[4]))
p2 = utils.convert_to_original_size(p2, original_height,
original_width,
for i in range(future): # if we should predict the future
h_t, c_t = self.lstm1(output, (h_t, c_t))
h_t2, c_t2 = self.lstm2(h_t, (h_t2, c_t2))
output = self.linear(h_t2)
outputs += [output]
outputs = torch.stack(outputs, 1).squeeze(2)
return outputs
if __name__ == '__main__':
# set random seed to 0
np.random.seed(0)
torch.manual_seed(0)
# load data and make training set
data = torch.load('traindata.pt')
input = Variable(torch.from_numpy(data[3:, :-1]), requires_grad=False)
target = Variable(torch.from_numpy(data[3:, 1:]), requires_grad=False)
test_input = Variable(torch.from_numpy(data[:3, :-1]), requires_grad=False)
test_target = Variable(torch.from_numpy(data[:3, 1:]), requires_grad=False)
# build the model
seq = Sequence()
seq.double()
criterion = nn.MSELoss()
# use LBFGS as optimizer since we can load the whole data to train
optimizer = optim.LBFGS(seq.parameters(), lr=0.8)
#begin to train
for i in range(15):
print('STEP: ', i)
def closure():
optimizer.zero_grad()
def __call__(self, sample):
sample['img'] = torch.from_numpy(sample['img'])
if sample['annot'] is not None:
sample['annot'] = torch.from_numpy(sample['annot'])
return sample
def subsequent_mask(size):
"Mask out subsequent positions."
attn_shape = (1, size, size)
# Note: Here np.triu is upper triabgle of an array
subsequent_mask = np.triu(np.ones(attn_shape), k=1).astype('uint8')
return torch.from_numpy(subsequent_mask) == 0
def get_data():
x_test, labels = [], []
data = np.genfromtxt("list_attr_celeba.csv",
dtype=float,
delimiter=',',
names=True)
print(data.shape)
train_labels = []
test_labels = []
train_paths = []
test_paths = []
n = 0
for row in data:
if n < 10000:
labels = np.array([
int(row['Eyeglasses']),
int(row['Bangs']),
int(row['Black_Hair']),
int(row['Blond_Hair']),
int(row['Brown_Hair']),
int(row['Gray_Hair']),
int(row['Male']),
int(row['Pale_Skin']),
int(row['Smiling']),
int(row['Young'])
])
train_labels.append(labels)
train_paths.append(row['image_id'])
elif n < 12000:
labels = np.array([
int(row['Eyeglasses']),
int(row['Bangs']),
int(row['Black_Hair']),
int(row['Blond_Hair']),
int(row['Brown_Hair']),
int(row['Gray_Hair']),
int(row['Male']),
int(row['Pale_Skin']),
int(row['Smiling']),
int(row['Young'])
])
test_labels.append(labels)
test_paths.append(row['image_id'])
else:
break
n = n + 1
y_train = train_labels
for row in train_paths:
img = Image.open('./img_align_celeba/' + train)
img = numpy.array(img.resize((48, 48)))
x_train.append(img)
x_train = torch.from_numpy(np.array(x_train) / 255)
y_train = torch.from_numpy(np.array(y_train))
print(x_train.size())
print(y_train.size())
pickle.dump(x_train, open("pic_tensor.p", "wb"))
pickle.dump(x_train, open("pic_tensor.p", "wb"))
return x_train.to(device), y_train.to(device)
print(model)
par_dict = {}
for par in model.named_parameters():
print("--------------------------------------------")
print("setting [",par[0],par[1].shape,"]")
if "weight" in par[0] and len(par[0].split("."))==3 and len(par[1].shape)==4:
layerid1 = int(par[0].split(".")[0])
layerid2 = int(par[0].split(".")[1])
vgglayername = "features.%d.weight"%(layerid2)
try:
np_vggweight = model_data[vgglayername].numpy()
except:
continue
vggweight = np.transpose( np_vggweight, (2,3,1,0) )
vggweight = torch.from_numpy( vggweight.reshape( (9,1,vggweight.shape[2],vggweight.shape[3]) ) ).to(device)
print(par[0],par[1].shape,'-->',vgglayername,vggweight.shape)
par_dict[par[0]] = vggweight[:,:,:par[1].shape[2],:par[1].shape[3]]
else:
par_dict[par[0]] = par[1]
#print(par[1])
if "weight" in par[0] and len(par[0].split("."))==3 and len(par[1].shape)==1:
par_dict[par[0].replace("weight","running_var")] = torch.ones( (par[1].shape[0]) )
if "bias" in par[0] and len(par[0].split("."))==3 and len(par[1].shape)==1:
par_dict[par[0].replace("bias","running_mean")] = torch.zeros( (par[1].shape[0]) )
#print(par_dict)
#par_dict["2.running_mean"] = torch.zeros( (32) )
#par_dict["2.running_var"] = torch.ones( (32) )
model.load_state_dict(par_dict)
def load_tf_weights_in_electra(model,
config,
tf_checkpoint_path,
discriminator_or_generator="discriminator"):
""" Load tf checkpoints in a pytorch model.
"""
try:
import re
import numpy as np
import tensorflow as tf
except ImportError:
logger.error(
"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions."
)
raise
tf_path = os.path.abspath(tf_checkpoint_path)
logger.info("Converting TensorFlow checkpoint from {}".format(tf_path))
# Load weights from TF model
init_vars = tf.train.list_variables(tf_path)
names = []
arrays = []
for name, shape in init_vars:
logger.info("Loading TF weight {} with shape {}".format(name, shape))
array = tf.train.load_variable(tf_path, name)
names.append(name)
arrays.append(array)
for name, array in zip(names, arrays):
original_name: str = name
try:
if isinstance(model, ElectraForMaskedLM):
name = name.replace("electra/embeddings/",
"generator/embeddings/")
if discriminator_or_generator == "generator":
name = name.replace("electra/", "discriminator/")
name = name.replace("generator/", "electra/")
name = name.replace("dense_1", "dense_prediction")
name = name.replace("generator_predictions/output_bias",
"generator_lm_head/bias")
name = name.split("/")
# print(original_name, name)
# adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
# which are not required for using pretrained model
if any(n in ["global_step", "temperature"] for n in name):
logger.info("Skipping {}".format(original_name))
continue
pointer = model
for m_name in name:
if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
scope_names = re.split(r"_(\d+)", m_name)
else:
scope_names = [m_name]
if scope_names[0] == "kernel" or scope_names[0] == "gamma":
pointer = getattr(pointer, "weight")
elif scope_names[0] == "output_bias" or scope_names[
0] == "beta":
pointer = getattr(pointer, "bias")
elif scope_names[0] == "output_weights":
pointer = getattr(pointer, "weight")
elif scope_names[0] == "squad":
pointer = getattr(pointer, "classifier")
else:
pointer = getattr(pointer, scope_names[0])
if len(scope_names) >= 2:
num = int(scope_names[1])
pointer = pointer[num]
if m_name.endswith("_embeddings"):
pointer = getattr(pointer, "weight")
elif m_name == "kernel":
array = np.transpose(array)
try:
assert pointer.shape == array.shape, original_name
except AssertionError as e:
e.args += (pointer.shape, array.shape)
raise
print("Initialize PyTorch weight {}".format(name), original_name)
pointer.data = torch.from_numpy(array)
except AttributeError as e:
print("Skipping {}".format(original_name), name, e)
continue
return model
def test(self):
import onnx
from ngraph_onnx.onnx_importer.importer import import_onnx_model
import ngraph as ng
global dim0, dim2, dim3
torch.set_grad_enabled(False)
epoch = self.optimizer.get_last_epoch() + 1
self.ckp.write_log('\nEvaluation:')
self.ckp.add_log(
torch.zeros(1, len(self.loader_test), len(self.scale))
)
self.model.eval()
timer_test = utility.timer()
if self.args.save_results: self.ckp.begin_background()
# print(self.loader_test)
for idx_data, d in enumerate(self.loader_test):
for idx_scale, scale in enumerate(self.scale):
d.dataset.set_scale(idx_scale)
print('idx_scale={}'.format(idx_scale))
# print("len: {}".format(len(d)))
# for lr, hr, filename, _ in tqdm(d, ncols=80):
for batch, (lr, hr, filename, _) in enumerate(d):
print('{} '.format(batch), end='', flush=True)
lr, hr = self.prepare(lr, hr)
print('test lr.size: {}'.format(lr.size()))
dim0 = lr.size()[0]
dim2 = lr.size()[2]
dim3 = lr.size()[3]
showbug = False
if showbug: print('stage1', flush=True)
if self.args.ngraph:
pytorch_model_name = self.args.ngraph
pytorch_edsr_model = torch.load(pytorch_model_name).cuda()
if showbug: print('stage2-1', flush=True)
# print(lr.size())
# dummy_input = torch.randn_like(lr, device='cuda')
if showbug: print('stage2-2', flush=True)
edsr_onnx_filename = '{}.onnx'.format(pytorch_model_name)
# print('Export to onnx model {}'.format(edsr_onnx_filename))
torch.onnx.export(pytorch_edsr_model, lr.to(torch.device('cuda')), edsr_onnx_filename, export_params=True, verbose=False, training=False)
if showbug: print('stage2-3', flush=True)
edsr_onnx_model = onnx.load(edsr_onnx_filename)
# print(onnx.helper.printable_graph(edsr_onnx_model.graph))
if showbug: print('stage2-4', flush=True)
ng_models = import_onnx_model(edsr_onnx_model)
# print('Convert to nGreph Model')
ng_model = ng_models[0]
if showbug: print('stage2-5', flush=True)
runtime = ng.runtime(backend_name='CPU')
if showbug: print('stage2-6', flush=True)
edsr_ng_model = runtime.computation(ng_model['output'], *ng_model['inputs'])
if showbug: print('stage2-7', flush=True)
sr = edsr_ng_model(lr, idx_scale)
if showbug: print('stage2-8', flush=True)
sr = torch.from_numpy(sr)
if showbug: print('stage2-9', flush=True)
elif self.args.tensorrt:
pytorch_model_name = self.args.tensorrt
pytorch_edsr_model = torch.load(pytorch_model_name)
# lr_np = lr.numpy().astype(np.float32)
dummy_input = torch.randn_like(lr, device='cuda')
edsr_onnx_filename = '{}.onnx'.format(pytorch_model_name)
print('Export to onnx model {}'.format(edsr_onnx_filename))
torch.onnx.export(pytorch_edsr_model, dummy_input, edsr_onnx_filename, export_params=True, verbose=False, training=False)
import os
import onnx
edsr_onnx_model = onnx.load(edsr_onnx_filename)
# print(onnx.helper.printable_graph(edsr_onnx_model.graph))
import tensorrt
import onnx_tensorrt.backend as backend
import numpy as np
tensorrt_engine = backend.prepare(edsr_onnx_model, device='CUDA:0')
# lr_np = lr_np.to(torch.device("cuda:0"))
# lr.numpy().astype(np.float32)
sr = tensorrt_engine.run(lr.numpy().astype(np.float32))[0]
sr = torch.from_numpy(sr)
print('complete one')
pytorch_model_name = self.args.tensorrt
pytorch_edsr_model = torch.load(pytorch_model_name)
# lr_np = lr.numpy().astype(np.float32)
dummy_input = torch.randn_like(lr, device='cuda')
edsr_onnx_filename = '{}.onnx'.format(pytorch_model_name)
print('Export to onnx model {}'.format(edsr_onnx_filename))
torch.onnx.export(pytorch_edsr_model, dummy_input, edsr_onnx_filename, export_params=True, verbose=False, training=False)
import os
import onnx
edsr_onnx_model = onnx.load(edsr_onnx_filename)
# print(onnx.helper.printable_graph(edsr_onnx_model.graph))
import tensorrt
import onnx_tensorrt.backend as backend
import numpy as np
tensorrt_engine = backend.prepare(edsr_onnx_model, device='CUDA:0')
# lr_np = lr_np.to(torch.device("cuda:0"))
# lr.numpy().astype(np.float32)
sr = tensorrt_engine.run(lr.numpy().astype(np.float32))[0]
sr = torch.from_numpy(sr)
print('complete two')
else:
sr = self.model(lr, idx_scale)
if showbug: print('stage3', flush=True)
sr = utility.quantize(sr, self.args.rgb_range)
if showbug: print('stage4', flush=True)
save_list = [sr]
if showbug: print('stage5', flush=True)
self.ckp.log[-1, idx_data, idx_scale] += utility.calc_psnr(
sr, hr, scale, self.args.rgb_range, dataset=d
)
if showbug: print('stage6', flush=True)
if self.args.save_gt:
save_list.extend([lr, hr])
if showbug: print('stage7', flush=True)
if self.args.save_results:
self.ckp.save_results(d, filename[0], save_list, scale)
if showbug: print('stage8', flush=True)
self.ckp.log[-1, idx_data, idx_scale] /= len(d)
best = self.ckp.log.max(0)
psnr = self.ckp.log[-1, idx_data, idx_scale].numpy()
print('')
self.ckp.write_log(
'[{} x{}]\tPSNR: {:.3f} (Best: {:.3f} @epoch {})'.format(
d.dataset.name,
scale,
self.ckp.log[-1, idx_data, idx_scale],
best[0][idx_data, idx_scale],
best[1][idx_data, idx_scale] + 1
)
)
self.ckp.write_log('Forward: {:.2f}s\n'.format(timer_test.toc()))
self.ckp.write_log('Saving...')
if self.args.save_results:
self.ckp.end_background()
if not self.args.test_only:
self.ckp.save(self, epoch, is_best=(best[1][0, 0] + 1 == epoch))
self.ckp.write_log(
'Total: {:.2f}s\n'.format(timer_test.toc()), refresh=True
)
torch.set_grad_enabled(True)
return psnr
video_list = [(train[0][k],augment)
for k in random_indices[i:(batch_size+i)]]
data = pool_threads.map(loadFrame,video_list)
next_batch = 0
for video in data:
if video.size==0: # there was an exception, skip this
next_batch = 1
if(next_batch==1):
continue
x = np.asarray(data,dtype=np.float32)
x = Variable(torch.FloatTensor(x)).cuda().contiguous()
y = train[1][random_indices[i:(batch_size+i)]]
y = torch.from_numpy(y).cuda()
output = model(x)
loss = criterion(output, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
prediction = output.data.max(1)[1]
accuracy = ( float( prediction.eq(y.data).sum() ) /float(batch_size))*100.0
if(epoch==0):
print(i,accuracy)
train_accu.append(accuracy)
accuracy_epoch = np.mean(train_accu)
real_boxes = np.array([[20, 30, 70, 80], [320, 23, 500, 120]], dtype = np.float32)
boxes = []
for box in real_boxes:
bndbox = []
for i, pt in enumerate(box):
cur_pt = pt / 500 if i % 2 == 0 else pt / 500
bndbox.append(cur_pt)
boxes.append(bndbox)
boxes = np.array(boxes, dtype = np.float32)
labels = np.array([[1], [7]], dtype = np.int64)
print("before augment", boxes)
imgs, boxes, labels = augment(imgs, boxes, labels)
print("after augment", boxes)
print(labels)
boxes, labels = torch.from_numpy(boxes), torch.from_numpy(labels)
conf_target, loc_target = encoder(boxes, labels)
print(" [*] Encoding results")
print(" [*] conf_target:")
print(conf_target)
print(" [*] loc_target:")
print(loc_target)
decoder = Decoder(input_size = 300)
print(" [*] Decoding result")
conf_pred = np.stack([one_hot(target, 9).numpy() for target in conf_target])
conf_pred = conf_pred.reshape(-1, 9)
# Sample the rest and keep the fixed one
_, con_noise = model.ginput_noise(n_con_samples)
z_noise = model.sample_fixed_noise(
'normal', 1, noise_dim=model.z_dim)
z_noise = z_noise.expand(
(n_con_samples, model.z_dim))
# Generate an image
gen_x = model.g_forward(z_noise, con_noise).detach()
filename = 'evalImages_fixusual_r{0}'.format(str(repeat))
model.plot_traj_grid(gen_x, filename, model.test_dir,
n=n_con_samples)
# Plot original
test_dataset = TrajDataset(path_to_data, device)
ref_np = torch.from_numpy(
test_dataset.get_subset(len(test_dataset), n_con_samples, reshape=False))
filename = 'evalImages_original'
model.plot_traj_grid(ref_np, filename, model.test_dir,
n=n_con_samples)
# Evaluate the model
if args.compute_prd:
eval_config = config_file['eval_config']
eval_config['filepath'] = eval_config['filepath'].format(args.config_name)
compute_chpnt_prds = eval_config['compute_chpnt_prds']
chnpt_list = eval_config['chnpt_list']
print(eval_config)
if not args.train:
def __init__(self, gender='female', num_betas=10):
super(STAR, self).__init__()
if gender not in ['male', 'female', 'neutral']:
raise RuntimeError('Invalid Gender')
if gender == 'male':
path_model = cfg.path_male_star
elif gender == 'female':
path_model = cfg.path_female_star
else:
path_model = cfg.path_neutral_star
if not os.path.exists(path_model):
raise RuntimeError('Path does not exist %s' % (path_model))
import numpy as np
star_model = np.load(path_model, allow_pickle=True)
J_regressor = star_model['J_regressor']
rows, cols = np.where(J_regressor != 0)
vals = J_regressor[rows, cols]
self.num_betas = num_betas
# Model sparse joints regressor, regresses joints location from a mesh
self.register_buffer('J_regressor',
torch.cuda.FloatTensor(J_regressor))
# Model skinning weights
self.register_buffer('weights',
torch.cuda.FloatTensor(star_model['weights']))
# Model pose corrective blend shapes
self.register_buffer(
'posedirs',
torch.cuda.FloatTensor(star_model['posedirs'].reshape((-1, 93))))
# Mean Shape
self.register_buffer('v_template',
torch.cuda.FloatTensor(star_model['v_template']))
# Shape corrective blend shapes
self.register_buffer(
'shapedirs',
torch.cuda.FloatTensor(
np.array(star_model['shapedirs'][:, :, :num_betas])))
# Mesh traingles
self.register_buffer(
'faces', torch.from_numpy(star_model['f'].astype(np.int64)))
self.f = star_model['f']
# Kinematic tree of the model
self.register_buffer(
'kintree_table',
torch.from_numpy(star_model['kintree_table'].astype(np.int64)))
id_to_col = {
self.kintree_table[1, i].item(): i
for i in range(self.kintree_table.shape[1])
}
self.register_buffer(
'parent',
torch.LongTensor([
id_to_col[self.kintree_table[0, it].item()]
for it in range(1, self.kintree_table.shape[1])
]))
self.verts = None
self.J = None
self.R = None
# top5 = AverageMeter()
#
# # load the best checkpoint
# self.load_checkpoint(best=self.best)
# self.model.eval()
# for i, (images, labels) in enumerate(self.test_loader):
# if self.use_gpu:
# images, labels = images.cuda(), labels.cuda()
# images, labels = Variable(images), Variable(labels)
#
# # forward pass
# outputs = self.model(images)
# loss = self.loss_fn(outputs, labels)
#
# # measure accuracy and record loss
# prec1, prec5 = accuracy(outputs.data, labels.data, topk=(1, 5))
# losses.update(loss.item(), images.size()[0])
# top1.update(prec1.item(), images.size()[0])
# top5.update(prec5.item(), images.size()[0])
#
# print(
# '[*] Test loss: {:.3f}, top1_acc: {:.3f}%, top5_acc: {:.3f}%'.format(
# losses.avg, top1.avg, top5.avg)
# )
import torch
import numpy
x = torch.from_numpy(numpy.array([[1, 2], [3, 4]]))
y = 1 * x + 2 * x[0, :]
print(y)
def plot(model, likelihood):
model.eval()
likelihood.eval()
with torch.no_grad():
model = model.double()
observed_pred = likelihood(model(test_x))
# Plot:
f, ax = plt.subplots(1, 1, figsize=(6, 5))
pred_labels = observed_pred.mean.view(81, 81)
ax.scatter(x1_torch[0, :], x1_torch[1, :], marker='o')
ax.scatter(x2_torch[0, :], x2_torch[1, :], marker='+')
ax.contour(test_x1, test_x2, pred_labels.numpy())
ax.contour(test_x1, test_x2, pred_labels.numpy(), [0.5], colors=['red'])
# Converting numpy data to torch:
train_x = torch.from_numpy(x)
train_x = train_x.double()
train_y = torch.from_numpy(y)
train_y = train_y.double()
test_x = torch.from_numpy(tgrid)
test_x1 = torch.from_numpy(t1)
test_x2 = torch.from_numpy(t2)
x1_torch = torch.from_numpy(x1)
x2_torch = torch.from_numpy(x2)
# Manual parameters: [No training]
model = GPClassificationModel(train_x)
init_lengthscale = 0.5
init_sigmaf_sq = 10.0
model.covar_module.base_kernel.lengthscale = init_lengthscale
model.covar_module.outputscale = init_sigmaf_sq
def binary_loss(y_pred, y):
y_pred = torch.from_numpy(y_pred)
logits = (y * y_pred.clamp(1e-12).log() + (1 - y) *
(1 - y_pred).clamp(1e-12).log()).mean()
return -logits
def load_fc(buf, start, fc_model):
num_w = fc_model.weight.numel()
num_b = fc_model.bias.numel()
fc_model.bias.data.copy_(torch.from_numpy(buf[start:start+num_b])); start = start + num_b
fc_model.weight.data.copy_(torch.from_numpy(buf[start:start+num_w])); start = start + num_w
return start
def main():
# Training settings
parser = argparse.ArgumentParser(description='MNIST')
parser.add_argument('--batch-size',
type=int,
default=16,
metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size',
type=int,
default=1000,
metavar='N',
help='input batch size for testing (default: 50)'
) # train itself 9221, test 3767
parser.add_argument('--epochs',
type=int,
default=60,
metavar='N',
help='number of epochs to train (default: 100)')
parser.add_argument('--lr',
type=float,
default=0.0001,
metavar='LR',
help='learning rate (default: 0.005)')
parser.add_argument('--momentum',
type=float,
default=0.3,
metavar='M',
help='SGD momentum (default: 0.005)')
parser.add_argument('--weight_decay',
type=float,
default=0,
metavar='M',
help='SGD momentum (default: 0.0005)')
parser.add_argument('--no-cuda',
action='store_true',
default=False,
help='disables CUDA training')
parser.add_argument('--seed',
type=int,
default=1,
metavar='S',
help='random seed (default: 1)')
parser.add_argument(
'--log-interval',
type=int,
default=2,
metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model',
action='store_true',
default=True,
help='For Saving the current Model')
args = parser.parse_args()
use_cuda = not args.no_cuda and torch.cuda.is_available()
torch.manual_seed(args.seed)
device = torch.device("cuda:0" if use_cuda else "cpu")
kwargs = {'num_workers': 2, 'pin_memory': True} if use_cuda else {}
#=================================Load Data=================================
X_train, y_train = read_dataset('MNIST_X_train.csv', 'MNIST_y_train.csv')
X_test, y_test = read_dataset('MNIST_X_test.csv', 'MNIST_y_test.csv')
X_train, X_test = normalize_features(X_train, X_test)
print('Trian:', X_train.shape)
print('Test:', X_test.shape)
print(y_train.shape)
print(y_test.shape)
#==================================Pack Data================================
train_data = torch.from_numpy(X_train).float()
test_data = torch.from_numpy(X_test).float()
trainset = torch.utils.data.TensorDataset(
train_data, torch.from_numpy(y_train.ravel()))
testset = torch.utils.data.TensorDataset(test_data,
torch.from_numpy(y_test.ravel()))
# Define data loader
train_loader = torch.utils.data.DataLoader(dataset=trainset,
batch_size=args.batch_size,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(dataset=testset,
batch_size=args.test_batch_size,
shuffle=False,
**kwargs)
#=================================Design Net================================
C_in = 1
H1 = 4
H2 = 256
K1 = (3, 3)
P1 = 2
H0 = H1 * 13 * 13
D_out = 10
model = CNNLayerNet(C_in, H0, H1, H2, K1, P1, D_out).to(device)
# optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
optimizer = optim.Adam(model.parameters(),
lr=args.lr,
eps=1e-08,
weight_decay=args.weight_decay,
amsgrad=False)
lr_adjust = optim.lr_scheduler.StepLR(optimizer,
step_size=20,
gamma=0.5,
last_epoch=-1)
for epoch in range(1, args.epochs + 1):
lr_adjust.step()
train(args, model, device, train_loader, optimizer, epoch)
test(args, model, device, epoch, test_loader)
if (args.save_model):
torch.save(model.state_dict(), "mnist_cnn.pt")
def __getitem__(self, index):
#---------
# Image
#---------
img_path = self.img_files[index % len(self.img_files)].rstrip()
img = np.array(Image.open(img_path))
# Handles images with less than three channels
while len(img.shape) != 3:
index += 1
img_path = self.img_files[index % len(self.img_files)].rstrip()
img = np.array(Image.open(img_path))
h, w, _ = img.shape
dim_diff = np.abs(h - w)
# Upper (left) and lower (right) padding
pad1, pad2 = dim_diff // 2, dim_diff - dim_diff // 2
# Determine padding
pad = ((pad1, pad2), (0, 0),
(0, 0)) if h <= w else ((0, 0), (pad1, pad2), (0, 0))
# Add padding
input_img = np.pad(img, pad, 'constant', constant_values=128) / 255.
padded_h, padded_w, _ = input_img.shape
# Resize and normalize
input_img = resize(input_img, (*self.img_shape, 3), mode='reflect')
# Channels-first
input_img = np.transpose(input_img, (2, 0, 1))
# As pytorch tensor
input_img = torch.from_numpy(input_img).float()
#---------
# Label
#---------
label_path = self.label_files[index % len(self.img_files)].rstrip()
labels = None
if os.path.exists(label_path):
labels = np.loadtxt(label_path).reshape(-1, 5)
# Extract coordinates for unpadded + unscaled image
x1 = w * (labels[:, 1] - labels[:, 3] / 2)
y1 = h * (labels[:, 2] - labels[:, 4] / 2)
x2 = w * (labels[:, 1] + labels[:, 3] / 2)
y2 = h * (labels[:, 2] + labels[:, 4] / 2)
# Adjust for added padding
x1 += pad[1][0]
y1 += pad[0][0]
x2 += pad[1][0]
y2 += pad[0][0]
# Calculate ratios from coordinates
labels[:, 1] = ((x1 + x2) / 2) / padded_w
labels[:, 2] = ((y1 + y2) / 2) / padded_h
labels[:, 3] *= w / padded_w
labels[:, 4] *= h / padded_h
# Fill matrix
filled_labels = np.zeros((self.max_objects, 5))
if labels is not None:
filled_labels[range(
len(labels))[:self.max_objects]] = labels[:self.max_objects]
filled_labels = torch.from_numpy(filled_labels)
return img_path, input_img, filled_labels
"tau": args.tau,
"n": args.n,
"length": args.length,
"device": device,
}
if model.is_attn:
context_voc = Vocabulary()
context_voc.load(model.params["context_voc"])
to_input["context_voc"] = context_voc
if model.params["use_ch"]:
ch_voc = Vocabulary()
ch_voc.load(model.params["ch_voc"])
to_input["ch_voc"] = ch_voc
for i in tqdm(range(len(dataset)), file=logfile):
if model.is_w2v:
to_input["input"] = torch.from_numpy(dataset.input_vectors[i])
if model.is_ada:
to_input["input"] = torch.from_numpy(
dataset.input_adaptive_vectors[i])
if model.is_attn:
to_input["word"] = dataset.data[i][0][0]
to_input["context"] = " ".join(dataset.data[i][2])
if model.params["use_ch"]:
to_input["CH_word"] = dataset.data[i][0][0]
if model.params["use_seed"]:
to_input["prefix"] = dataset.data[i][0][0]
else:
to_input["prefix"] = BOS
defsave.write("Word: {0}\nContext: {1}\n".format(
dataset.data[i][0][0], " ".join(dataset.data[i][2])))
defsave.write(generate(**to_input) + "\n")
def test(args, shared_model):
log = {}
logger = setup_logger("test_log", "./logs/test_log")
torch.manual_seed(args.seed)
env = Tetris(50)
model = agentNET()
model.eval()
test_time = 0
reward_num = 0
clean_sum = 0
max_reward = -1
while (1):
model.load_state_dict(shared_model.state_dict())
if args.gpu:
model = model.cuda()
cx = Variable(torch.zeros(1, 78).cuda())
hx = Variable(torch.zeros(1, 78).cuda())
else:
cx = Variable(torch.zeros(1, 78))
hx = Variable(torch.zeros(1, 78))
state = env.reset() #50 100 3
state = torch.from_numpy(state).float()
while (1):
if args.gpu:
value, logit, (hx, cx) = model(
(Variable(state.unsqueeze(0)).cuda(), (hx, cx)))
else:
value, logit, (hx, cx) = model(
(Variable(state.unsqueeze(0)), (hx, cx)))
prob = F.softmax(logit)
if args.gpu:
action = prob.max(1)[1].data.cpu()
else:
action = prob.max(1)[1].data
state, done, reward, clean = env.step(action.numpy()[0])
state = torch.from_numpy(state).float()
reward_num += reward
clean_sum += clean.get('1', -1000)
if done:
#print('dead', test_time)
test_time += 1
break
if test_time % 50 == 0:
if reward_num > max_reward:
if args.gpu:
model = model.cpu()
state_to_save = model.state_dict()
torch.save(state_to_save, "./tetris.dat")
logger.info('save')
max_reward = reward_num
if args.gpu:
model = model.cuda()
logger.info('reward = ' + str(reward_num / test_time))
logger.info('cleaned = ' + str(clean_sum / test_time))
test_time = 0
reward_num = 0
clean_sum = 0
# implement depthwise horizontal layer as a grouped convolution, preserve the stride of the original convolution
depthwise_horizontal_layer = \
torch.nn.Conv2d(in_channels=horizontal.shape[1], \
out_channels=horizontal.shape[1],
kernel_size=(1, horizontal.shape[0]), stride=layer.stride,
padding=(0, layer.padding[0]),
dilation=layer.dilation, groups=horizontal.shape[1], bias=False)
# iplement pointwise layer as conv2d with 1x1 kernel also
pointwise_r_to_t_layer = torch.nn.Conv2d(in_channels=last.shape[1], \
out_channels=last.shape[0], kernel_size=1, stride=1,
padding=0, dilation=layer.dilation, bias=False)
# convert if using GPU
if USE_GPU:
horizontal = torch.from_numpy(horizontal).type(
'torch.FloatTensor').cuda()
vertical = torch.from_numpy(vertical).type(
'torch.FloatTensor').cuda()
first = torch.from_numpy(first).type('torch.FloatTensor').cuda()
last = torch.from_numpy(last).type('torch.FloatTensor').cuda()
else:
horizontal = torch.from_numpy(horizontal).type('torch.FloatTensor')
vertical = torch.from_numpy(vertical).type('torch.FloatTensor')
first = torch.from_numpy(first).type('torch.FloatTensor')
last = torch.from_numpy(last).type('torch.FloatTensor')
# initialize weights of the layers constructed with the tensors obtained from the CP decomposition
depthwise_horizontal_layer.weight.data = \
torch.transpose(horizontal, 1, 0).unsqueeze(1).unsqueeze(1)
depthwise_vertical_layer.weight.data = \
torch.transpose(vertical, 1, 0).unsqueeze(1).unsqueeze(-1)
TRAIN_X = VECTORIZER.transform(TRAIN_X_RAW).todense()
for index, row in enumerate(TRAIN_X):
if np.sum(row) < 1e-2:
TRAIN_X[index, :] = np.ones((1, VOCABULARY_SIZE),dtype=np.float32)
TRAIN_X = TRAIN_X / TRAIN_X.sum(axis=1)
TRAIN_X = np.matmul(EMBEDDINGS, TRAIN_X.T)
DEV_X = VECTORIZER.transform(DEV_X_RAW).T
DEV_X = DEV_X / DEV_X.sum(axis=0)
DEV_X = np.matmul(EMBEDDINGS, DEV_X)
TEST_X = VECTORIZER.transform(TEST_X_RAW).T
TEST_X = TEST_X / TEST_X.sum(axis=0)
TEST_X = np.matmul(EMBEDDINGS, TEST_X)
TRAIN_DATA = data_utils.TensorDataset(torch.from_numpy(TRAIN_X.T.astype(np.float32)), torch.from_numpy(TRAIN_Y))
TRAIN_LOADER = data_utils.DataLoader(TRAIN_DATA, batch_size=173,
shuffle=True, num_workers=2, drop_last=True)
DEV_DATA = data_utils.TensorDataset(torch.from_numpy(DEV_X.T.astype(np.float32)), torch.from_numpy(DEV_Y))
DEV_LOADER = data_utils.DataLoader(DEV_DATA, batch_size=1,
shuffle=False, num_workers=2)
TEST_DATA = data_utils.TensorDataset(torch.from_numpy(TEST_X.T.astype(np.float32)), torch.from_numpy(TEST_Y))
TEST_LOADER = data_utils.DataLoader(TEST_DATA, batch_size=1,
shuffle=False, num_workers=2)
##################################################################
# Define the network
##################################################################
class SentimentNet(nn.Module):
def __init__(self, hidden_dim):
super(SentimentNet, self).__init__()
self.linear1 = nn.Linear(EMBEDDING_SIZE, hidden_dim)
'crazing', 'inclusion', 'patches',
'pitted_surface', 'rolled-in_scale', 'scratches')
fasterRCNN = vgg16(classes, pretrained=False, class_agnostic=False)
fasterRCNN.create_architecture()
fasterRCNN.cuda()
fasterRCNN.eval()
im_data = torch.FloatTensor(1).to(device)
im_info = torch.FloatTensor(1).to(device)
num_boxes = torch.LongTensor(1).to(device)
gt_boxes = torch.FloatTensor(1).to(device)
input = np.random.random((1,args.img_size, args.img_size,3))
im_blob = input
im_info_np = np.array([[im_blob.shape[1], im_blob.shape[2], 1.5]], dtype=np.float32)
im_data_pt = torch.from_numpy(im_blob)
im_data_pt = im_data_pt.permute(0, 3, 1, 2)
im_info_pt = torch.from_numpy(im_info_np)
# summary(net1, input_size=(3,300,300))
im_data.resize_(im_data_pt.size()).copy_(im_data_pt)
im_info.resize_(im_info_pt.size()).copy_(im_info_pt)
gt_boxes.resize_(1, 1, 5).zero_()
num_boxes.resize_(1).zero_()
total = sum([param.nelement() for param in fasterRCNN.parameters()])
print(' + fasterRCNN Number of params: %.2fM' % (total / 1e6))
print_model_parm_flops(fasterRCNN,im_data,im_info,gt_boxes,num_boxes)
def id_to_torch(speaker_id):
if speaker_id is not None:
speaker_id = np.asarray(speaker_id)
speaker_id = torch.from_numpy(speaker_id).unsqueeze(0)
return speaker_id
def compute_hausdorff_distance(
y_pred: Union[np.ndarray, torch.Tensor],
y: Union[np.ndarray, torch.Tensor],
include_background: bool = False,
distance_metric: str = "euclidean",
percentile: Optional[float] = None,
directed: bool = False,
):
"""
Compute the Hausdorff distance.
Args:
y_pred: input data to compute, typical segmentation model output.
It must be one-hot format and first dim is batch, example shape: [16, 3, 32, 32]. The values
should be binarized.
y: ground truth to compute mean the distance. It must be one-hot format and first dim is batch.
The values should be binarized.
include_background: whether to skip distance computation on the first channel of
the predicted output. Defaults to ``False``.
distance_metric: : [``"euclidean"``, ``"chessboard"``, ``"taxicab"``]
the metric used to compute surface distance. Defaults to ``"euclidean"``.
percentile: an optional float number between 0 and 100. If specified, the corresponding
percentile of the Hausdorff Distance rather than the maximum result will be achieved.
Defaults to ``None``.
directed: whether to calculate directed Hausdorff distance. Defaults to ``False``.
"""
if not include_background:
y_pred, y = ignore_background(y_pred=y_pred, y=y)
if isinstance(y, torch.Tensor):
y = y.float()
if isinstance(y_pred, torch.Tensor):
y_pred = y_pred.float()
if y.shape != y_pred.shape:
raise ValueError(
f"y_pred and y should have same shapes, got {y_pred.shape} and {y.shape}."
)
batch_size, n_class = y_pred.shape[:2]
hd = np.empty((batch_size, n_class))
for b, c in np.ndindex(batch_size, n_class):
(edges_pred, edges_gt) = get_mask_edges(y_pred[b, c], y[b, c])
if not np.any(edges_gt):
warnings.warn(
f"the ground truth of class {c} is all 0, this may result in nan/inf distance."
)
if not np.any(edges_pred):
warnings.warn(
f"the prediction of class {c} is all 0, this may result in nan/inf distance."
)
distance_1 = compute_percent_hausdorff_distance(
edges_pred, edges_gt, distance_metric, percentile)
if directed:
hd[b, c] = distance_1
else:
distance_2 = compute_percent_hausdorff_distance(
edges_gt, edges_pred, distance_metric, percentile)
hd[b, c] = max(distance_1, distance_2)
return torch.from_numpy(hd)
def forward(self, states):
if not isinstance(states, torch.Tensor):
states = torch.from_numpy(states).float().to(
tu.global_device())
dist = super().forward(states)
return torch.softmax(dist, dim=-1)
if SAVE_MODEL and not LOAD_MODEL:
torch.save(
{
"epoch": epoch_i,
"hyper": hyper,
"loss": (l_mem_mean, l_mem_var),
"state_dict": diff_net.state_dict()
}, f"./models/{filename}_loss_{i:01d}.torch")
print("\n################################################")
print("Evaluating Performance:\n")
t0_batch = time.perf_counter()
with torch.no_grad():
# Convert NumPy samples to torch:
x_torch = torch.from_numpy(x_test).float().to(diff_net.device)
y_hat, dydx_hat = diff_net(x_torch)
y_hat = y_hat.cpu().numpy()
dydx_hat = dydx_hat.transpose(dim0=1,
dim1=2).cpu().numpy().squeeze()
t_batch = (time.perf_counter() - t0_batch) / (float(x_test.shape[0]))
# Compute Errors:
err_y = 1. / float(x_test.shape[0]) * np.sum((y_hat - y_test)**2)
err_dydx = 1. / float(x_test.shape[0]) * np.sum(
(dydx_hat - dydx_test)**2)
print(f"Performance:")
print(f" y MSE = {err_y:.3e}")
print(f" dy/dx MSE = {err_dydx:.3e}")
iter = iter + 1
if (loss.data[0] < best_loss):
print('saving')
best_loss = loss.data[0]
torch.save(model, 'noise2.pkl')
#Evalute model performance on test set
if (epoch % 5 == 0):
t = 0
for a, b in train_test_image_folder:
x = np.asarray(
a) #There is some proble in PIL to tensor conversion
x = x[:, :, :1].reshape(
1, 1, 180, 180
) #Greyscale conversion and need approriate dim of (x,1,180,180)
#for model to work
test_img = torch.from_numpy(x / 255.0).float().cuda()
std = np.random.uniform(20, 30)
nse = torch.FloatTensor(test_img.size()).normal_(mean=0,
std=std /
255.0).cuda()
#torch.sum(test_img**2).cpu().item()
nssy_img = test_img + nse
out = model(nssy_img)
est_image = nssy_img - out
print("PSNR of test image" + str(t) + " is " +
str(psnr(est_image, test_img).cpu().item()))
t = t + 1
