def __init__(self, ntokens, rescale_loss, bptt, emsize,
nhid, nlayers, dropout, num_proj, batch_size, k):
out = rnn(bptt, ntokens, emsize, nhid, nlayers,
dropout, num_proj, batch_size)
rnn_out, self.last_states, self.lstm_args, self.state_names = out
# decoder weight and bias
decoder_w = S.var("decoder_weight", stype='row_sparse')
decoder_b = S.var("decoder_bias", shape=(ntokens, 1), stype='row_sparse')
# sampled softmax for training
sample = S.var('sample', shape=(k,))
prob_sample = S.var("prob_sample", shape=(k,))
prob_target = S.var("prob_target")
self.sample_names = ['sample', 'prob_sample', 'prob_target']
logits, new_targets = sampled_softmax(ntokens, k, num_proj,
rnn_out, decoder_w, decoder_b,
[sample, prob_sample, prob_target])
self.train_loss = cross_entropy_loss(logits, new_targets, rescale_loss=rescale_loss)
# full softmax for testing
eval_logits = S.FullyConnected(data=rnn_out, weight=decoder_w,
num_hidden=ntokens, name='decode_fc', bias=decoder_b)
label = S.Variable('label')
label = S.reshape(label, shape=(-1,))
self.eval_loss = cross_entropy_loss(eval_logits, label)
def rnn(bptt, vocab_size, num_embed, nhid, num_layers, dropout, num_proj, batch_size):
""" word embedding + LSTM Projected """
state_names = []
data = S.var('data')
weight = S.var("encoder_weight", stype='row_sparse')
embed = S.sparse.Embedding(data=data, weight=weight, input_dim=vocab_size,
output_dim=num_embed, name='embed', sparse_grad=True)
states = []
outputs = S.Dropout(embed, p=dropout)
for i in range(num_layers):
prefix = 'lstmp%d_' % i
init_h = S.var(prefix + 'init_h', shape=(batch_size, num_proj), init=mx.init.Zero())
init_c = S.var(prefix + 'init_c', shape=(batch_size, nhid), init=mx.init.Zero())
state_names += [prefix + 'init_h', prefix + 'init_c']
lstmp = mx.gluon.contrib.rnn.LSTMPCell(nhid, num_proj)
outputs, next_states = lstmp.unroll(bptt, outputs, begin_state=[init_h, init_c], \
layout='NTC', merge_outputs=True)
outputs = S.Dropout(outputs, p=dropout)
states += [S.stop_gradient(s) for s in next_states]
outputs = S.reshape(outputs, shape=(-1, num_proj))
trainable_lstm_args = []
for arg in outputs.list_arguments():
if 'lstmp' in arg and 'init' not in arg:
trainable_lstm_args.append(arg)
return outputs, states, trainable_lstm_args, state_names
def sampled_softmax(num_classes, num_samples, in_dim, inputs, weight, bias,
sampled_values, remove_accidental_hits=True):
""" Sampled softmax via importance sampling.
This under-estimates the full softmax and is only used for training.
"""
# inputs = (n, in_dim)
sample, prob_sample, prob_target = sampled_values
# (num_samples, )
sample = S.var('sample', shape=(num_samples,), dtype='float32')
# (n, )
label = S.var('label')
label = S.reshape(label, shape=(-1,), name="label_reshape")
# (num_samples+n, )
sample_label = S.concat(sample, label, dim=0)
# lookup weights and biases
# (num_samples+n, dim)
sample_target_w = S.sparse.Embedding(data=sample_label, weight=weight,
input_dim=num_classes, output_dim=in_dim,
sparse_grad=True)
# (num_samples+n, 1)
sample_target_b = S.sparse.Embedding(data=sample_label, weight=bias,
input_dim=num_classes, output_dim=1,
sparse_grad=True)
# (num_samples, dim)
sample_w = S.slice(sample_target_w, begin=(0, 0), end=(num_samples, None))
target_w = S.slice(sample_target_w, begin=(num_samples, 0), end=(None, None))
sample_b = S.slice(sample_target_b, begin=(0, 0), end=(num_samples, None))
target_b = S.slice(sample_target_b, begin=(num_samples, 0), end=(None, None))
# target
# (n, 1)
true_pred = S.sum(target_w * inputs, axis=1, keepdims=True) + target_b
# samples
# (n, num_samples)
sample_b = S.reshape(sample_b, (-1,))
sample_pred = S.FullyConnected(inputs, weight=sample_w, bias=sample_b,
num_hidden=num_samples)
# remove accidental hits
if remove_accidental_hits:
label_v = S.reshape(label, (-1, 1))
sample_v = S.reshape(sample, (1, -1))
neg = S.broadcast_equal(label_v, sample_v) * -1e37
sample_pred = sample_pred + neg
prob_sample = S.reshape(prob_sample, shape=(1, num_samples))
p_target = true_pred - S.log(prob_target)
p_sample = S.broadcast_sub(sample_pred, S.log(prob_sample))
# return logits and new_labels
# (n, 1+num_samples)
logits = S.concat(p_target, p_sample, dim=1)
new_targets = S.zeros_like(label)
return logits, new_targets
def cross_entropy_loss(inputs, labels, rescale_loss=1):
""" cross entropy loss with a mask """
criterion = mx.gluon.loss.SoftmaxCrossEntropyLoss(weight=rescale_loss)
loss = criterion(inputs, labels)
mask = S.var('mask')
loss = loss * S.reshape(mask, shape=(-1,))
return S.make_loss(loss.mean())
def get_symbol():
data = sym.var('data')
out1 = get_conv(data, 'conv1-1', 64, (2, 2))
out = get_conv(out1, 'conv1-2', 64)
out = get_conv(out, 'conv1-3', 64) + out1
out1 = get_conv(out, 'conv2-1', 128, (2, 2))
out = get_conv(out1, 'conv2-2', 128)
out1 = get_conv(out, 'conv2-3', 128) + out1
out = get_conv(out1, 'conv2-4', 128)
out1 = get_conv(out, 'conv2-5', 128) + out1
out1 = get_conv(out1, 'conv3-1', 256, (2, 2))
out = get_conv(out1, 'conv3-2', 256)
out1 = get_conv(out, 'conv3-3', 256) + out1
out = get_conv(out1, 'conv3-4', 256)
out1 = get_conv(out, 'conv3-5', 256) + out1
out = get_conv(out1, 'conv3-6', 256)
out1 = get_conv(out, 'conv3-7', 256) + out1
out = get_conv(out1, 'conv3-8', 256)
out1 = get_conv(out, 'conv3-9', 256) + out1
out1 = get_conv(out1, 'conv4-1', 512, (2, 2))
out = get_conv(out1, 'conv4-2', 512)
out = get_conv(out, 'conv4-3', 512) + out1
return sym.FullyConnected(out, name='fc5', num_hidden=512)
def get_conv(inpt, name, num_filter, stride=(1, 1), act_type='prelu'):
act_name = name.replace('conv', act_type)
return sym.LeakyReLU(sym.Convolution(inpt,
name=name,
num_filter=num_filter,
stride=stride,
kernel=(3, 3),
pad=(1, 1)),
name=act_name,
act_type=act_type,
gamma=sym.var(act_name))
def debug_net(net):
data = symbol.var('flow')
internals = net(data).get_internals()
hooks = [internals[i] for i in params.debug_nodes]
new = gluon.SymbolBlock(hooks, data, params=net.collect_params())
return new
"alexnet0_conv2_fwd_output", "alexnet0_conv3_fwd_output",
"alexnet0_conv4_fwd_output", "alexnet0_dense0_fwd_output",
"alexnet0_dense1_fwd_output", "alexnet0_dense2_fwd_output",
"alexnet0_dropout0_fwd_output", "alexnet0_dropout1_fwd_output",
"alexnet0_flatten0_flatten0_output", "alexnet0_pool0_fwd_output",
"alexnet0_pool1_fwd_output", "alexnet0_pool2_fwd_output"
])
# 预处理模型以及图片
net = get_model(params.model, pretrained=True)
img = image.imread(params.input_pic)
img = transform_eval(img, resize_short=224, crop_size=224)
wxf.export(img.asnumpy(), 'input.wxf', target_format='wxf')
# 列出可选的输出节点
nodes = net(symbol.var('flow')).get_internals().list_outputs()
wxf.export(nodes, 'nodes.wxf', target_format='wxf')
def debug_net(net):
data = symbol.var('flow')
internals = net(data).get_internals()
hooks = [internals[i] for i in params.debug_nodes]
new = gluon.SymbolBlock(hooks, data, params=net.collect_params())
return new
debug = debug_net(net)
ndarray = [i.asnumpy() for i in debug(img)]
wxf.export(ndarray, 'debug.wxf', target_format='wxf')
kernel=(1, 1), stride=(1, 1), no_bias=True)
conv2_relu = mx.symbol.Activation(name='res2a_branch2a_relu' + suffix, data=conv2, act_type='relu')
emb = mx.symbol.flatten(conv2_relu)
return emb
im_1 = cv2.imread('/home/alex/image1.png')
im_tr_1 = transform(im_1)
im_tr_1 = mx.nd.array(im_tr_1, mx.сpu(0))
im_2 = cv2.imread('/home/alex/image1.png')
im_tr_2 = transform(im_2)
im_tr_2 = mx.nd.array(im_tr_2, mx.сpu(0))
d1 = mxs.var('data_a')
emb1 = embedder(d1, '_a')
infer_shape(emb1, data_a=(1, 3, 32, 32), data=None)
emb1_arguments = emb1.list_arguments()
emb1_arguments.pop(0)
emb1_auxiliary = emb1.list_auxiliary_states()
mod = mx.module.Module(emb1, ['data_a'], context=[mx.сpu(0)])
mod.bind([('data_a', (1, 3, 32, 32))])
mod.init_params()
arg_params, aux_params = mod.get_params()
d2 = mxs.var('data_b')
emb2 = embedder(d2, '_b')
infer_shape(emb2, data_b=(1, 3, 32, 32), data=None)
emb2_arguments = emb2.list_arguments()
def _infer_weight_shape(op_name, data_shape, kwargs):
op = getattr(symbol, op_name)
sym = op(symbol.var('data', shape=data_shape), **kwargs)
return sym.infer_shape_partial()[0]