I have a Deep CNN based on ResNet, and a dataset(10000, 50,50,1) to classify digits . when I run it to start leanrning , accuracy just stops in some value and gently occilating(around 0.2). I am wondering if it has overfitting or there is another issue involved ?
here is the identity block :
def identity_block(X, f, filters, stage, block):
# defining name basics
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
# retrieve filters
F1, F2, F3 = filters
# save the shortcut
X_shortcut = X
# first component
X = Conv2D(filters=F1, kernel_size=(1, 1), strides=(1, 1), padding='valid', name=conv_name_base + '2a',
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name=bn_name_base + '2a')(X)
X = Activation('relu')(X)
# second component
X = Conv2D(filters=F2, kernel_size=(f, f), strides=(1, 1), padding='same', name=conv_name_base + '2b',
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name=bn_name_base + '2b')(X)
X = Activation('relu')(X)
# third component
X = Conv2D(filters=F3, kernel_size=(1, 1), strides=(1, 1), padding='valid', name=conv_name_base + '2c',
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name=bn_name_base + '2c')(X)
# final component
X = Add()([X, X_shortcut])
X = Activation('relu')(X)
return X
and convolutional block :
def conv_block(X, f, filters, stage, block, s=2):
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
# Retivr filters
F1, F2, F3 = filters
# Save shortcut
X_shortcut = X
# First component
X = Conv2D(F1, kernel_size=(1, 1), strides=(s, s), name=conv_name_base + '2a',
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name=bn_name_base + '2a')(X)
X = Activation('relu')(X)
# Second component
X = Conv2D(F2, kernel_size=(f, f), strides=(1, 1), padding='same', name=conv_name_base + '2b',
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name=bn_name_base + '2b')(X)
X = Activation('relu')(X)
# third component
X = Conv2D(F3, kernel_size=(1, 1), strides=(1, 1), name=conv_name_base + '2c', padding='valid',
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name=bn_name_base + '2c')(X)
# short cut
X_shortcut = Conv2D(F3, kernel_size=(1, 1), strides=(s, s), name=conv_name_base + '1',
kernel_initializer=initializers.glorot_uniform(seed=0))(X_shortcut)
X_shortcut = BatchNormalization(axis=3, name=bn_name_base + '1')(X_shortcut)
# finaly
X = Add()([X, X_shortcut])
X = Activation('relu')(X)
return X
and finaly the ResNet:
def ResNet( input_shape=(50, 50, 1), classes=10):
inp = Input(shape=(50,50,1))
# zero padding
X = ZeroPadding2D((3, 3), name='pad0')(inp)
# stage1
X = Conv2D(32, (5,5), name='conv1', input_shape=input_shape,
kernel_initializer=initializers.glorot_uniform(seed=0))(X)
X = BatchNormalization(axis=3, name='bn1')(X)
X = Activation('relu')(X)
X = MaxPooling2D((2,2), name='pool1')(X)
# Stage 2
stage2_filtersize = 32
X = conv_block(X, 3, filters=[stage2_filtersize, stage2_filtersize, stage2_filtersize], stage=2, block='a', s=1)
X = identity_block(X, 3, [stage2_filtersize,stage2_filtersize, stage2_filtersize], stage=2, block='b')
X = identity_block(X, 3, [stage2_filtersize, stage2_filtersize, stage2_filtersize], stage=2, block='c')
# Stage 3
stage3_filtersize = 64
X = conv_block(X, 3, filters=[stage3_filtersize, stage3_filtersize, stage3_filtersize], stage=3, block='a', s=1)
X = identity_block(X, 3, [stage3_filtersize, stage3_filtersize, stage3_filtersize], stage=3, block='b')
X = identity_block(X, 3, [stage3_filtersize, stage3_filtersize, stage3_filtersize], stage=3, block='c')
# Stage 4
stage4_filtersize = 128
X = conv_block(X, 3, filters=[stage4_filtersize, stage4_filtersize, stage4_filtersize], stage=4, block='a', s=1)
X = identity_block(X, 3, [stage4_filtersize, stage4_filtersize, stage4_filtersize], stage=4, block='b')
X = identity_block(X, 3, [stage4_filtersize, stage4_filtersize, stage4_filtersize], stage=4, block='c')
# final
X = AveragePooling2D((2, 2), padding='same', name='Pool0')(X)
# FC
X = Flatten(name='D0')(X)
X = Dense(classes, activation='softmax', kernel_initializer=initializers.glorot_uniform(seed=0), name='D2')(X)
# creat model
model = Model(inputs=inp, outputs=X)
return model
update 1 : here are the fitting and compile methods :
model.compile(optimizer='adam',
loss=tensorflow.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
model.compile(optimizer='adam',
loss=tensorflow.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=['accuracy'])
print("model compiled settings imported successfully")
early_stopping = EarlyStopping(monitor='val_loss', patience=2)
model.fit(X_train, Y_train, validation_split=0.2, callbacks=[early_stopping], epochs=10)
test_loss, test_acc = model.evaluate(X_test, Y_test, verbose=2)
First try normalizing the values of the digit image (50x50).
Then also consider how a neural network learns its weights. Convolutional Neural Networks learns by continually adding gradient error vectors that are multiplied by a learning rate computed from backpropagation to various weight matrices throughout the network as training examples are passed through.
The most important thing to consider is the multiplication of the learning rate, because once we didn't scale the training inputs the range of distributions of the feature values will be likely different from each feature, thus the learning rate would cause corrections in each dimension that would differ from one another. This is random, so the machine could be overcompensating a correction in one weight dimension and under compensating in another. Which is very non-ideal as this might result in an oscillation state or a very slow training state.
Oscillating means that the model is unable to locate the center for the better maxima in weights.
Slow training means moving too slow to achieve a better maxima.
This is why it is a common practice to normalize images before using it as an input for Neural Network or any Models that is Gradient-Based.
TF_Support's answer:
Provide few samples of the dataset, loss curve, accuracy plot so we can clearly understand what you're trying to learn, it's more important than the code you provided.
I would guess, you are trying to learn very hard samples, 50by50 grayscale is not much. Is your network overfitting? (We would only figure that out after looking into some plots of the validation metrics) (0.2 is it your training accuracy?)
First do a sanity check on the dataset, by training a very simple CNN. I see you have 10 classes (not sure, just guessing from the function's default value), the randomized accuracy is 10%, so set a baseline first with a simple CNN and then try to improve with ResNet.
Increase the learning rate and see how the accuracy fluctuates. After a few epochs, reduce the learning rate when the accuracy better than the baseline.
Related
My vae class looks like this:
class Encoder(nn.Module):
def __init__(self):
super(Encoder, self).__init__()
c = capacity
self.conv1 = nn.Conv2d(in_channels=1, out_channels=c, kernel_size=4, stride=2, padding=1) # out: c x 14 x 14
self.conv2 = nn.Conv2d(in_channels=c, out_channels=c*2, kernel_size=4, stride=2, padding=1) # out: c x 7 x 7
self.fc_mu = nn.Linear(in_features=c*2*7*7, out_features=latent_dims)
self.fc_logvar = nn.Linear(in_features=c*2*7*7, out_features=latent_dims)
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = x.view(x.size(0), -1) # flatten batch of multi-channel feature maps to a batch of feature vectors
x_mu = self.fc_mu(x)
x_logvar = self.fc_logvar(x)
return x_mu, x_logvar
class Decoder(nn.Module):
def __init__(self):
super(Decoder, self).__init__()
c = capacity
self.fc = nn.Linear(in_features=latent_dims, out_features=c*2*7*7)
self.conv2 = nn.ConvTranspose2d(in_channels=c*2, out_channels=c, kernel_size=4, stride=2, padding=1)
self.conv1 = nn.ConvTranspose2d(in_channels=c, out_channels=1, kernel_size=4, stride=2, padding=1)
def forward(self, x):
x = self.fc(x)
x = x.view(x.size(0), capacity*2, 7, 7) # unflatten batch of feature vectors to a batch of multi-channel feature maps
x = F.relu(self.conv2(x))
x = torch.sigmoid(self.conv1(x)) # last layer before output is sigmoid, since we are using BCE as reconstruction loss
return x
class VariationalAutoencoder(nn.Module):
def __init__(self):
super(VariationalAutoencoder, self).__init__()
self.encoder = Encoder()
self.decoder = Decoder()
def forward(self, x):
latent_mu, latent_logvar = self.encoder(x)
latent = self.latent_sample(latent_mu, latent_logvar)
x_recon = self.decoder(latent)
return x_recon, latent_mu, latent_logvar
def latent_sample(self, mu, logvar):
if self.training:
# the reparameterization trick
std = logvar.mul(0.5).exp_()
eps = torch.empty_like(std).normal_()
return eps.mul(std).add_(mu)
else:
return mu
def vae_loss(recon_x, x, mu, logvar):
# recon_x is the probability of a multivariate Bernoulli distribution p.
# -log(p(x)) is then the pixel-wise binary cross-entropy.
# Averaging or not averaging the binary cross-entropy over all pixels here
# is a subtle detail with big effect on training, since it changes the weight
# we need to pick for the other loss term by several orders of magnitude.
# Not averaging is the direct implementation of the negative log likelihood,
# but averaging makes the weight of the other loss term independent of the image resolution.
recon_loss = F.binary_cross_entropy(recon_x.view(-1, 784), x.view(-1, 784), reduction='sum')
kldivergence = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
return recon_loss + variational_beta * kldivergence
I train it on MNIST dataset.
I want to sample it, or generate an array and give it to the decoder and see what the output will be.
The problem is that I don't really understand, what my z array should look like and what shape should it need.
Here is the code for sampling:
z = ...
input = torch.FloatTensor(z).to(device)
vae.eval()
output = vae.decoder(input)
plot_gallery(output.data.cpu().numpy(), 24, 24, n_row=5, n_col=5)
I have created a CNN for classification of three classes based on input images of size 39 x 39. I'm optimizing the parameters of the network using Optuna. For Optuna I'm defining the following parameters to optimize:
num_blocks = trial.suggest_int('num_blocks', 1, 4)
num_filters = [int(trial.suggest_categorical("num_filters", [32, 64, 128, 256]))]
kernel_size = trial.suggest_int('kernel_size', 2, 7)
num_dense_nodes = trial.suggest_categorical('num_dense_nodes', [64, 128, 256, 512, 1024])
dense_nodes_divisor = trial.suggest_categorical('dense_nodes_divisor', [1, 2, 4, 8])
batch_size = trial.suggest_categorical('batch_size', [16, 32, 64, 128])
drop_out = trial.suggest_discrete_uniform('drop_out', 0.05, 0.5, 0.05)
lr = trial.suggest_loguniform('lr', 1e-6, 1e-1)
dict_params = {'num_blocks': num_blocks,
'num_filters': num_filters,
'kernel_size': kernel_size,
'num_dense_nodes': num_dense_nodes,
'dense_nodes_divisor': dense_nodes_divisor,
'batch_size': batch_size,
'drop_out': drop_out,
'lr': lr}
My network looks as follows:
input_tensor = Input(shape=(39,39,3))
# 1st cnn block
x = Conv2D(filters=dict_params['num_filters'],
kernel_size=dict_params['kernel_size'],
strides=1, padding='same')(input_tensor)
x = BatchNormalization()(x, training=training)
x = Activation('relu')(x)
x = MaxPooling2D(padding='same')(x)
x = Dropout(dict_params['drop_out'])(x)
# additional cnn blocks
for i in range(1, dict_params['num_blocks']):
x = Conv2D(filters=dict_params['num_filters']*(2**i), kernel_size=dict_params['kernel_size'], strides=1, padding='same')(x)
x = BatchNormalization()(x, training=training)
x = Activation('relu')(x)
x = MaxPooling2D(padding='same')(x)
x = Dropout(dict_params['drop_out'])(x)
# mlp
x = Flatten()(x)
x = Dense(dict_params['num_dense_nodes'], activation='relu')(x)
x = Dropout(dict_params['drop_out'])(x)
x = Dense(dict_params['num_dense_nodes'] // dict_params['dense_nodes_divisor'], activation='relu')(x)
output_tensor = Dense(self.number_of_classes, activation='softmax')(x)
# instantiate and compile model
cnn_model = Model(inputs=input_tensor, outputs=output_tensor)
opt = Adam(lr=dict_params['lr'])
loss = 'categorical_crossentropy'
cnn_model.compile(loss=loss, optimizer=opt, metrics=['accuracy', tf.keras.metrics.AUC()])
I'm optimizing (minimizing) the validation loss with Optuna. There is a maximum of 4 blocks in the network and the number of filters is doubled for each block. That means e.g. 64 in the first block, 128 in the second block, 256 in the third and so on. There are two problems. First, when we start with e.g. 256 filters and a total of 4 blocks, in the last block there will be 2048 filters which is too much.
Is it possible to make the num_filters parameter dependent on the num_blocks parameter? That means if there are more blocks, the starting filter size should be smaller. So, for example, if num_blocks is chosen to be 4, num_filters should only be sampled from 32, 64 and 128.
Second, I think it is common to double the filter size but there are also networks with constant filter sizes or two convolutions (with same number of filters) before a max pooling layer (similar to VGG) and so on. Is it possible to adapt the Optuna optimization to cover all these variations?
I have a dataset containing 1000 examples where each example has 5 features (a,b,c,d,e). I want to feed 7 examples to an LSTM so it predicts the feature (a) of the 8th day.
Reading Pytorchs documentation of nn.LSTM() I came up with the following:
input_size = 5
hidden_size = 10
num_layers = 1
output_size = 1
lstm = nn.LSTM(input_size, hidden_size, num_layers)
fc = nn.Linear(hidden_size, output_size)
out, hidden = lstm(X) # Where X's shape is ([7,1,5])
output = fc(out[-1])
output # output's shape is ([7,1])
According to the docs:
The input of the nn.LSTM is "input of shape (seq_len, batch, input_size)" with "input_size – The number of expected features in the input x",
And the output is: "output of shape (seq_len, batch, num_directions * hidden_size): tensor containing the output features (h_t) from the last layer of the LSTM, for each t."
In this case, I thought seq_len would be the sequence of 7 examples, batchis 1 and input_size is 5. So the lstm would consume each example containing 5 features refeeding the hidden layer every iteration.
What am I missing?
When I extend your code to a full example -- I also added some comments to may help -- I get the following:
import torch
import torch.nn as nn
input_size = 5
hidden_size = 10
num_layers = 1
output_size = 1
lstm = nn.LSTM(input_size, hidden_size, num_layers)
fc = nn.Linear(hidden_size, output_size)
X = [
[[1,2,3,4,5]],
[[1,2,3,4,5]],
[[1,2,3,4,5]],
[[1,2,3,4,5]],
[[1,2,3,4,5]],
[[1,2,3,4,5]],
[[1,2,3,4,5]],
]
X = torch.tensor(X, dtype=torch.float32)
print(X.shape) # (seq_len, batch_size, input_size) = (7, 1, 5)
out, hidden = lstm(X) # Where X's shape is ([7,1,5])
print(out.shape) # (seq_len, batch_size, hidden_size) = (7, 1, 10)
out = out[-1] # Get output of last step
print(out.shape) # (batch, hidden_size) = (1, 10)
out = fc(out) # Push through linear layer
print(out.shape) # (batch_size, output_size) = (1, 1)
This makes sense to me, given your batch_size = 1 and output_size = 1 (I assume, you're doing regression). I don't know where your output.shape = (7, 1) come from.
Are you sure that your X has the correct dimensions? Did you create nn.LSTM maybe with batch_first=True? There are lot of little things that can sneak in.
I am new to pytorch. The following is the basic example of using nn module to train a simple one-layer model with some random data (from here)
import torch
N, D_in, H, D_out = 64, 1000, 100, 10
x = torch.randn(N, D_in)
y = torch.randn(N, D_out)
model = torch.nn.Sequential(
torch.nn.Linear(D_in, H),
torch.nn.ReLU(),
torch.nn.Linear(H, D_out),
)
loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
for t in range(500):
y_pred = model(x)
loss = loss_fn(y_pred, y)
print(t, loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
As far as I understand, the batch size is equal to 1 in the example, in other words, a single point (out of 64) is used to calculate gradients and update parameters. My question is: how to modify this example to train the model with the batch size greater than one?
In fact N is the batch size. So you just need to modify N currently its set to 64. So you have in every training batch 64 vectors with size / dim D_in.
I checked the link you posted, you can also take a look at the comments - there is some explanation too :)
# -*- coding: utf-8 -*-
import numpy as np
# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
N, D_in, H, D_out = 64, 1000, 100, 10
# Create random input and output data
x = np.random.randn(N, D_in)
y = np.random.randn(N, D_out)
# Randomly initialize weights
w1 = np.random.randn(D_in, H)
w2 = np.random.randn(H, D_out)
learning_rate = 1e-6
for t in range(500):
# Forward pass: compute predicted y
h = x.dot(w1)
h_relu = np.maximum(h, 0)
y_pred = h_relu.dot(w2)
# Compute and print loss
loss = np.square(y_pred - y).sum()
print(t, loss)
# Backprop to compute gradients of w1 and w2 with respect to loss
grad_y_pred = 2.0 * (y_pred - y)
grad_w2 = h_relu.T.dot(grad_y_pred)
grad_h_relu = grad_y_pred.dot(w2.T)
grad_h = grad_h_relu.copy()
grad_h[h < 0] = 0
grad_w1 = x.T.dot(grad_h)
# Update weights
w1 -= learning_rate * grad_w1
w2 -= learning_rate * grad_w2
To include batch size in PyTorch basic examples, the easiest and cleanest way is to use PyTorch torch.utils.data.DataLoader and torch.utils.data.TensorDataset.
Dataset stores the samples and their corresponding labels, and DataLoader wraps an iterable around the Dataset to enable easy access to the samples.
DataLoader will take care of creating batches for you.
Building on your question, there is a complete code snippet, where we iterate over a dataset of 10000 examples for 2 epochs with a batch size of 64:
import torch
from torch.utils.data import DataLoader, TensorDataset
# Create the dataset with N_SAMPLES samples
N_SAMPLES, D_in, H, D_out = 10000, 1000, 100, 10
x = torch.randn(N_SAMPLES, D_in)
y = torch.randn(N_SAMPLES, D_out)
# Define the batch size and the number of epochs
BATCH_SIZE = 64
N_EPOCHS = 2
# Use torch.utils.data to create a DataLoader
# that will take care of creating batches
dataset = TensorDataset(x, y)
dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=True)
# Define model, loss and optimizer
model = torch.nn.Sequential(
torch.nn.Linear(D_in, H),
torch.nn.ReLU(),
torch.nn.Linear(H, D_out),
)
loss_fn = torch.nn.MSELoss(reduction='sum')
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
# Get the dataset size for printing (it is equal to N_SAMPLES)
dataset_size = len(dataloader.dataset)
# Loop over epochs
for epoch in range(N_EPOCHS):
print(f"Epoch {epoch + 1}\n-------------------------------")
# Loop over batches in an epoch using DataLoader
for id_batch, (x_batch, y_batch) in enumerate(dataloader):
y_batch_pred = model(x_batch)
loss = loss_fn(y_batch_pred, y_batch)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Every 100 batches, print the loss for this batch
# as well as the number of examples processed so far
if id_batch % 100 == 0:
loss, current = loss.item(), (id_batch + 1)* len(x_batch)
print(f"loss: {loss:>7f} [{current:>5d}/{dataset_size:>5d}]")
The output should be something like:
Epoch 1
-------------------------------
loss: 643.433716 [ 64/10000]
loss: 648.195435 [ 6464/10000]
Epoch 2
-------------------------------
loss: 613.619873 [ 64/10000]
loss: 625.018555 [ 6464/10000]
I have trained the two versions of Squeezenet, both with success, thanks #forresti !
When training the one with residual connections, I am stucked. Whatever learning policy I took, the one shipped in this repo, or the plainly step, I cannot train it to the results given in the paper. The accuracy is a bit lower than Squeezenet v1.0....
I know that I should post this in that repo, but I can't find issues tab there....
Anyone could shed me some light? Thanks in advance!
====================EDIT=============================
I firstly adopted the solver hyperparameters shipped with SqueezeNet-v1.0. Then, I changed the learning policy from poly to step, keeping the remaining parameters untouched and closely monitored the loss and accuracy, when they became apparently flat, I changed the learning rate by a factor of 0.4. In both cases, I got top-5 accuracies 81.9x% and 79.8x%, lower than the benchmark provided in the paper, seems rather weird....
You can use newest SqueezeNet v1.1 version of Squezenet from: https://github.com/rcmalli/keras-squeezenet
Model Definition:
from keras import backend as K
from keras.layers import Input, Convolution2D, MaxPooling2D, Activation, concatenate, Dropout
from keras.layers import GlobalAveragePooling2D, GlobalMaxPooling2D
from keras.models import Model
from keras.utils.layer_utils import get_source_inputs #https://stackoverflow.com/questions/68862735/keras-vggface-no-module-named-keras-engine-topology
from tensorflow.keras.utils import get_file
from keras.utils import layer_utils
sq1x1 = "squeeze1x1"
exp1x1 = "expand1x1"
exp3x3 = "expand3x3"
relu = "relu_"
WEIGHTS_PATH = "https://github.com/rcmalli/keras-squeezenet/releases/download/v1.0/squeezenet_weights_tf_dim_ordering_tf_kernels.h5"
WEIGHTS_PATH_NO_TOP = "https://github.com/rcmalli/keras-squeezenet/releases/download/v1.0/squeezenet_weights_tf_dim_ordering_tf_kernels_notop.h5"
# Modular function for Fire Node
def fire_module(x, fire_id, squeeze=16, expand=64):
s_id = 'fire' + str(fire_id) + '/'
if K.image_data_format() == 'channels_first':
channel_axis = 1
else:
channel_axis = 3
x = Convolution2D(squeeze, (1, 1), padding='valid', name=s_id + sq1x1)(x)
x = Activation('relu', name=s_id + relu + sq1x1)(x)
left = Convolution2D(expand, (1, 1), padding='valid', name=s_id + exp1x1)(x)
left = Activation('relu', name=s_id + relu + exp1x1)(left)
right = Convolution2D(expand, (3, 3), padding='same', name=s_id + exp3x3)(x)
right = Activation('relu', name=s_id + relu + exp3x3)(right)
x = concatenate([left, right], axis=channel_axis, name=s_id + 'concat')
return x
# Original SqueezeNet from paper.
def SqueezeNet(include_top=True, weights='imagenet',
input_tensor=None, input_shape=None,
pooling=None,
classes=1000):
"""Instantiates the SqueezeNet architecture."""
if weights not in {'imagenet', None}:
raise ValueError('The `weights` argument should be either '
'`None` (random initialization) or `imagenet` '
'(pre-training on ImageNet).')
input_shape = input_shape
if input_tensor is None:
img_input = Input(shape=input_shape)
else:
if not K.is_keras_tensor(input_tensor):
img_input = Input(tensor=input_tensor, shape=input_shape)
else:
img_input = input_tensor
x = Convolution2D(64, (3, 3), strides=(2, 2), padding='valid', name='conv1')(img_input)
x = Activation('relu', name='relu_conv1')(x)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool1')(x)
x = fire_module(x, fire_id=2, squeeze=16, expand=64)
x = fire_module(x, fire_id=3, squeeze=16, expand=64)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool3')(x)
x = fire_module(x, fire_id=4, squeeze=32, expand=128)
x = fire_module(x, fire_id=5, squeeze=32, expand=128)
x = MaxPooling2D(pool_size=(3, 3), strides=(2, 2), name='pool5')(x)
x = fire_module(x, fire_id=6, squeeze=48, expand=192)
x = fire_module(x, fire_id=7, squeeze=48, expand=192)
x = fire_module(x, fire_id=8, squeeze=64, expand=256)
x = fire_module(x, fire_id=9, squeeze=64, expand=256)
if include_top:
# It's not obvious where to cut the network...
# Could do the 8th or 9th layer... some work recommends cutting earlier layers.
x = Dropout(0.5, name='drop9')(x)
x = Convolution2D(classes, (1, 1), padding='valid', name='conv10')(x)
x = Activation('relu', name='relu_conv10')(x)
x = GlobalAveragePooling2D()(x)
x = Activation('softmax', name='loss')(x)
else:
if pooling == 'avg':
x = GlobalAveragePooling2D()(x)
elif pooling=='max':
x = GlobalMaxPooling2D()(x)
elif pooling==None:
pass
else:
raise ValueError("Unknown argument for 'pooling'=" + pooling)
#x = Dense(10, activation= 'softmax')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
if input_tensor is not None:
inputs = get_source_inputs(input_tensor)
else:
inputs = img_input
model = Model(inputs, x, name='squeezenet')
# load weights
if weights == 'imagenet':
if include_top:
weights_path = get_file('squeezenet_weights_tf_dim_ordering_tf_kernels.h5',
WEIGHTS_PATH,
cache_subdir='models')
else:
weights_path = get_file('squeezenet_weights_tf_dim_ordering_tf_kernels_notop.h5',
WEIGHTS_PATH_NO_TOP,
cache_subdir='models')
model.load_weights(weights_path)
if K.backend() == 'theano':
layer_utils.convert_all_kernels_in_model(model)
return model
Example Usage:
import numpy as np
from keras_squeezenet import SqueezeNet
from keras.applications.imagenet_utils import preprocess_input, decode_predictions
from keras.preprocessing import image
model = SqueezeNet()
img = image.load_img('../images/cat.jpeg', target_size=(227, 227))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = model.predict(x)
print('Predicted:', decode_predictions(preds))