I'm very new to tensorflow, and I couldn't figure this one out.
I have this while loop:
def process_tree_tf(n_child, reprs, weights, bias, embed_dim, activation = tf.nn.relu):
n_child, reprs = n_child, reprs
parent_idxs = generate_parents_numpy(n_child)
loop_idx = reprs.shape[0] - 1
loop_vars = loop_idx, reprs, parent_idxs, weights, embed_dim
def loop_condition(loop_ind, *_):
return tf.greater(0, loop_idx)
def loop_body(loop_ind, reprs, parent_idxs, weights, embed_dim):
x = reprs[loop_ind]
x_expanded = tf.expand_dims(x, axis=-1)
w = weights
out = tf.squeeze(tf.add(tf.matmul(x_expanded,w,transpose_a=True), bias))
activated = activation(out)
par_idx = parent_idxs[loop_ind]
reprs = update_parent(reprs, par_idx, embed_dim, activated)
reprs = tf.Print(reprs, [reprs]) #This doesn't work
loop_ind = loop_ind-1
return loop_ind, reprs, parent_idxs, weights, embed_dim
return tf.while_loop(loop_condition, loop_body, loop_vars)
And I'm evaluating it this way:
embed_dim = 2
hidden_dim = 2
n_nodes = 4
batch = 2
reprs = np.ones((n_nodes, embed_dim+hidden_dim))
n_child = np.array([1, 1, 1, 0])
weights = np.ones((embed_dim+hidden_dim, hidden_dim))
bias = np.ones(hidden_dim)
with tf.Session() as sess:
_, r, *_ = process_tree_tf(n_child, reprs, weights, bias, embed_dim, activation=tf.nn.relu)
print(r.eval())
I want to check the value of reprs inside the while loop, but tf.Print doesn't seem to work, and print just tells me it's a tensor and gives me its shape.
How do I go about doing this?
Thank you so much!
Take a look at this webpage: https://www.tensorflow.org/api_docs/python/tf/Print
You can see that tf.Print is an identity operator with the side effect of printing data when evaluating. You should therefore use this line to print:
reprs = tf.Print(reprs, [reprs])
Hope this helps, and good luck!
The approach suggested by rmeertens is the one I think is correct. I would just add (as a response to your comments) that if something is printing "Tensor("while/update_parent:0, ...... " then that implies that that value in the graph is not being evaluated.
You are likely seeing that as the output of your "print(r.eval())" statement, NOT the tf.Print() statement.
Note that the output of tf.Print() appears in PyCharm (the IDE I am using) in red, while the output of a normal python print operation appears in black. So the tf.Print() output looks like a warning message. It could be that it is indeed printing out, but you are simply overlooking it.
Related
Tuning is done as follows:
tuner = kt.RandomSearch(
MyHyperModel(),
objective="mae",
max_trials=30,
overwrite=True,
directory=results_dir,
project_name="tune_hypermodel",
)
And I'm iterating over how many features to use:
data[name]=pd.read_pickle(os.path.join(root, name)+'/'+name+'.data.reindexed_by_pc.pkl')
CpG_num_lst=[100,500,1000,5000,10000,20000,40000]
train_score, valid_score = [], []
hps_by_CpG_num = []
for CpG_num in CpG_num_lst:
print("CpG_num:",CpG_num)
# force overwrite tune search (to start new search). Cause even with overwrite=True it doesn't overwrite
if os.path.exists(results_dir+'/tune_hypermodel'):
shutil.rmtree(results_dir+'/tune_hypermodel')
# initialize
X1, y1, X2, y2 = [dict() for _ in range(4)]
X1[name] = data[name][fold1_ids].head(CpG_num).values.T
X2[name] = data[name][fold2_ids].head(CpG_num).values.T
# get the ages of the corresponding persons. Notice info_1 and info_2 only contain "Ctrl" and not "Test" samples
y1[name] = info_1[name].query("`Train.Test`=='Train'")['Age'].values.astype(float)
y2[name] = info_2[name].query("`Train.Test`=='Train'")['Age'].values.astype(float)
# Split the data
X1_train, X1_valid, y1_train, y1_valid = train_test_split(X1[name], y1[name], test_size=0.2, shuffle= True)
# Grid search
tuner.search(X1_train, y1_train, validation_data = (X1_valid,y1_valid))
best_hps = tuner.get_best_hyperparameters(num_trials=1)[0]
# Get best hyperparameters
hp_dict = dict()
hp_dict['num_layers'] = best_hps.get('num_layers')
hp_dict['batch_size'] = best_hps.get('batch_size')
hp_dict['act_1'] = best_hps.get('act_1')
hp_dict['act_2'] = best_hps.get('act_2')
hp_dict['units_1'] = best_hps.get('units_1')
hp_dict['units_2'] = best_hps.get('units_2')
hps_by_CpG_num.append(hp_dict)
# Build best model
best_model = MyHyperModel().build(hp=best_hps)
history = best_model.fit(X1_train, y1_train, validation_data = (X1_valid,y1_valid), batch_size=best_hps.get('batch_size'), epochs=200)
However for each element in the for loop the tuner search doesn't restart, it just uses the best hyperparameters from the first search (CpG_num = 500)
What am I missing? Why is Keras taking old hyperparameters?
The solution was to include the tuner instantiation within the for loop.
tuner = kt.RandomSearch(
MyHyperModel(),
objective="mae",
max_trials=30,
overwrite=True,
directory=results_dir,
project_name="tune_hypermodel",
)
Not sure why though, but works... If somebody has any insight on this let me know. I thought the tuner would be overwritten.
Is there a better way of doing this?
I am trying to train SRGAN from scratch. I have read solutions for this type of problem, but it would be great if someone could help me debug my code. The exact error is: "RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad()" Here is the snippet I am trying to train:
gen_model = Generator().to(device, non_blocking=True)
disc_model = Discriminator().to(device, non_blocking=True)
opt_gen = optim.Adam(gen_model.parameters(), lr=0.01)
opt_disc = optim.Adam(disc_model.parameters(), lr=0.01)
from torch.nn.modules.loss import BCELoss
def train_model(gen, disc):
for epoch in range(20):
run_loss_disc = 0
run_loss_gen = 0
for data in train:
low_res, high_res = data[0].to(device, non_blocking=True, dtype=torch.float).permute(0, 3, 1, 2),data[1].to(device, non_blocking=True, dtype=torch.float).permute(0, 3, 1, 2)
#--------Discriminator-----------------
gen_image = gen(low_res)
gen_image = gen_image.detach()
disc_gen = disc(gen_image)
disc_real = disc(high_res)
p=nn.BCEWithLogitsLoss()
loss_gen = p(disc_real, torch.ones_like(disc_real))
loss_real = p(disc_gen, torch.zeros_like(disc_gen))
loss_disc = loss_gen + loss_real
opt_disc.zero_grad()
loss_disc.backward()
run_loss_disc+=loss_disc
#---------Generator--------------------
cont_loss = vgg_loss(high_res, gen_image)
adv_loss = 1e-3*p(disc_gen, torch.ones_like(disc_gen))
gen_loss = cont_loss+(10^-3)*adv_loss
opt_gen.zero_grad()
gen_loss.backward()
opt_disc.step()
opt_gen.step()
run_loss_gen+=gen_loss
print("Run Loss Discriminator: %d", run_loss_disc)
print("Run Loss Generator: %d", run_loss_gen)
train_model(gen_model, disc_model)
Apparently your disc_gen value was discarded by the first backward() call, as it says.
It should work if you change the discriminator part a bit:
gen_image = gen(low_res)
disc_gen = disc(gen_image.detach())
and add this at the start of the generator part:
disc_gen = disc(gen_image)
I have this similarity matrix plot of some documents. I want to sort the values of the matrix, which is a numpynd array, to group colors, while maintaining their relative position (diagonal yellow line), and labels as well.
path = "C:\\Users\\user\\Desktop\\texts\\dataset"
text_files = os.listdir(path)
#print (text_files)
tfidf_vectorizer = TfidfVectorizer()
documents = [open(f, encoding="utf-8").read() for f in text_files if f.endswith('.txt')]
sparse_matrix = tfidf_vectorizer.fit_transform(documents)
labels = []
for f in text_files:
if f.endswith('.txt'):
labels.append(f)
pairwise_similarity = sparse_matrix * sparse_matrix.T
pairwise_similarity_array = pairwise_similarity.toarray()
fig, ax = plt.subplots(figsize=(20,20))
cax = ax.matshow(pairwise_similarity_array, interpolation='spline16')
ax.grid(True)
plt.title('News articles similarity matrix')
plt.xticks(range(23), labels, rotation=90);
plt.yticks(range(23), labels);
fig.colorbar(cax, ticks=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1])
plt.show()
Here is one possibility.
The idea is to use the information in the similarity matrix and put elements next to each other if they are similar. If two items are similar they should also be similar with respect to other elements ie have similar colors.
I start with the element which has the most in common with all other elements (this choice is a bit arbitrary) [a] and as next element I choose from the remaining elements the one which is closest to the current [b].
import numpy as np
import matplotlib.pyplot as plt
def create_dummy_sim_mat(n):
sm = np.random.random((n, n))
sm = (sm + sm.T) / 2
sm[range(n), range(n)] = 1
return sm
def argsort_sim_mat(sm):
idx = [np.argmax(np.sum(sm, axis=1))] # a
for i in range(1, len(sm)):
sm_i = sm[idx[-1]].copy()
sm_i[idx] = -1
idx.append(np.argmax(sm_i)) # b
return np.array(idx)
n = 10
sim_mat = create_dummy_sim_mat(n=n)
idx = argsort_sim_mat(sim_mat)
sim_mat2 = sim_mat[idx, :][:, idx] # apply reordering for rows and columns
# Plot results
fig, ax = plt.subplots(1, 2)
ax[0].imshow(sim_mat)
ax[1].imshow(sim_mat2)
def ticks(_ax, ti, la):
_ax.set_xticks(ti)
_ax.set_yticks(ti)
_ax.set_xticklabels(la)
_ax.set_yticklabels(la)
ticks(_ax=ax[0], ti=range(n), la=range(n))
ticks(_ax=ax[1], ti=range(n), la=idx)
After meTchaikovsky's answer I also tested my idea on a clustered similarity matrix (see first image) this method works but is not perfect (see second image).
Because I use the similarity between two elements as approximation to their similarity to all other elements, it is quite clear why this does not work perfectly.
So instead of using the initial similarity to sort the elements one could calculate a second order similarity matrix which measures how similar the similarities are (sorry).
This measure describes better what you are interested in. If two rows / columns have similar colors they should be close to each other. The algorithm to sort the matrix is the same as before
def add_cluster(sm, c=3):
idx_cluster = np.array_split(np.random.permutation(np.arange(len(sm))), c)
for ic in idx_cluster:
cluster_noise = np.random.uniform(0.9, 1.0, (len(ic),)*2)
sm[ic[np.newaxis, :], ic[:, np.newaxis]] = cluster_noise
def get_sim_mat2(sm):
return 1 / (np.linalg.norm(sm[:, np.newaxis] - sm[np.newaxis], axis=-1) + 1/n)
sim_mat = create_dummy_sim_mat(n=100)
add_cluster(sim_mat, c=4)
sim_mat2 = get_sim_mat2(sim_mat)
idx = argsort_sim_mat(sim_mat)
idx2 = argsort_sim_mat(sim_mat2)
sim_mat_sorted = sim_mat[idx, :][:, idx]
sim_mat_sorted2 = sim_mat[idx2, :][:, idx2]
# Plot results
fig, ax = plt.subplots(1, 3)
ax[0].imshow(sim_mat)
ax[1].imshow(sim_mat_sorted)
ax[2].imshow(sim_mat_sorted2)
The results with this second method are quite good (see third image)
but I guess there exist cases where this approach also fails, so I would be happy about feedback.
Edit
I tried to explain it and did also link the ideas to the code with [a] and [b], but obviously I did not do a good job, so here is a second more verbose explanation.
You have n elements and a n x n similarity matrix sm where each cell (i, j) describes how similar element i is to element j. The goal is to order the rows / columns in such a way that one can see existing patterns in the similarity matrix. My idea to achieve this is really simple.
You start with an empty list and add elements one by one. The criterion for the next element is the similarity to the current element. If element i was added in the last step, I chose the element argmax(sm[i, :]) as next, ignoring the elements already added to the list. I ignore the elements by setting the values of those elements to -1.
You can use the function ticks to reorder the labels:
labels = np.array(labels) # make labels an numpy array, to index it with a list
ticks(_ax=ax[0], ti=range(n), la=labels[idx])
#scleronomic's solution is very elegant, but it also has one shortage, which is we cannot set the number of clusters in the sorted correlation matrix. Assume we are working with a set of variables, in which some of them are weakly correlated
import string
import numpy as np
import pandas as pd
n_variables = 20
n_clusters = 10
n_samples = 100
np.random.seed(100)
names = list(string.ascii_lowercase)[:n_variables]
belongs_to_cluster = np.random.randint(0,n_clusters,n_variables)
latent = np.random.randn(n_clusters,n_samples)
variables = np.random.rand(n_variables,n_samples)
for ind in range(n_clusters):
mask = belongs_to_cluster == ind
# weakening the correlation
if ind % 2 == 0:variables[mask] += latent[ind]*0.1
variables[mask] += latent[ind]
df = pd.DataFrame({key:val for key,val in zip(names,variables)})
corr_mat = np.array(df.corr())
As you can see, there are 10 clusters of variables by construction, however, variables within clusters that has an even index are weakly correlated. If we only want to see roughly 5 clusters in the sorted correlation matrix, maybe we need to find another way.
Based on this post, which is the accepted answer to the question "Clustering a correlation matrix", to sort a correlation matrix into blocks, what we need to find are blocks, where correlations within blocks are high and correlations between blocks are low. However, the solution provided by this accepted answer works best when we know how many blocks are there in the first place, and more importantly, the sizes of the underlying blocks are the same, or at least similar. Therefore, I improved the solution with a new function sort_corr_mat
def sort_corr_mat(corr_mat,clusters_guess):
def _swap_rows(corr_mat, var1, var2):
rs = corr_mat.copy()
rs[var2, :],rs[var1, :]= corr_mat[var1, :],corr_mat[var2, :]
cs = rs.copy()
cs[:, var2],cs[:, var1] = rs[:, var1],rs[:, var2]
return cs
# analysis
max_iter = 500
best_score,current_score,best_count = -1e8,-1e8,0
num_minimua_to_visit = 20
best_corr = corr_mat
best_ordering = np.arange(n_variables)
for i in range(max_iter):
for row1 in range(n_variables):
for row2 in range(n_variables):
if row1 == row2: continue
option_ordering = best_ordering.copy()
option_ordering[row1],option_ordering[row2] = best_ordering[row2],best_ordering[row1]
option_corr = _swap_rows(best_corr,row1,row2)
option_score = score(option_corr,n_variables,clusters_guess)
if option_score > best_score:
best_corr = option_corr
best_ordering = option_ordering
best_score = option_score
if best_score > current_score:
best_count += 1
current_corr = best_corr
current_ordering = best_ordering
current_score = best_score
if best_count >= num_minimua_to_visit:
return best_corr#,best_ordering
return best_corr#,best_ordering
With this function and the corr_mat constructed in the first place, I compared the result obtained with my function (on the right) with that obtained with #scleronomic's solution (in the middle)
sim_mat_sorted = corr_mat[argsort_sim_mat(corr_mat), :][:, argsort_sim_mat(corr_mat)]
corr_mat_sorted = sort_corr_mat(corr_mat,clusters_guess=5)
# Plot results
fig, ax = plt.subplots(1,3,figsize=(18,6))
ax[0].imshow(corr_mat)
ax[1].imshow(sim_mat_sorted)
ax[2].imshow(corr_mat_sorted)
Clearly, #scleronomic's solution works much better and faster, but my solution offers more control to the pattern of the output.
I have implemented sequence to sequence model with attention layer if I 300000 data points I'm not getting any error if I use all of my data points I'm getting following error model.fit
TypeError: Expected int32, got None of type 'NoneType' instead.
what would be the reason for this?
the code before model.fit is
class encoder_decoder(tf.keras.Model):
def __init__(self,embedding_size,encoder_inputs_length,output_length,vocab_size,output_vocab_size,score_fun,units):
super(encoder_decoder,self).__init__()
self.vocab_size = vocab_size
self.enc_units = units
self.embedding_size = embedding_size
self.encoder_inputs_length = encoder_inputs_length
self.output_length = output_length
self.lstm_output = 0
self.state_h = 0
self.state_c = 0
self.output_vocab_size = output_vocab_size
self.dec_units = units
self.score_fun = score_fun
self.att_units = units
self.encoder=Encoder(self.vocab_size,self.embedding_size,self.enc_units,self.encoder_inputs_length)
self.decoder = Decoder(self.output_vocab_size, self.embedding_size, self.output_length, self.dec_units ,self.score_fun ,self.att_units)
# self.dense = Dense(self.output_vocab_size,activation = "softmax")
def call(self,data):
input,output = data[0],data[1]
encoder_hidden = self.encoder.initialize_states(input.shape[0])
encoder_output,encoder_hidden,encoder_cell = self.encoder(input,encoder_hidden)
decoder_hidden = encoder_hidden
decoder_cell =encoder_cell
decoder_output = self.decoder(output,encoder_output,decoder_hidden,decoder_cell)
return decoder_output
Inside the call function I'm initializing states for the encoder where I'm getting
the number of rows from input using the following line of code
encoder_hidden = self.encoder.initialize_states(input.shape[0])
If I print input, I'm getting shape as (None,55)
That's the reason I'm getting this error.
Here my total number data points is 330614 when I use all my data I getting this
error, when I use only 330000 data points I'm getting this error,
if I print batch inside def method I'm getting shape as (64,55)
Please find my below code for creating dataset for my sequence to sequence model
the function to reprocess the data and the function to create the dataset
and a function the load the dataset
def preprocess_sentence(w):
# w = unicode_to_ascii(w.lower().strip())
w = re.sub(r"([?.!,¿])", r" \1 ", w)
w = re.sub(r'[" "]+', " ", w)
w = re.sub(r"[^a-zA-Z?.!,¿]+", " ", w)
w = w.strip()
w = '<start> ' + w + ' <end>'
return w
def create_dataset(path, num_examples):
lines = io.open(path, encoding='UTF-8').read().strip().split('\n')
# lines1 = lines[330000:]
# lines = lines[0:323386]+lines1
word_pairs = [[preprocess_sentence(w) for w in l.split('\t')] for l in lines[:num_examples]]
word_pairs = [[i[0],i[1]] for i in word_pairs]
return zip(*word_pairs)
def tokenize(lang):
lang_tokenizer = tf.keras.preprocessing.text.Tokenizer(
filters='')
lang_tokenizer.fit_on_texts(lang)
tensor = lang_tokenizer.texts_to_sequences(lang)
tensor = tf.keras.preprocessing.sequence.pad_sequences(tensor,padding='post')
return tensor, lang_tokenizer
def load_dataset(path, num_examples=None):
# creating cleaned input, output pairs
targ_lang, inp_lang = create_dataset(path, num_examples)
input_tensor, inp_lang_tokenizer = tokenize(inp_lang)
target_tensor, targ_lang_tokenizer = tokenize(targ_lang)
return input_tensor, target_tensor, inp_lang_tokenizer, targ_lang_tokenizer,targ_lang,inp_lang
# Try experimenting with the size of that dataset
num_examples = None
input_tensor, target_tensor, inp_lang, targ_lang,targ_lang_text,inp_lang_text = load_dataset(path, num_examples)
# Calculate max_length of the target tensors
max_length_targ, max_length_inp = target_tensor.shape[1], input_tensor.shape[1]
max_length_targ,max_length_inp
input_tensor_train, input_tensor_val, target_tensor_train, target_tensor_val = train_test_split(input_tensor, target_tensor, test_size=0.2)
the shape of datasets as follows
shape of input train (269291, 55)
shape of target train (269291, 53)
shape of input test (67323, 55)
shape of target test (67323, 53)
You can share the code block before the model.fit.
NoneType error is indicating that the final array which is passed to the model is for some reason empty. You can add print statements at previous steps to understand where along the way your array became empty.
Compare the scenario to the case where you are taking all your data points so that you can understand where the array is changing and how it is handled prior to passing it through model.fit.
I noticed that the sklearn.linear_model.SGDClassifier implements gradient descent for a linear model, therefore one could state that the class combines the fitting procedure (SGD) and the model (linear model) in one class.
SGD is, however, not inherent to linear models, and linear_models can be trained using many other optimizations with particular pro's (memory usage, convergence speed, local optima avoidance*, ...). One could state that such optimization techniques implement how to iterate over the training data, whether to do it online or offline, in what feature-dimension to apply an update and when to stop (possibly based on an error function of a validation set).
In particular, I implemented a model using theano and wrapped it in a fit/predict interface. Theano is cool because it allows one to define a callable which applies gradient descent on one sample, or on a set of samples, as well as a callable which returns the error on a validation set. But this coolness is not inherent to theano, a lot more models can simply define an update and error-evaluation function, which can then be used by different iteration- and stopping policies for fitting.
The theano examples often use minibatch, and the minibatch code is copy-pasted or reimplemented a lot with just minor adjustments which can easily be factored out. So I was hoping that sklearn implements something that you initialize with some parameters and an update/error callable to fit 'any model'. Or possible there is some good practice on how to do this yourself (especially w.r.t. the interface of the fitter).
Is there anything like this (in sklearn), that is Fitters which do not define the model?
*In the particular case of linear models and a l2 cost function, local optima do not exist of course, but still.
EDIT
Fair enough, this calls for a suggestion. I coded these two classes, which are not 100% clean, but they given an idea of what I mean:
import numpy
class StochasticUpdate():
def __init__(self, model, update, n_epochs, n_data_points, error=None, test_fraction=None):
self.update = update
self.n_epochs = n_epochs
self.n_data_points = n_data_points
self.error = error
self.model = model
if self.error is None and test_fraction is not None:
raise ValueError('error parameter must be specified if a test_faction (value: %s) should be used.' % test_fraction)
self.do_test = test_fraction is not None
self.n_train_samples = int(n_data_points - test_fraction) if self.do_test else n_data_points
if self.do_test:
self.test_range = numpy.arange(self.n_train_samples, n_data_points)
self.n_test_samples = int(n_data_points * test_fraction)
self.train_range = numpy.arange(0, self.n_train_samples)
def fit(self):
if self.do_test: self.test_errors = []
self.train_errors = []
self.mean_cost_values = []
for epoch in range(self.n_epochs):
order = numpy.random.permutation(self.n_train_samples)
mean_cost_value = 0
for i in range(self.n_train_samples):
mean_cost_value += self.update([order[i]])
self.mean_cost_values.append(mean_cost_value/ self.n_data_points)
if self.error is not None:
self.train_errors.append(self.error(self.train_range))
if self.do_test:
self.test_errors.append(self.error(self.test_range))
return self.model
from math import ceil
class MinibatchStochasticUpdate(StochasticUpdate):
def __init__(self, model, update, n_epochs, n_data_points, error, batch_size, patience=5000, patience_increase=2,
improvement_threshold = 0.995, validation_frequency=None, validate_faction=0.1, test_fraction=None):
super().__init__(self, model, update, n_data_points, error, test_fraction)
self.update = update
self.n_epochs = n_epochs
self.n_data_points = n_data_points
self.model = model
self.batch_size = batch_size
self.patience = patience
self.patience_increase = patience_increase
self.improvement_threshold = improvement_threshold
self.n_validation_samples = int(n_data_points * validate_faction)
self.validation_range = numpy.arange(self.n_train_samples, self.n_train_samples + self.n_validation_samples)
self.n_train_batches = int(ceil(n_data_points / batch_size))
self.n_train_batches = int(ceil(self.n_train_samples / self.batch_size))
self.train_batch_ranges = [
numpy.arange(minibatch_index * self.batch_size, min((minibatch_index+1) * self.batch_size, self.n_train_samples))
for minibatch_index in range(self.n_train_batches)
]
self.validation_frequency = min(self.n_train_batches, patience/2) if validation_frequency is None else validation_frequency
def fit(self):
self.best_validation_error = numpy.inf
best_params = None
iteration = 0
for epoch in range(self.n_epochs):
for minibatch_index in range(self.n_train_batches):
self.update(self.train_batch_ranges[minibatch_index])
if (iter + 1) % self.validation_frequency == 0:
current_validation_error = self.error(self.validation_error)
if current_validation_error < self.best_validation_error:
if current_validation_error < self.best_validation_error * self.improvement_threshold:
patience = max(self.patience, iter * self.patience_increase)
best_params = self.model.copy_parameters()
self.best_validation_error = current_validation_error
if iteration <= patience:
self.model.set_parameters(best_params)
return self.model
iteration += 1
self.model.set_parameters(best_params)
return self.model
Then for in the fit of the model one could support different training approaches and stopping criteria like this:
def fit(self, X, y):
X_shared = theano.shared(X, borrow=True)
y_shared = theano.shared(y, borrow=True)
learning_rate = self.training_method_options['learning_rate']
trainer = {
'stochastic_gradient_descent': lambda: StochasticUpdate(
self,
update=self.update_stochastic_gradient_descent_function(X_shared, y_shared, learning_rate),
n_epochs=self.training_method_options['n_epochs'],
n_data_points=X.shape[0],
error=self.evaluation_function(X_shared, y_shared),
),
'minibatch_gradient_descent': lambda: MinibatchStochasticUpdate(
self,
update=self.update_stochastic_gradient_descent_function(X_shared, y_shared, learning_rate),
n_epochs=self.training_method_options['n_epochs'],
n_data_points=X.shape[0],
error=self.evaluation_function(X_shared, y_shared),
batch_size=self.training_method_options['batch_size']
)
}[self.training_method]()
trainer.fit()
return self
Obviosuly the hash-map part is hacky, and could be done more elegantly using a standardized interface for the two classes above (since the hashmaps are still O(N*M) in size for N fitters and M models).