I tried to compile a tflite model to edgetpu model and run into some error like that.
Edge TPU Compiler version 16.0.384591198
Started a compilation timeout timer of 180 seconds.
ERROR: Attempting to use a delegate that only supports static-sized tensors with a graph that has dynamic-sized tensors.
Compilation failed: Model failed in Tflite interpreter. Please ensure model can be loaded/run in Tflite interpreter.
Compilation child process completed within timeout period.
Compilation failed!
I define my model like that:
preprocess_input = tf.keras.applications.efficientnet.preprocess_input
def Model(image_size=IMG_SIZE):
input_shape = image_size + (3,)
inputs = tf.keras.Input(shape=input_shape)
x = preprocess_input(inputs)
base_model = tf.keras.applications.efficientnet.EfficientNetB0(input_shape=input_shape, include_top=False, weights="imagenet")
base_model.trainable = False
x = base_model(x, training=False)
x = tfl.GlobalAvgPool2D()(x)
x = tfl.Dropout(rate=0.2)(x)
outputs = tfl.Dense(90, activation='softmax')(x)
model = tf.keras.Model(inputs, outputs)
return model
The model summary is like that:
I convert to tflite model like that:
converter = tf.lite.TFLiteConverter.from_keras_model(model)
# Defining the representative dataset from training images.
def representative_dataset_gen():
for image, label in train_dataset.take(100):
yield [image]
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
# Using Integer Quantization.
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
# Setting the input and output tensors to uint8.
converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8
tflite_model = converter.convert()
if not os.path.isdir('exported'):
os.mkdir('exported')
with open('/workspace/eff/exported/groups_1.tflite', 'wb') as f:
f.write(tflite_model)
Environment:
Edge TPU Compiler version 16.0.384591198
Python version 3.6.9
tensorflow 1.15.3
When looking for solutions on google, someone said you need to get rid of the preprocess_input, I'm not sure what that means.
How can I check if there is a dynamic shape tensor in the model
and how can I fix it?
Related
I'm a beginner using coral devboard mini.
I want to start a Smart Bird Feeder project.
https://coral.ai/projects/bird-feeder/
I've been trying to execute the code by referring to
I can't run bird_classify.py.
The error is as follows
untimeError: Internal: Unsupported data type in custom op handler: 0Node number 0 (edgetpu-custom-op) failed to prepare.
Originally, the samples in this project seemed to be deprecated, and
The edgetpu requires an old runtimeversion of 13, instead of the current 14.
(tflite is 2.5 ) I have downloaded it directly and re-installed it in
/usr/lib/python3/dist-packagesm
, but I cannot uninstall the new version and cannot match the version.
Is there a better way to do this?
Also, I've decided to give up on running the same environment as the sample, and use the pycoralapi to run the
If there is a good way to rewrite the code to use pycoral, please let me know.
Thanks
#!/usr/bin/python3
"""
Coral Smart Bird Feeder
Uses ClassificationEngine from the EdgeTPU API to analyze animals in
camera frames. Sounds a deterrent if a squirrel is detected.
Users define model, labels file, storage path, deterrent sound, and
optionally can set this to training mode for collecting images for a custom
model.
"""
import argparse
import time
import re
import imp
import logging
import gstreamer
import sys
sys.path.append('/usr/lib/python3/dist-packages/edgetpu')
from edgetpu.classification.engine import ClassificationEngine
from PIL import Image
from playsound import playsound
from pycoral.adapters import classify
from pycoral.adapters import common
from pycoral.utils.dataset import read_label_file
from pycoral.utils.edgetpu import make_interpreter
def save_data(image,results,path,ext='png'):
"""Saves camera frame and model inference results
to user-defined storage directory."""
tag = '%010d' % int(time.monotonic()*1000)
name = '%s/img-%s.%s' %(path,tag,ext)
image.save(name)
print('Frame saved as: %s' %name)
logging.info('Image: %s Results: %s', tag,results)
def load_labels(path):
"""Parses provided label file for use in model inference."""
p = re.compile(r'\s*(\d+)(.+)')
with open(path, 'r', encoding='utf-8') as f:
lines = (p.match(line).groups() for line in f.readlines())
return {int(num): text.strip() for num, text in lines}
def print_results(start_time, last_time, end_time, results):
"""Print results to terminal for debugging."""
inference_rate = ((end_time - start_time) * 1000)
fps = (1.0/(end_time - last_time))
print('\nInference: %.2f ms, FPS: %.2f fps' % (inference_rate, fps))
for label, score in results:
print(' %s, score=%.2f' %(label, score))
def do_training(results,last_results,top_k):
"""Compares current model results to previous results and returns
true if at least one label difference is detected. Used to collect
images for training a custom model."""
new_labels = [label[0] for label in results]
old_labels = [label[0] for label in last_results]
shared_labels = set(new_labels).intersection(old_labels)
if len(shared_labels) < top_k:
print('Difference detected')
return True
def user_selections():
parser = argparse.ArgumentParser()
parser.add_argument('--model', required=True,
help='.tflite model path')
parser.add_argument('--labels', required=True,
help='label file path')
parser.add_argument('--top_k', type=int, default=3,
help='number of classes with highest score to display')
parser.add_argument('--threshold', type=float, default=0.1,
help='class score threshold')
parser.add_argument('--storage', required=True,
help='File path to store images and results')
parser.add_argument('--sound', required=True,
help='File path to deterrent sound')
parser.add_argument('--print', default=False, required=False,
help='Print inference results to terminal')
parser.add_argument('--training', default=False, required=False,
help='Training mode for image collection')
args = parser.parse_args()
return args
def main():
"""Creates camera pipeline, and pushes pipeline through ClassificationEngine
model. Logs results to user-defined storage. Runs either in training mode to
gather images for custom model creation or in deterrent mode that sounds an
'alarm' if a defined label is detected."""
args = user_selections()
print("Loading %s with %s labels."%(args.model, args.labels))
engine = ClassificationEngine(args.model)
labels = load_labels(args.labels)
storage_dir = args.storage
#Initialize logging file
logging.basicConfig(filename='%s/results.log'%storage_dir,
format='%(asctime)s-%(message)s',
level=logging.DEBUG)
last_time = time.monotonic()
last_results = [('label', 0)]
def user_callback(image,svg_canvas):
nonlocal last_time
nonlocal last_results
start_time = time.monotonic()
results = engine.classify_with_image(image, threshold=args.threshold, top_k=args.top_k)
end_time = time.monotonic()
results = [(labels[i], score) for i, score in results]
if args.print:
print_results(start_time,last_time, end_time, results)
if args.training:
if do_training(results,last_results,args.top_k):
save_data(image,results, storage_dir)
else:
#Custom model mode:
#The labels can be modified to detect/deter user-selected items
if results[0][0] !='background':
save_data(image, storage_dir,results)
if 'fox squirrel, eastern fox squirrel, Sciurus niger' in results:
playsound(args.sound)
logging.info('Deterrent sounded')
last_results=results
last_time = end_time
result = gstreamer.run_pipeline(user_callback)
if __name__ == '__main__':
main()
enter code here
I suggest that you follow one of the examples available from the coral examples. There is an example named classify_image.py which uses the edgetpu (tflite) that I found works. After you install the coral examples, you have to drill down through the directory hierarchy. So, in my case, from root it is: /home/pi/ml-projects/coral/pycoral/tensorflow/examples/lite/examples. There are 17 files in that last examples directory. I'm using: numpy 1.19.3, pycoral 2.0.0, scipy 1.7.1, tensorflow 2.4.0, tflite-runtime 2.5.0.post1. I've installed the following edgetpu-runtime: edgetpu_runtime_20201105.zip.
I am using nn.DataParallel() for my model but facing an error.
I am doing something like
self.model = self.model.to(device)
self.model = nn.DataParallel(self.model)
If the device is cuda:1 then I get RuntimeError: all tensors must be on devices[0].
But if I change the device to cuda:0 the parallel training on multiple GPUs works with no error. I am wondering what the problem is.
I changed this: self.model = DataParallel(self.model) to this: self.model = DataParallel(self.model, device_ids=[1,0])
Now it is working fine. But if I write: self.model = DataParallel(self.model, device_ids=[0,1]) the error pops up. If there are more GPUs, say 4, then write: self.model = DataParallel(self.model, device_ids=[1,2,3,0])
I'm working on the exercise by Matlab on dct https://www.mathworks.com/help/images/discrete-cosine-transform.html
and i ran the following code:
I = imread('cameraman.tif');
I = im2double(I);
imshow(I)
T = dctmtx(8);
dct = #(block_struct) T * block_struct.data * T';
B = blockproc(I,[8 8],dct,'BorderSize',[4 4],'Trim',false) %modification made here
The code was from the link above, but i included the 'BorderSize' criteria to allow for overlapping of the blocks. Upon running the code, i get the error saying:
Error using blockprocFunDispatcher
BLOCKPROC encountered an error while evaluating the user-supplied function handle,FUN.
Error in blockprocInMemory
[u1_output fun_nargout] = blockprocFunDispatcher(fun,block_struct,...
Error in blockproc
result_image = blockprocInMemory(source,fun,options);
What could be the issue causing the error message above? If i remove 'BorderSize',[4 4], 'Trim',false it works as normal like in the link but am i using it wrongly?
Are you accounting for the fact that with BorderSize specified, your block size will be bigger in your definition of the function dct? Your function will receive a 12x12 block, sounds to me like a size mismatch in the matrix multiplies with T and T'. Do you need to change the definition of T to:
T = dctmx(12)
?
I have a training set of 614 images which have already been shuffled. I want to read the images in order in batches of 5. Because my labels are arranged in the same order, any shuffling of the images when being read into the batch will result in incorrect labelling.
These are my functions to read and add the images to the batch:
# To add files from queue to a batch:
def add_to_batch(image):
print('Adding to batch')
image_batch = tf.train.batch([image],batch_size=5,num_threads=1,capacity=614)
# Add to summary
tf.image_summary('images',image_batch,max_images=30)
return image_batch
# To read files in queue and process:
def get_batch():
# Create filename queue of images to read
filenames = [('/media/jessica/Jessica/TensorFlow/StreetView/training/original/train_%d.png' % i) for i in range(1,614)]
filename_queue = tf.train.string_input_producer(filenames,shuffle=False,capacity=614)
reader = tf.WholeFileReader()
key, value = reader.read(filename_queue)
# Read and process image
# Image is 500 x 275:
my_image = tf.image.decode_png(value)
my_image_float = tf.cast(my_image,tf.float32)
my_image_float = tf.reshape(my_image_float,[275,500,4])
return add_to_batch(my_image_float)
This is my function to perform the prediction:
def inference(x):
< Perform convolution, pooling etc.>
return y_conv
This is my function to calculate loss and perform optimisation:
def train_step(y_label,y_conv):
""" Calculate loss """
# Cross-entropy
loss = -tf.reduce_sum(y_label*tf.log(y_conv + 1e-9))
# Add to summary
tf.scalar_summary('loss',loss)
""" Optimisation """
opt = tf.train.AdamOptimizer().minimize(loss)
return loss
This is my main function:
def main ():
# Training
images = get_batch()
y_conv = inference(images)
loss = train_step(y_label,y_conv)
# To write and merge summaries
writer = tf.train.SummaryWriter('/media/jessica/Jessica/TensorFlow/StreetView/SummaryLogs/log_5', graph_def=sess.graph_def)
merged = tf.merge_all_summaries()
""" Run session """
sess.run(tf.initialize_all_variables())
tf.train.start_queue_runners(sess=sess)
print "Running..."
for step in range(5):
# y_1 = <get the correct labels here>
# Train
loss_value = sess.run(train_step,feed_dict={y_label:y_1})
print "Step %d, Loss %g"%(step,loss_value)
# Save summary
summary_str = sess.run(merged,feed_dict={y_label:y_1})
writer.add_summary(summary_str,step)
print('Finished')
if __name__ == '__main__':
main()
When I check my image_summary the images do not seem to be in sequence. Or rather, what is happening is:
Images 1-5: discarded, Images 6-10: read, Images 11-15: discarded, Images 16-20: read etc.
So it looks like I am getting my batches twice, throwing away the first one and using the second one? I have tried a few remedies but nothing seems to work. I feel like I am understanding something fundamentally wrong about calling images = get_batch() and sess.run().
Your batch operation is a FIFOQueue, so every time you use it's output, it advances the state.
Your first session.run call uses the images 1-5 in the computation of train_step, your second session.run asks for the computation of image_summary which pulls images 5-6 and uses them in the visualization.
If you want to visualize things without affecting the state of input, it helps to cache queue values in variables and define your summaries with variables as inputs rather than depending on live queue.
(image_batch_live,) = tf.train.batch([image],batch_size=5,num_threads=1,capacity=614)
image_batch = tf.Variable(
tf.zeros((batch_size, image_size, image_size, color_channels)),
trainable=False,
name="input_values_cached")
advance_batch = tf.assign(image_batch, image_batch_live)
So now your image_batch is a static value which you can use both for computing loss and visualization. Between steps you would call sess.run(advance_batch) to advance the queue.
Minor wrinkle with this approach -- default saver will save your image_batch variable to checkpoint. If you ever change your batch-size, then your checkpoint restore will fail with dimension mismatch. To work-around you would need to specify the list of variables to restore manually, and run initializers for the rest.
I am currently trying to implement some functions using sparkR version 1.5.1. I have seen older (version 1.3) examples, where people used the apply function on DataFrames, but it looks like this is no longer directly available. Example:
x = c(1,2)
xDF_R = data.frame(x)
colnames(xDF_R) = c("number")
xDF_S = createDataFrame(sqlContext,xDF_R)
Now, I can use the function sapply on the data.frame object
xDF_R$result = sapply(xDF_R$number, ppois, q=10)
When I use a similar logic on the DataFrame
xDF_S$result = sapply(xDF_S$number, ppois, q=10)
I get the error message "Error in as.list.default(X) :
no method for coercing this S4 class to a vector"
Can I somehow do this?
This is possible with user defined functions in Spark 2.0.
wrapper = function(df){
+ out = df
+ out$result = sapply(df$number, ppois, q=10)
+ return(out)
+ }
> xDF_S2 = dapplyCollect(xDF_S, wrapper)
> identical(xDF_S2, xDF_R)
[1] TRUE
Note you need a wrapper function like this because you can't pass the extra arguments in directly, but that may change in the future.
The native R functions do not support Spark DataFrames. We can use user defined functions in SparkR to execute native R modules. These are executed on the executors and thus the libraries must be available on all the executors.
For example, suppose we have a custom function holt_forecast which takes in a data.table as an argument.
Sample R code
sales_R_df %>%
group_by(product_id) %>%
do(holt_forecast(data.table(.))) %>%
data.table(.) -> dt_holt
For using UDFs, we need to specify the schema of the output data.frame returned by the execution of the native R method. This schema is used by Spark to generate back the Spark DataFrame.
Equivalent SparkR code
Define the schema
structField("product_id", "integer"),
structField("audit_date", "date"),
structField("holt_unit_forecast", "double"),
structField("holt_unit_forecast_std", "double")
)
Execute the method
library(data.table)
library(lubridate)
library(dplyr)
library(forecast)
sales <- data.table(x)
y <- data.frame(key,holt_forecast(sales))
}, dt_holt_schema)
Reference: https://shbhmrzd.medium.com/stl-and-holt-from-r-to-sparkr-1815bacfe1cc