a3c | Tensorflow implementation of Asynchronous Advantage Actor | Reinforcement Learning library

Since you have 12 inputs so make sure you dont use too many parameters, also try changing activation function. i dont use Torch so i can not understand model architecture. why your first layer is LSTM? is your data time series? try using only Dense layer,

• 1 Dense only with 12 neurons and output layer
• 2 Dense Layers with 12 neurons each and output layer

As for activation function use leaky relu, since your data is -5000, or you can make your data positive only by adding 5000 to all data samples.

throws a domain error, even thoug I guard against negative numbers

But if (lensq > FLT_EPSILON) is too late as earlier dx*dx + dy*dy + dz*dz caused int overflow. "and indeed overflow, causing lensq to be negative" - which is undefined behavior UB.

Compiler can take advantage that sqrtf(lensq) can always work because it can assume dx*dx + dy*dy + dz*dz >= 0 and so lensq >= 0.0f.

Get rid of UB.

Short answer

Since you are running this on windows, all code that is not function or class definitions should be inside the if __name__ is "__main__" block.

Long answer

On POSIX operating systems, the multiprocessing module is implemented using the fork() system call, which creates a copy of a process. This is very handy because the second process is completely initialized out of the box.

Microsoft windows does not have this system call. So Python tries to mimick this by starting a new Python interpreter and importing your program as a module. For this to work well importing your program should not have side effects. The best way to achieve that is to put anything that is not a class or function definition inside the if __name__ is "__main__" block.

Since part of your code is outside the main block, it will be executed when your program is imported in the newly created Python processes. That is why you are seeing the multiple "epoch" prints.

The pytorch error you get means "you can only call backward on scalars, i.e 0-dimensional tensors". Here, according to your prints, policy_lossis not scalar, it's a 1x4 matrix. As a consequence, so is policy_loss + 0.5 * value_loss. Thus your call to backward yields an error.

You probably forgot to reduce your losses to a scalar (with functions like norm or MSELoss ...). See example here

The reason it does not work is the way the gradient propagation works internally (it's basically a Jacobian multiplication engine). You can call backward on a non-scalar tensor, but then you have to provide a gradient yourself, like :

Simple implementation using reduce

I was able to extend the first-network by adding a new Orderer Organization as follows:

1. Start the first-network through the byfn.sh script in the fabric-samples repo in the etcdraft mode.
2. I generated crypto-material like described in the crypto-config.yaml in the question above.
3. Use the configtxgen tool to print the new orderer organization's MSP into JSON format.
4. Mount or docker cp this JSON file to the running cli container.
5. Set the environment inside the cli container corresponding to existing ordering node. Import the latest system-channel configuration. Decode it to JSON format.

6. Edit the system channel configuration block's Orderer section to add the new orderer organization's MSP as follows:

   jq -s '.[0] * {"channel_group":{"groups":{"Orderer":{"groups": {"Orderer1Org":.[1]}}}}}' config.json orderer1org.json > config1.json

7. Edit the system channel configuration block's Consortiums section to add the new orderer organization's MSP as follows:

   jq -s '.[0] * {"channel_group":{"groups":{"Consortiums":{"groups":{"SampleConsortium":{"groups": {"Orderer1MSP":.[1]}}}}}}}' config1.json orderer1org.json > config2.json

8. Edit the system channel configuration block's Consenters section to add the TLS credentials for the new orderer organization's orderer.example1.com node as follows:

   cert=`base64 ../crypto/ordererOrganizations/example1.com/orderers/orderer.example1.com/tls/server.crt | sed ':a;N;$!ba;s/\n//g'`

   cat config2.json | jq '.channel_group.groups.Orderer.values.ConsensusType.value.metadata.consenters += [{"client_tls_cert": "'$cert'", "host": "orderer.example1.com", "port": 7050, "server_tls_cert": "'$cert'"}] ' > modified_config.json

9. Encode the block, find delta, create channel update transaction, encode it as protobuf envelope and submit the channel update transaction.

10. Fetch the latest system channel configuration block.
11. Start one of the orderers (the one who was added to consenters list previously) using this latest fetched system channel configuration block as it's genesis.block file.
12. Perform docker exec into the cli container. Using the environment of an existing orderer node, fetch the latest system channel configuration.

13. Edit the system channel configuration block to add the new orderer's endpoint in the OrdererAddresses section as follows:

    cat config.json | jq '.channel_group.values.OrdererAddresses.value.addresses += ["orderer.example1.com:7050"] ' > modified_config.json

14. Encode the block, find delta, create channel update transaction, encode it as protobuf envelope and get the block signed by Orderer1Org admin to satify the mod_policy for /Channel/OrdererAddresses resource which is set to Admins policy. This implicit meta policy expects the signature of MAJORITY Admins at that level of update. So, as the number of orderer organizations are 2 now, we need both the organization's admins to sign this system channel update transaction. Set the environment corresponding to Orderer1Org admin and run the following command:

    peer channel signconfigtx -f ordorg_update_in_envelope.pb

15. Set the environment back to OrdererOrg admin and submit the channel update transaction. The peer channel update will automatically sign the transaction on behalf of OrdererOrg admin.

    peer channel update -f ordorg_update_in_envelope.pb -c $CHANNEL_NAME -o orderer.example.com:7050 --tls true --cafile $ORDERER_CA

For updating any application channel, just replace the step 7 by updating the application channel configuration block's Application section to add the new orderer organization's MSP there.

Hope this helps!

use this dependency for glide

According to a Keras contributor in this GitHub issue:

Keras models can't be guaranteed to be thread-safe. Consider having independent copies of the model in each thread for CPU inference.

In my case, it seems that the crawler is just having a bug. I have my own canonical answer in this question, hope that helps someone: FB OpenGraph og:image not pulling images (possibly https?)