Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

fix exercise 6.5 #165

Closed
wants to merge 1 commit into from
Closed

fix exercise 6.5 #165

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
207 changes: 66 additions & 141 deletions labs/notebooks/reinforcement_learning/exercise_5.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,188 +9,113 @@
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {
"collapsed": true
},
"outputs": [],
"source": [
"%load_ext autoreload\n",
"%autoreload 2"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"As a last exercise, apply what you have learned to the RNN model seen in previous days. Implement REINFORCE to replace the maximum likelihood loss used on the RNN day. For this you can modify the PolicyRNN class in lxmls/deep learning/pytorch\\_models/rnn.py"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### WSJ Data"
]
},
{
"cell_type": "code",
"execution_count": null,
"execution_count": 2,
"metadata": {},
"outputs": [],
"source": [
"# Load Part-of-Speech data \n",
"# Load Part-of-Speech data\n",
"from lxmls.readers.pos_corpus import PostagCorpusData\n",
"data = PostagCorpusData()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Train model with Cross entropy"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"print( data.input_size)\n",
"print( data.output_size)"
"data = PostagCorpusData()\n",
"# Get batch iterators for train and test\n",
"batch_size = 1\n",
"train_batches = data.batches('train', batch_size=batch_size)\n",
"dev_set = data.batches('dev', batch_size=batch_size)\n",
"test_set = data.batches('test', batch_size=batch_size)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"execution_count": 4,
"metadata": {
"scrolled": true
},
"outputs": [],
"source": [
"# Alterbative native CuDNN native implementation of RNNs\n",
"from lxmls.deep_learning.pytorch_models.rnn import FastPytorchRNN\n",
"model = FastPytorchRNN(\n",
" input_size=data.input_size,\n",
" embedding_size=50,\n",
" hidden_size=20,\n",
" output_size=data.output_size,\n",
" learning_rate=0.1\n",
"reinforce = True\n",
"# Load version of the RNN\n",
"from lxmls.deep_learning.pytorch_models.rnn import PolicyRNN\n",
"model = PolicyRNN(\n",
"input_size=data.input_size,\n",
"embedding_size=50,\n",
"hidden_size=100,\n",
"output_size=data.output_size,\n",
"learning_rate=0.05,\n",
"gamma=0.8,\n",
"RL=reinforce,\n",
"maxL=data.maxL\n",
")"
]
},
{
"cell_type": "code",
"execution_count": null,
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"num_epochs = 10"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Epoch 1: dev accuracy 79.95 %\n",
"Epoch 2: dev accuracy 85.14 %\n",
"Epoch 3: dev accuracy 85.20 %\n",
"Epoch 4: dev accuracy 85.20 %\n",
"Epoch 5: dev accuracy 85.20 %\n",
"Epoch 6: dev accuracy 85.20 %\n",
"Epoch 7: dev accuracy 85.23 %\n",
"Epoch 8: dev accuracy 85.20 %\n",
"Epoch 9: dev accuracy 85.23 %\n",
"Epoch 10: dev accuracy 85.20 %\n",
"Epoch 11: dev accuracy 85.23 %\n",
"Epoch 12: dev accuracy 85.20 %\n",
"Epoch 13: dev accuracy 85.23 %\n",
"Epoch 14: dev accuracy 85.20 %\n",
"Epoch 15: dev accuracy 85.23 %\n",
"Training took 3.39 seconds per epoch\n",
"Test accuracy 84.11 %\n"
]
}
],
"source": [
"import numpy as np\n",
"import time\n",
"\n",
"# Run training\n",
"num_epochs = 15\n",
"batch_size = 1\n",
"\n",
"# Get batch iterators for train and test\n",
"train_batches = data.batches('train', batch_size=1)\n",
"dev_set = data.batches('dev', batch_size=1)\n",
"test_set = data.batches('test', batch_size=1)\n",
"train_batches = data.sample('train', batch_size=batch_size)\n",
"dev_set = data.sample('dev', batch_size=batch_size)\n",
"test_set = data.sample('test', batch_size=batch_size)\n",
"\n",
"# Epoch loop\n",
"start = time.time()\n",
"for epoch in range(num_epochs):\n",
"\n",
" # Batch loop\n",
" for batch in train_batches:\n",
" model.update(input=batch['input'], output=batch['output'])\n",
"\n",
" # Evaluation dev\n",
" is_hit = []\n",
" for batch in dev_set:\n",
" is_hit.extend(model.predict(input=batch['input']) == batch['output'])\n",
" accuracy = 100*np.mean(is_hit)\n",
"\n",
" loss = model.predict_loss(batch['input'], batch['output'])\n",
" is_hit.extend(loss)\n",
" accuracy = 100 * np.mean(is_hit)\n",
" # Inform user\n",
" print(\"Epoch %d: dev accuracy %2.2f %%\" % (epoch+1, accuracy))\n",
"\n",
"print(\"Training took %2.2f seconds per epoch\" % ((time.time() - start)/num_epochs)) \n",
" print(\"Epoch %d: dev accuracy %2.2f %%\" % (epoch + 1, accuracy))\n",
" \n",
"print(\"Training took %2.2f seconds per epoch\" % ((time.time() - start) / num_epochs))\n",
"# Evaluation test\n",
"is_hit = []\n",
"for batch in test_set:\n",
" is_hit.extend(model.predict(input=batch['input']) == batch['output'])\n",
"accuracy = 100*np.mean(is_hit)\n",
"\n",
"# Inform user\n",
"print(\"Test accuracy %2.2f %%\" % accuracy)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Train model with [LOSS REQUIRING RL]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Example of sampling\n",
"print(train_batches[3]['input'])\n",
"samples, log_probs = model._sample(input=train_batches[3]['input'])\n",
"samples, log_probs"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Epoch loop\n",
"start = time.time()\n",
"for epoch in range(num_epochs):\n",
"\n",
" # Batch loop\n",
" for batch in train_batches:\n",
" # TODO: Use this here to create an RL inside model.update()\n",
" samples, log_probs = model._sample(input=batch['input']) #sample actions and its neg log probs\n",
" raise NotImplementedError\n",
"\n",
" # Evaluation dev\n",
" is_hit = []\n",
" for batch in dev_set:\n",
" is_hit.extend(model.predict(input=batch['input']) == batch['output'])\n",
" accuracy = 100*np.mean(is_hit)\n",
"\n",
" # Inform user\n",
" print(\"Epoch %d: dev accuracy %2.2f %%\" % (epoch+1, accuracy))\n",
"\n",
"print(\"Training took %2.2f seconds per epoch\" % ((time.time() - start)/num_epochs)) \n",
" \n",
"# Evaluation test\n",
"is_hit = []\n",
"for batch in test_set:\n",
" is_hit.extend(model.predict(input=batch['input']) == batch['output'])\n",
"accuracy = 100*np.mean(is_hit)\n",
" is_hit.extend(model.predict_loss(batch['input'],batch['output']))\n",
"accuracy = 100 * np.mean(is_hit)\n",
"\n",
"# Inform user\n",
"print(\"Test accuracy %2.2f %%\" % accuracy)"
]
}
],
"metadata": {
"git": {
"suppress_outputs": true
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
Expand All @@ -206,9 +131,9 @@
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.6.5"
"version": "3.7.3"
}
},
"nbformat": 4,
"nbformat_minor": 2
}
}
Loading