Andrew Drozdov mrdrozdov

"There's no good textbooks. Everything has changed." - Fergus talking about how Computer Vision is all about Neural Nets now.

Famous Professors

Fei Fei Li - Leads Stanford Vision Group - Known for ImageNet
Yann Lecun - Leads NYU ML Group - Head of Facebook AI - Known for Convolutional Neural Nets
Yoshua Bengio - Leads U. of Montreal ML Group
George Hinton - Leads U. of Toronto ML Group - Google Researcher - Lecun, Bengio, Hinton are like the deep learning 3 musketeers.

	class RNN_Sentence(Chain):
	def __init__(self, model_dim, word_embedding_dim, vocab_size, compose_network,
	seq_length,
	num_classes,
	initial_embeddings,
	):
	super(RNN_Sentence, self).__init__(
	embed=L.EmbedID(vocab_size, word_embedding_dim, initialW=initial_embeddings), # word embedding
	mid=L.LSTM(word_embedding_dim, model_dim), # the first LSTM layer
	out=L.Linear(model_dim, num_classes), # the feed-forward output layer

	$ python -m spinn.models.fat_classifier \
	--batch_size 32 \
	--ckpt_path rnn_scratch_6 \
	--connect_tracking_comp \
	--data_type snli \
	--embedding_data_path ../glove/glove.840B.300d.txt \
	--embedding_keep_rate 0.9 \
	--eval_data_path ../snli_1.0/snli_1.0_dev.jsonl \
	--eval_seq_length 25 \
	--experiment_name model0_scratch_6 \

	$ cd path/to/spinn/checkpoints
	$ export PYTHONPATH=$PYTHONPATH:../python
	$ python -m spinn.models.fat_classifier \
	--batch_size 32 \
	--ckpt_path rnn_scratch_6.ckpt_best \
	--connect_tracking_comp \
	--data_type snli \
	--embedding_data_path ../python/spinn/tests/test_embedding_matrix.5d.txt \
	--embedding_keep_rate 0.9 \
	--eval_data_path ../snli_1.0/snli_1.0_dev.jsonl \

	created with: https://gist.github.com/mrdrozdov/d11739bdf7a9ff5d7c4783de7f8ac161

	PADDING: 16
	INITIAL SENTENCE: A person on a horse jumps over a broken down airplane
	PADDING SENTENCE: EMPTY EMPTY EMPTY EMPTY EMPTY A person on a horse jumps over a broken down airplane
	TREE DEPTH: 5

	TREE STATES:

	These are the hidden states generated with the

	import math

	def recursive_tree(root, spacing=0, depth=1):
	print("{:3}: {} {}".format(depth, "-"*spacing, str(root)))
	if isinstance(root, tuple):
	for c in root:
	recursive_tree(c, spacing+4, depth+1)

	# Preprocessing
	padding = 16

	from tf_logger import TFLogger


	""" Example of using TFLogger to save train & dev statistics. To visualize
	in tensorboard simply do:

	tensorboard --logdir /path/to/summaries

	This code does depend on Tensorflow, but does not require that your model
	is built using Tensorflow. For instance, could build a model in Chainer, then

	[user]
	name = Andrew Drozdov
	email = andrew@mrdrozdov.com
	[credential]
	helper = osxkeychain
	[color]
	ui = true
	[alias]
	gr = log --graph --full-history --all --color --pretty=tformat:"%x1b[31m%h%x09%x1b[32m%d%x1b[0m%x20%s%x20%x1b[33m(%an)%x1b[0m"
	undo-commit = reset --soft HEAD^