homework7

homework6/code/run_experiments.bat

+%# batch size%
+python gmf.py --batch_size 512 --lr 0.01 --n_emb 8 --epochs 30
+python gmf.py --batch_size 1024 --lr 0.01 --n_emb 8 --epochs 30
+python gmf.py --batch_size 1024 --lr 0.01 --n_emb 8 --epochs 30 --validate_every 2
+
+%# learning rates%
+python gmf.py --batch_size 1024 --lr 0.001 --n_emb 8 --epochs 30 --validate_every 2
+python gmf.py --batch_size 1024 --lr 0.005 --n_emb 8 --epochs 30 --validate_every 2
+python gmf.py --batch_size 1024 --lr 0.01 --n_emb 8 --lr_scheduler --epochs 30 --validate_every 2
+
+%# Embeddings%
+python gmf.py --batch_size 1024 --lr 0.01 --n_emb 16 --epochs 30 --validate_every 2
+python gmf.py --batch_size 1024 --lr 0.01 --n_emb 32 --epochs 30 --validate_every 2
+python gmf.py --batch_size 1024 --lr 0.01 --n_emb 64 --epochs 30 --validate_every 2
+
+%# batch size%
+python mlp.py --batch_size 512 --lr 0.01 --layers "[32, 16, 8]" --epochs 30 --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.01 --layers "[32, 16, 8]" --epochs 30 --validate_every 2
+
+%# learning rates%
+python mlp.py --batch_size 1024 --lr 0.001 --layers "[32, 16, 8]" --epochs 30 --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.005 --layers "[32, 16, 8]" --epochs 30 --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.01 --layers "[32, 16, 8]" --epochs 30 --lr_scheduler --validate_every 2
+
+%# Embeddings%
+python mlp.py --batch_size 1024 --lr 0.01 --layers "[64, 32, 16]" --epochs 30 --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.01 --layers "[128, 64, 32]" --epochs 30 --validate_every 2
+
+%# higher lr and lr_scheduler%
+python mlp.py --batch_size 1024 --lr 0.03 --layers "[64, 32, 16]" --epochs 30 --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.03 --layers "[128, 64, 32]" --epochs 30 --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.03 --layers "[64, 32, 16]" --epochs 30 --lr_scheduler --validate_every 2
+python mlp.py --batch_size 1024 --lr 0.03 --layers "[128, 64, 32]" --epochs 30 --lr_scheduler --validate_every 2
+
+%# neumf%
+python neumf.py --batch_size 1024 --lr 0.01 --n_emb 8 --lr_scheduler --layers "[32, 16, 8]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_512_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_512_reg_00_lr_001_n_emb_16_ll_8_dp_wodp_lrnr_adam_lrs_wolrs.pt" \
+--epochs 1 --learner "SGD"
+
+python neumf.py --batch_size 1024 --lr 0.01 --n_emb 8 --lr_scheduler --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--epochs 20 --learner "SGD" --validate_every 2
+
+python neumf.py --batch_size 1024 --lr 0.01 --n_emb 8 --lr_scheduler --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--epochs 20 --learner "SGD" --validate_every 2
+
+python neumf.py --batch_size 1024 --lr 0.01 --n_emb 8 --lr_scheduler --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--epochs 20 --learner "SGD" --validate_every 2
+
+python neumf.py --batch_size 1024 --lr 0.01 --n_emb 8 --lr_scheduler --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--epochs 20 --validate_every 2
+
+%# I repeated this experiment 3 times: with and without momentum and a 3rd time
+# with MSE but did not save it%
+python neumf.py --batch_size 1024 --lr 0.001 --n_emb 8 --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--freeze 1 --epochs 4 --learner "SGD"
+
+python neumf.py --batch_size 1024 --lr 0.001 --n_emb 8 --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--freeze 1 --epochs 4 --learner "SGD"
+
+python neumf.py --batch_size 1024 --lr 0.001 --n_emb 8 --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--freeze 1 --epochs 4 --learner "SGD"
+
+python neumf.py --batch_size 1024 --lr 0.001 --n_emb 8 --layers "[128, 64, 32]" --dropouts "[0.,0.]" \
+--mf_pretrain "GMF_bs_1024_lr_001_n_emb_8_lrnr_adam_lrs_wolrs.pt" \
+--mlp_pretrain "MLP_bs_1024_reg_00_lr_003_n_emb_64_ll_32_dp_wodp_lrnr_adam_lrs_wlrs.pt" \
+--freeze 1 --epochs 4
+
+

homework7/readme.md

+# LinUCB for Disjoint Linear Models
+
+### An implementation of the Disjoint Linear UCB bandit algorithm for the multi-arm contextual bandit problem with unbiased offline evaluation following Li(2012) [1] and [2]. Sample data can be found at: https://goo.gl/mRPGUp
+
+### References:
+#### 1. A Contextual-Bandit Approach to Personalized News Article Recommendation
+#### 2. Unbiased Offline Evaluation of Contextual-bandit-based News Article Recommendation Algorithms
\ No newline at end of file

homework7/utils.py

+"""
+Colby Wise
+LinUCB1 Multi-Arm Bandit Problem
+Re-inforcement Learning
+
+Data helper functions that return the correct
+column or row from a pandas dataframe
+"""
+
+import pandas as pd
+
+
+"""
+Read in CSV data file and remove 'timestamp' column
+@param:
+    file - file directory
+@return:
+    data - pandas dataframe of action & context vectors
+"""
+def getData(file):
+        data = pd.read_csv(file, sep=" ", header=None)
+        data = data.loc[:,:101]
+        return data 
+
+"""
+Helps sequentially retrieve a context vector at each time step
+in the algorithm
+@param:
+    data - pandas df
+    idx - time step (row) wanted
+@return:
+    x - context vector for given time step
+"""
+def getContext(data, idx):
+    return data.loc[idx, 2:]
+
+"""
+Helps sequentially retrieve an action vector at each time step
+in the algorithm
+@param:
+    data - pandas df
+    idx - time step (row) wanted
+@return:
+    action - action vector for given time step
+"""
+def getAction(data, idx):
+    return data.loc[idx, 0]
+
+"""
+Helps sequentially retrieve a reward vector at each time step
+in the algorithm
+@param:
+    data - pandas df
+    idx - time step (row) wanted
+@return:
+    reward - reward vector for given time step
+"""
+def getReward(data, idx):
+    return data.loc[idx, 1]
\ No newline at end of file